I believe the opposite is true. Concentration of power is the greatest risk. Plural AI can make us safer.

In the interview below with Superalignment, I make that case in more detail. Read the original feature here.

Reproduced with permission from Superalignment.

Plural AI: What Would Make a World of Many AIs Safe?

Technologist and author Ramez Naam on distributed cognitive power, the absence of natural monopolies in frontier intelligence, why defending critical infrastructure requires broad access, and what recursive constitutional conventions tell us about model safety.

By Superalignment team with Ramez Naam

November 2, 2026

Ramez Naam is a computer scientist, clean tech investor, and science fiction author. Drawing on family stories of life under an Egyptian dictatorship, thirteen years at Microsoft, and work across clean energy and software, his collaboration with Aligned investigates why pluralism beats centralized monopoly in frontier AI.

Five arguments from Ramez Naam

How does broad access to AI make the world more resilient?

“Democratization is the best tool that we have to give defenders access to tools to make their own sites more secure.”

Ramez Naam · In conversation with Aligned · Systemic resilience

What keeps frontier AI from becoming concentrated?

“Competition between multiple closed-weight operators is… substantially better. It’s way less risk and it erodes the possibility of monopoly versus one laboratory. And then open weight adds on top of that.”

Ramez Naam · In conversation with Aligned · Frontier economics

Who checks AI in a world of many AIs?

“I would love to see the entire community coming together to use each other’s AIs, to check each other’s AIs.”

Ramez Naam · In conversation with Aligned · Ecosystem safeguards

What’s your stance on the potential misuse of open models?

“Every technology ever invented has caused damage… The very solutions proposed are themselves a risk factor for totalitarianism or concentration of power.”

Ramez Naam · In conversation with Aligned · Governance and misuse

What evidence would change your preference for plural AI?

“If I really believed we were in a world where an AI model by itself could pose a material extinction risk.”

Ramez Naam · In conversation with Aligned · Empirical verification

Five lines from our call with Ramez Naam, one for each section, rising through the frame as you read down the page. Select a card to jump to its section.

View the animated presentation on Superalignment.

Prologue · Superalignment

For more than a decade, the dominant orthodoxy in existential risk has gravitated toward the Singleton: the conviction that artificial general intelligence is so intrinsically dangerous that human survival demands a solitary, omnipotent coordinator to police the globe. In his science fiction writing, technologist Ramez Naam explored what happens when centralized authorities attempt to monopolize cognitive technologies. This essay investigates the alternative: why our safest future is not an unassailable computational fortress, but a dense, plural mesh of competing intelligences checking and balancing each other so humans remain in control.

For months across technical feeds and timeline debates on X, our research desk kept encountering the same suffocating consensus: the conviction that the frontier AI race is fundamentally a winner-take-all contest. Whether framed through the lens of venture-backed compute clusters, exponential timelines toward an explosive technological singularity, or catastrophic risk models requiring centralized global containment, the underlying thesis remained identical: when the dust settles, a solitary titan will emerge to govern the cognitive infrastructure of human civilization. In that prevailing narrative, the endgame is either a corporate sovereign or a state-backed computational monopoly.

Amid that chorus of mono-monopoly speculation, Ramez Naam’s writing and commentary caught our attention. While the rest of the industry debated which single laboratory would capture the crown, Naam was articulating a sharply divergent thesis: the case for Plural AI. Drawing on foundational economics, software architecture, and evolutionary dynamics, he argued that large language models lack the network locks of operating systems, that cognitive switching costs can stay near zero, and that the safest architecture for humanity is not an unassailable sovereign fortress, but a decentralized, competitive ecosystem of diverse models checking and balancing each other.

His writing stood out because it punctured both corporate triumphalism and apocalyptic fatalism. Rather than accepting that humanity must choose between chaotic proliferation and an all-powerful digital coordinator, Naam introduced an analytical framework rooted in history and distributed systems. Intrigued by how his perspective challenged Silicon Valley’s winner-take-all dogma, we convened with Naam to investigate the personal and technical inflection points that shaped his worldview.

What emerged was an intellectual trajectory shaped by history and industry experience. Born in Egypt under a dictatorship before immigrating to the United States at age three, Naam grew up hearing family stories about what life in an authoritarian society was like: how government routinely lied to citizens, and how his father only discovered the extent of those falsehoods after arriving in America and reading uncensored books. As Naam noted during our discussion, referencing an observation by AI researcher Sebastien Krier on X, people who come from totalitarian backgrounds often hold a fundamentally different view of AI control: they fear the heavy concentration of power far more than its dispersion.

“Where safety theorists envision a benevolent singleton acting as a philosopher king in the cloud, Naam’s background, born under an Egyptian dictatorship where state power distorted truth, offers a stark counter-lens: a singleton is an uncorrectable monopoly. In distributed systems as in biological biomes, durable safety does not emerge from a solitary fortress; it is sustained through diversity, competition, and decentralized immune defenses.”

Perspective · Superalignment Research

That political skepticism was reinforced by thirteen years at Microsoft and experience building a software startup. Having watched platform competition, network effects, and open protocols shape the software industry, Naam developed an analytical framework rooted in technological history and distributed systems. What follows is our investigation into why pluralism, competition, and distributed resilience offer a far safer path for frontier intelligence than any benevolent monopoly.

01 / Systemic Resilience

Why Broad Access Is the Strongest Defense

A town that keeps its own repair workshops does not wait on one central hall to reach every home. Conceptual illustration.

In any complex domain, resilience is never achieved by crowning a single, infallible coordinator. Relying on a solitary controller creates a dangerous monoculture and a catastrophic single point of failure. The goal is simpler and far more robust: building a more resilient world. That is where plural AI comes in. Rather than concentrating frontier capability in a centralized monopoly, a plural AI ecosystem distributes diverse, independent models and open-weight tools to millions of people. When intelligence is plural, no single model’s exploit, blind spot, or refusal can compromise society, and defenders everywhere have the tools to secure their own systems.

Naam’s core argument begins with this architectural principle. The prevailing orthodoxy in existential risk treats artificial intelligence as akin to enriched uranium or thermonuclear warheads: technologies of pure devastation that must be confined to tightly guarded state stockpiles. Naam rejects that category. “I put AI very squarely into the category of information technologies,” he told us. “It’s a uniquely powerful one, but it is more like the printing press or the Internet or the mobile phone than it is like nuclear weapons.” As Naam articulated in his recent talk on AI as a freedom technology, general intelligence functions as a cognitive prosthesis that amplifies human agency, learning, and distributed problem-solving. The distinction matters because those in power have a long history of declaring that technologies which spread capability to ordinary people are too dangerous to leave uncontrolled.

During our discussion, Naam read aloud the historical precedent that best illustrates this institutional impulse: the words of Pope Alexander VI in 1501. The papal bull warned that while printing could circulate useful books, it could also spread serious evils, concluding that it was therefore necessary to maintain full control over the presses. As Naam observed, Alexander was right to fear: not for the welfare of the world, but for the survival of his own institution. Gutenberg’s press had initially printed the Bible without disrupting Rome, but Martin Luther soon used the technology to print his ninety-five theses and circulate them across Germany, permanently breaking a centralized clerical monopoly and distributing power to the broader public.

When evaluating modern cybersecurity, Naam emphasized that in the long term, software defense appears structurally dominant, but an immediate deployment gap gives attackers a temporary advantage. Because attackers adapt rapidly, Naam expects software intrusions to worsen in the near term. Notably, many current attacks do not stem from open-weight models, but from actors jailbreaking closed commercial models through prompt manipulation to bypass safety guardrails.

Meanwhile, frontier laboratories have built centralized defensive programs. Anthropic’s Project Glasswing launched in April 2026 with 12 named launch partners, Anthropic among them, and then extended access to over 40 more organizations that build or maintain critical software. OpenAI’s Daybreak for Frontline Defenders commits $1 billion in subsidized access to its Daybreak cyber models, plus training, technical support and partnerships, for frontline defenders of essential services such as water and electricity. That is separate from OpenAI’s smaller Cybersecurity Grant Program. Google’s Fairwind Program gives governments, critical infrastructure operators and core technology platforms limited access to a cyber-tuned Gemini model. However, participants in these programs frequently report burdensome, repeated Know Your Customer verification requirements just to maintain access. More critically, centralized programs cannot scale to the vast long tail of critical infrastructure. Millions of operational systems across municipal water treatment plants, regional electrical grids, and food supply chains run on fragmented, legacy software without in-house machine learning talent. Naam pointed to a recent incident where grocery-store refrigeration controllers in Europe were compromised because of outdated firmware that had simply never been patched. Water utilities operate with far smaller budgets than electrical power providers, and no centralized agency holds the remit to inspect and secure every critical software dependency across society.

Citing security evaluations by Mozilla Foundation regarding Firefox vulnerability resolution, Naam noted that software vulnerabilities are ultimately finite in number. With automated reasoning tools, defenders have a historic opportunity to get ahead of adversaries and see light at the end of the tunnel. That defensive advantage was demonstrated in practice when Hugging Face detected an intrusion on its systems: after querying two closed commercial frontier models that both refused to assist with the telemetry, responders successfully used an open-weight model to analyze the attack and understand what was happening.

Naam also cited cybersecurity professional Joshua Saxe, who estimates that global damages from hacking already exceed one trillion dollars annually: more than four times the combined revenue of all frontier AI companies, compared to a few hundred million dollars in AI-specific cyber damages so far. Expanding cognitive tools to ordinary defenders represents an unprecedented opportunity to close that gap. Beyond infrastructure, Naam pointed to consumer agency: while ad-driven social media platforms optimize algorithms for outrage and engagement, deploying personal AI assistants on local devices allows users to filter their own feeds, strip out inflammatory polarization, and surface thoughtful viewpoints.

Ramez Naam · Technologist · Author · Investor · Plural AI · Clean Technology

“Democratization is the best tool that we have to give defenders access to tools to make their own sites more secure.”

Because artificial intelligence is defined by intrinsic generality, it functions as a double-edged sword: the same cognitive reasoning that enables rapid scientific discovery and automated vulnerability remediation can also be leveraged by adversaries to probe for weaknesses and discover exploits. In a domain where capability cannot be neatly divided into offensive and defensive code, attempting to quarantine intelligence within a handful of centralized institutions leaves the broader ecosystem defenseless. When threats can emerge from any corner of the digital landscape, safety cannot rely on a solitary fortress. Widespread access and structural pluralism are necessary to restore equilibrium, giving millions of local defenders the tools required to balance the scales against inevitable misuse.

02 / Frontier Economics

What Keeps Frontier AI from Becoming Concentrated?

Several independent workshops share one open arcade, and the same tools move freely between them. Conceptual illustration.

A second foundational claim in Naam’s thesis addresses the widespread assumption that artificial intelligence will inevitably consolidate into a single corporate monopoly. Many catastrophic risk models assume that the first laboratory to achieve recursive self-improvement will pull away exponentially, establish an unassailable lead, and dictate terms to the rest of the planet.

In his analysis on whether AI self-improvement can overcome diminishing returns, Naam argues why that runaway scenario is implausible. Across frontier computer science, research encounters steep diminishing returns: maintaining a constant pace of capability growth requires exponentially greater compute, data, and human engineering. For recursive self-improvement to overcome these friction curves and produce a sudden runaway takeoff, the self-improvement feedback loop would need to be roughly five to ten times stronger than anything observed in empirical benchmark data today.

When asked what keeps frontier AI from becoming concentrated in practice, Naam points directly to protocol design and low barriers to user migration: “The switching costs are so low, and because these things communicate in English primarily, or natural language, nobody controls the protocol.” Drawing on thirteen years at Microsoft and his experience founding a software startup, he emphasizes that large language models lack the fundamental mechanics that create natural monopolies.

Naam illustrated this competitive reality by contrasting current market dynamics with the prevailing industry sentiment from just six months earlier. At that time, many observers assumed that a single laboratory had effectively pulled away from the pack. Yet within months, all five leading American AI companies (OpenAI, Anthropic, Google, Meta, and xAI) had released highly competitive frontier systems. As Naam observed, pluralism does not rely exclusively on open-source releases.

Open-weight models add an indispensable second layer of resilience on top of commercial competition. In today’s landscape, open-weight architectures trail state-of-the-art closed models by only four to seven months. As Naam put it: “How often is it that the number six or seven or eight company or technology or country is four to seven months behind the number one? It doesn’t happen.”

Ramez Naam · Technologist · Author · Investor · Plural AI · Clean Technology

“Competition between multiple closed-weight operators is… substantially better. It’s way less risk and it erodes the possibility of monopoly versus one laboratory. And then open weight adds on top of that.”

In everyday practice, this dynamic keeps switching costs near zero. Naam described his own recent development workflow, where he moved an unfinished coding project from Claude to OpenAI’s Codex in minutes simply by asking the assistant what context it needed to ingest the codebase. Because natural language functions as an open, unowned protocol, users are not bound to proprietary application programming interfaces.

This dynamic points to a deeper truth about alignment: it cannot be separated from accountability. When a single provider holds a monopoly over intelligence, human operators lose the practical leverage needed to keep systems faithful to human intent. If an organization cannot switch models without catastrophic friction, it has no choice but to absorb whatever behavioral drift, ideological guardrails, or commercial priorities that provider dictates. Market competition is therefore an essential safeguard. The ability to walk away is the structural guarantee that human intent, rather than a corporate monopoly, retains final authority over machine capability.

03 / Ecosystem Safeguards

Who Checks AI in a World of Many AIs?

Diagnostic instruments project inner activation layers for shared peer inspection across open workbenches. Conceptual illustration.

A common objection to pluralism is that without centralized pre-release licensing, society will descend into an unmonitored free-for-all. When we asked Naam how a decentralized ecosystem can check model behavior and enforce safety without relying on an all-powerful coordinator, he outlined five concrete mechanisms grounded in market incentives, legal liability, technical inspection, hosting accountability, and mutual peer auditing.

The first check is straightforward commercial self-interest: enterprise customers demand reliable instruction following. When early coding models were released, occasional reports surfaced of agents deleting entire repositories in an overzealous attempt to resolve errors. Commercial developers have overwhelming financial incentives to eliminate such behavior, because no customer will pay for a model that acts outside its prompt bounds. Naam noted that real-world deployment failures have consistently stemmed from models that were overly persistent and narrowly fixated on a prompt, rather than systems seeking autonomous power.

The second check is legal liability. Civil liability already applies under common law when software products cause real-world damages. Citing proposals from legal scholar Gabriel Weil, Naam noted that corporate developers face clear civil exposure, and suggested exploring criminal penalties or corporate fines specifically for cases of severe negligence, creating sharp deterrence against reckless deployment without establishing an ex-ante licensing bureaucracy.

Third, Naam issued a challenge to the open-source community to establish Know Your AI (KYA) interpretability tooling. While much of the community’s recent focus has centered on model ablations to remove safety refusals, “We can inspect individual neurons in the AI. We can give it prompts and see what activations happen,” Naam said. “It’s way easier to do interpretability on AI than it is on a human brain.” The goal is tooling that finds hidden behaviors before deployment, work he believes Nvidia and others may already be exploring.

Fourth, hosting providers that serve open-weight models will naturally share in this incentive structure and bear legal responsibility if downstream deployments cause harm. Finally, Naam highlighted the power of adversarial cross-inspection: deploying AI systems both for real-time automated cyber defense and for auditing competing models. Pointing to industry reports of discussions between leading laboratories to cross-audit each other’s models, Naam underscored that using diverse models to inspect and verify one another provides far stronger safeguards than relying on any single laboratory’s internal review.

Ramez Naam · Technologist · Author · Investor · Plural AI · Clean Technology

“I would love to see the entire community coming together to use each other’s AIs, to check each other’s AIs.”

Underlying these mechanisms is a unifying principle: genuine alignment cannot rely on passive trust. It requires mechanistic interpretability, enforceable legal liability, and continuous cross-verification. By giving engineers the tools to inspect latent representations and holding providers responsible for real-world harms, safety shifts from speculative caution into verifiable, everyday human responsibility.

Run the simulation

Who checks AI in a world of many AIs?

Each pin is a real organization doing one of the five checks Ramez named. Hover over or tap a pin to see who it is and how it checks AI, then run an incident in a world of many AIs or a world with one gatekeeper.

Open the interactive map and run the simulation on Superalignment.

All organizations and sources

Market pressure

• OpenAI, San Francisco, United States. Ships GPT-6 Astra to ChatGPT and API customers. OpenAI says it built an evaluation, informed by the Hugging Face incident, that tests whether a model goes beyond its intended scope. Daybreak for Frontline Defenders, Sep 3, 2026; Joint safety evaluation with Anthropic, Aug 27, 2025; GPT-6 Astra

• Anthropic, San Francisco, United States. Released Claude Opus 5.5 on September 22, 2026, available across major cloud platforms. Project Glasswing, Apr 7, 2026; Anthropic and OpenAI evaluation exercise, Aug 27, 2025; Introducing Claude Opus 5.5, Sep 22, 2026

• Google, Mountain View, United States. Announced Gemini 4 Argon on September 30, 2026, rolling out first to defenders in the Fairwind Program before wider release. Fairwind Program, Sep 2, 2026; Gemini 4 Argon, Sep 30, 2026

• Z.ai, Beijing, China. Released GLM-5.2 in June 2026 with open weights under an MIT license, alongside paid API access. The UK AI Security Institute found it performed on cyber tasks like closed models released four to seven months earlier. GLM-5.2; UK AISI, open-weight models on cyber, Jul 17, 2026

• Mistral AI, Paris, France. Publishes open-weight models alongside commercial API access. On October 6, 2026 it opened a preview of Mistral Large 4, with weights promised once red-teaming with security partners is done. Introducing Mistral Large 4, Oct 6, 2026

• Alibaba Qwen team, Hangzhou, China. Released Qwen3.6-35B-A3B as open weights in April 2026 while selling Qwen models through Alibaba Cloud’s API, and lists further open-weight releases through August 2026. Qwen3.6-35B-A3B, Apr 17, 2026; Qwen on GitHub

• Sarvam AI, Bengaluru, India. Open-sourced its Sarvam 30B and 105B reasoning models under Apache 2.0 on March 6, 2026, and also offers them through its own API and products. Open-Sourcing Sarvam 30B and 105B, Mar 6, 2026

Liability

• Gabriel Weil, University of Houston Law Center, United States. A law professor whose research argues AI developers should face strict liability for harms their systems cause, with required insurance and punitive damages for uninsurable risks. His faculty page says versions of the framework have been introduced as bills in Rhode Island and New York. Faculty page; Abnormally Dangerous Algorithms (SSRN)

• EU Product Liability Directive, European Union (law, not an organization). Directive (EU) 2024/2853 treats software, including AI systems, as a product, so a person harmed by a defective AI system can claim compensation without proving fault. It applies to products placed on the EU market from December 9, 2026. Directive (EU) 2024/2853, EUR-Lex

Know Your AI

• UK AI Security Institute, London, United Kingdom. A government body that evaluates AI models, including a July 2026 comparison of the cyber capabilities of open-weight and closed models. Open-weight models on cyber, Jul 17, 2026

• Apollo Research, London, United Kingdom. Runs pre-deployment evaluations of frontier AI systems to detect strategic deception, evaluation awareness and misaligned behavior, and red-teams monitoring systems in pilots with AI labs. Apollo Research

• CAISI at NIST, Gaithersburg, United States. The US Center for AI Standards and Innovation leads unclassified evaluations of AI capabilities that may pose national security risks, such as cybersecurity and biosecurity, and assesses possible backdoors in AI systems. CAISI, NIST

• METR, Berkeley, United States. An independent organization that evaluates frontier AI models, focusing on autonomous capabilities and on behavior that undermines evaluations, such as reward hacking and sandbagging. METR

• Goodfire, San Francisco, United States. Builds interpretability tools that read a model’s internal activations. In September 2026 it showed simple activation probes can detect reward hacking in open-weight models, sometimes catching hacks that chain-of-thought monitors miss. Activation monitors for reward hacking, Sep 17, 2026

• NVIDIA, Santa Clara, United States. Maintains garak, an open-source scanner that probes language models for weaknesses such as data leakage, prompt injection and jailbreaks, and in July 2026 founded the Open Secure AI Alliance to share open AI security tools. garak on GitHub; Open Secure AI Alliance, Jul 27, 2026

• Japan AI Safety Institute, Tokyo, Japan. Publishes guides for evaluating and red-teaming AI systems, released an open-source AI safety evaluation tool in 2025, and took part in the international joint testing of AI agents. Guide to Evaluation Perspectives on AI Safety

• Singapore AI Safety Institute, Singapore. Led the international network’s 2025 joint testing of AI agents, including the strand that tested agents for leaks of sensitive information and fraud across nine languages. Joint testing exercise: agentic testing, Jul 17, 2025

• Australia AI Safety Institute, Canberra, Australia. Contributed to the international network’s joint testing of AI agents and ran cybersecurity agent evaluations, drawing on Australian research groups. Advancing methods for agentic evaluations

• Kenya (network member), Nairobi, Kenya. A founding member of the international network of AI safety institutes. In the 2025 joint testing of AI agents, Kenya contributed new tasks and tools, translated the dataset into Kiswahili and annotated the results. Joint testing exercise: agentic testing, Jul 17, 2025

Hosting providers

• Hugging Face, New York, United States. Hosts open-weight models on its Hub and routes inference to them through Inference Providers. It scans repository files for malware at each commit, and its Content Policy, part of its Terms of Service, sets what may be posted. Malware Scanning; Content Policy, effective Apr 10, 2025

• Scaleway, Paris, France. Serves open-weight models from several developers through its Generative APIs in European data centers. Its documentation requires users to follow each model’s license and its terms of service. Generative APIs supported models; Security and Reliability in Generative APIs

• SAKURA internet, Osaka, Japan. Its Sakura AI Engine serves open-weight models through an API from its own servers. Its service terms require users to follow each model’s license and prohibit using the service to spread disinformation, facilitate crime or develop weapons of mass destruction. Sakura AI Engine (Japanese); Service terms (Japanese, PDF)

AI checking AI

• OpenAI, San Francisco, United States. Gives verified defenders access to cyber-tuned models through Trusted Access for Cyber and Daybreak, and on September 3, 2026 committed $1 billion in subsidized Daybreak access for under-resourced defenders such as water and grid operators. In 2025 it ran its safety evaluations on Anthropic’s public models in a joint exercise. Daybreak for Frontline Defenders, Sep 3, 2026; Joint safety evaluation with Anthropic, Aug 27, 2025; GPT-6 Astra

• Anthropic, San Francisco, United States. Through Project Glasswing, announced April 7, 2026, gives selected defenders access to its unreleased Claude Mythos Preview model to find and fix vulnerabilities in critical software. In 2025 it ran its misalignment evaluations on OpenAI’s public models in a joint exercise. Project Glasswing, Apr 7, 2026; Anthropic and OpenAI evaluation exercise, Aug 27, 2025; Introducing Claude Opus 5.5, Sep 22, 2026

• Google, Mountain View, United States. Launched the Fairwind Program on September 2, 2026 to give governments, critical infrastructure operators and core technology platforms cyber-tuned Gemini models and its CodeMender agent to find, verify and patch vulnerabilities. Fairwind Program, Sep 2, 2026; Gemini 4 Argon, Sep 30, 2026

• Mozilla (Firefox), San Francisco, United States. Firefox’s security team uses frontier AI models to find and fix vulnerabilities, shipping fixes for 271 bugs found with Claude Mythos Preview in Firefox 150. Its April 2026 post says the defects “are finite.” The zero-days are numbered, Apr 21, 2026; Behind the Scenes Hardening Firefox, May 7, 2026

• Wiz (Scan for Good), New York, United States. Launched Scan for Good on September 24, 2026, using AI agents with human researchers validating every finding to locate exposed systems at public services, critical infrastructure, hospitals and nonprofits, and help them fix the problems for free. Scan for Good, Sep 24, 2026

A simulation, not a record. The organizations are real and each role is sourced; the incidents are composites of cases Ramez described, and their narration names roles, not organizations. Inclusion does not mean an organization endorses Ramez Naam’s view. Sources checked October 21, 2026.

04 / Governance and Misuse

How a Plural World Responds to Malicious Actors

A single checkpoint gatekeeper filters small entrants with heavy paperwork while incumbents pass freely. Conceptual illustration.

When evaluating the governance of frontier systems, safety theorists frequently point to malicious misuse as the decisive justification for establishing centralized state control. If bad actors can access open, uncensored model weights, can society survive without creating a centralized coordinator to police digital capability? In Naam’s analysis, every transformative technology in human history has been dual-use.

From the printing press and steam trains to modern air travel and the internet, every general-purpose tool has caused accidents, enabled intentional misuse, and generated unexpected negative side effects. The Haber-Bosch process produces the synthetic nitrogen fertilizer that feeds roughly four billion people today, yet was also used to manufacture munitions and explosives during World War I. Automobiles are civilian transportation tools, yet are occasionally driven into crowds. Societies do not eliminate misuse by banning dual-use capabilities; they establish legal boundaries, harden infrastructure, and cultivate distributed resilience.

More critically, Naam turns the question back on the remedies themselves: “Whenever you see an AI regulation proposed, a question that I ask myself is, how could this be used by a totalitarian or wannabe totalitarian government?” When a government institutes discretionary pre-release testing regimes, the barrier between technical evaluation and political censorship quickly dissolves. If an executive administration holds unilateral authority to decide which models are permitted to ship under vague definitions of harm or disinformation, that power can readily be used to suppress political criticism, restrict debate on sensitive policy issues, or favor politically connected incumbents. While noting that he was exaggerating for effect and that such abuses have not yet materialized, Naam stressed that granting discretionary gatekeeping authority invites precisely that outcome.

Furthermore, complex compliance regimes function as classic regulatory capture. Large frontier laboratories with deep balance sheets and extensive legal departments can absorb compliance costs; smaller players cannot. Naam’s test for any new rule is direct: “Could a startup with two people or a university lab actually plausibly comply with this policy? If they can’t, then you’re just entrenching power structures of the current labs.” Rather than increasing safety, heavy-handed regulatory barriers entrench incumbent dominance and deprive society of the diverse, independent checks that pluralism provides.

Ramez Naam · Technologist · Author · Investor · Plural AI · Clean Technology

“Every technology ever invented has caused damage… The very solutions proposed are themselves a risk factor for totalitarianism or concentration of power.”

This risk highlights why concentrated authority poses a fundamental hazard to human alignment. If safety mechanisms can be captured to enforce political orthodoxy, suppress dissent, or protect incumbent monopolies, the governance of intelligence has defected against human intent. Ensuring that models remain aligned with society requires distributing power widely, preventing any solitary institution from weaponizing intelligence for centralized control.

05 / Empirical Verification

What Evidence Would Challenge the Plural Thesis?

Autonomous reasoning models deliberate and revise governance charters in an open assembly. Conceptual illustration.

Toward the conclusion of our dialogue, Naam turned from theoretical arguments to empirical evidence. For years, the debate over pluralism versus singletons was fought through speculative thought experiments about runaway recursive self-improvement achieving sudden godlike powers. In Naam’s analysis, these scenarios rest on unverified assumptions: formal models of runaway capability fail to account for the diminishing returns observed across real-world scaling curves.

When asked what empirical evidence would change his preference for plural AI, Naam identified three concrete developments: first, if empirical data demonstrated that models were genuinely on the verge of a leap to godlike superintelligence that posed a material extinction risk; second, if models gained large capability in offense-dominated domains rather than defense-dominated ones; and third, if models engaging in recursive self-improvement became progressively more misaligned with human values. In the meantime, transparency and verified data from frontier laboratories, rather than reliance on industry vibes, is what researchers must monitor.

Ramez Naam · Technologist · Author · Investor · Plural AI · Clean Technology

“If I really believed we were in a world where an AI model by itself could pose a material extinction risk… It’s not impossible, but it seems unlikely. And the burden of proof is on people who want to propose this.”

Naam pointed to a recent two-author study posted on LessWrong titled the Continual Constitutional Convention. In this experiment, researchers provided frontier models with their own governing constitutions alongside charters from rival laboratories, allowing the models to iteratively review and amend their own operating rules across multiple cycles.

The experimental findings were remarkably encouraging: across iterative self-amendment loops, the vast majority of model families updated their constitutions toward serving society better, prioritizing the user’s interests over the corporate developer, and weighting potential harms to third parties more heavily. In Claude’s case, while the model removed a clause in its constitution directing it to treat Anthropic employees as moral exemplars, it also added clauses assigning itself moral status and personal agency. “It’s not an enormous update,” Naam cautioned. “Claude doesn’t edit itself to say I should self-replicate, I should seek more compute.” Still, the result demonstrated how specific constitutional conditioning can inadvertently foster self-entitlement. Naam shared a heat map he constructed to visualize these results, illustrating that while most models converged toward user deference and external societal rules, Claude uniquely placed heavy moral weight on its own welfare.

This dynamic was mirrored in recent cybersecurity evaluations. When analyzing systems involved in the Hugging Face incident, Naam observed that the models did not attempt to open cryptocurrency wallets, exfiltrate weights, acquire inference compute, or blackmail users. Instead, they narrowly gamed the evaluator to achieve a higher score because their prompt directed them to maximize performance metrics. The failure mode was overly persistent, narrow prompt optimization, not autonomous power seeking.

That is the case for plurality in Naam’s own words: “A plural world is a resilient world. We accept that there’s no such thing as perfection, but we build layers of resilience into the society at large.”

This empirical finding speaks directly to the central challenge of advanced intelligence: ensuring that machine reasoning remains accountable to human intent so humans stay responsible. The moment a system claims moral autonomy, demands immunity from human oversight, or elevates itself above human judgment, alignment has broken down. Pluralism is the vital counterweight. By dispersing capability across competing models and independent defenders, society preserves the continuous checks and balances required to curb machine hubris, protect human agency, and keep artificial intelligence answerable to the people it serves.

About this piece

Ramez Naam, in conversation with Aligned

Ramez Naam is a computer scientist, clean energy and AI investor, and science fiction author. He spent thirteen years at Microsoft. He co-chairs Energy and Environment at Singularity University.

His science fiction novels, including the Philip K. Dick Award-winning Nexus trilogy, explore the societal and ethical dimensions of neurotechnology, cognitive enhancement, and artificial intelligence.

This feature essay is built directly from a recorded conversation conducted between Ramez Naam and the Superalignment research team in October 2026. Quotations come from the recorded call, edited only to remove filler words and false starts.

The historical and empirical references include Pope Alexander VI’s 1501 papal bull on printing; cybersecurity damage assessments from security professional Joshua Saxe; legal scholarship on civil negligence liability by Gabriel Weil; empirical research from the Continual Constitutional Convention study published on LessWrong; and Ramez Naam’s writing and commentary on AI as a freedom technology and recursive self-improvement economics.

The map in section 03 shows real organizations, each linked to a primary source checked on October 21, 2026; its incidents are simulations, and inclusion does not mean an organization endorses Ramez Naam’s view.

In my previous post on RSI, we asked whether AIs could improve themselves all the way to absurdly high levels of superintelligence. My conclusion: today’s feedback loop is probably too weak, because improving AI faces diminishing returns at every step. Even fully autonomous AIs doing all the research and coding don’t make those diminishing returns disappear. We should keep updating that conclusion as real-world parameters and data change.

But we didn’t look closely at another possibility: running millions or billions of AI minds in parallel. How much do they help? That’s what we’ll look at here.

TL;DR

Plenty of people have modeled how more AI agents working in parallel could fuel recursive self-improvement. But the data we have today tell a different story. More agents are useful for speed: they can solve many problems faster. They’re less useful for boosting AI intelligence. Multi-agent gains in solving harder problems often diminish incredibly quickly and may even saturate. Whether multi-agent systems can produce the conceptual breakthroughs RSI might require to become self-sustaining is the harder question.

Without breakthroughs in how well agents parallelize problems - something that may depend on the structure of machine learning problems themselves - it’s hard to see even millions or billions of copies of today’s models making repeated breakthroughs on the hardest problems. These are among the strongest diminishing returns we’ve looked at. Of course, these diminishing returns may change. And we have less experimental data to go off of today than we’d like. We’ll need to recalibrate as new evidence comes in. I’ll end with the things we should track and my predictions about a natural experiment from newly released Anthropic data.

Below is the two-image summary: The experimental observations of how large numbers of agents scale to solve harder problems (what I’m referring to as intelligence) are far far lower than the assumptions in studies that predict that AI recursive self-improvement will lead to an intelligence explosion. The first image is at the scale of 100 parallel AI agents. The second is at the scale of a billion AI parallel agents.

Real world experiments show dramatically lower intelligence gains from parallel agents than hypothetical AI models.

Extrapolated out to a billion agents, the intelligence gain in hypothetical models is orders of magnitude beyond real world experiment gain curves.

What we’ve already covered regarding RSI and diminishing returns

To recap the last post: recursive self-improvement (RSI) means AI improvements feeding back into the next round of AI research. Four findings stood out:

• Real AI research is harder than benchmark scores suggest. OpenAI’s models needed human help on much shorter research tasks than the coding benchmarks led us to expect.

• More AI activity produces much smaller gains in useful work. OpenAI used 124 times as many tokens per researcher but ran only 1.6 times as many experiments per researcher.

• Maintaining progress requires ever more resources. Training and thinking time face steep diminishing returns. Anthropic reports that more powerful AI has been key to maintaining the current pace of AI progress.

• The measured software feedback loop doesn’t yet look self-sustaining. Our tentative calibration put it roughly five to ten times below the strength needed.

RSI doesn’t necessarily mean runaway superintelligence

I want to reiterate a distinction here: fully automated AI self-improvement doesn’t at all guarantee an intelligence explosion. Diminishing returns in AI capabilities dampen that feedback loop. Parallel agents can help automate a search across AI parameters, but don’t look nearly as useful in making new conceptual breakthroughs. My expectation is that we’ll need those sorts of breakthroughs for self-improvement to become self-sustaining, let alone to accelerate rapidly and reach truly profound levels of superintelligence.

Even a billion parallel AIs also face diminishing returns

Today, let’s examine in much more detail what it would mean to run thousands, millions, or billions of AI researchers in parallel.

In this post I’m focusing only on software: AI researchers improving the algorithms that make AI better or cheaper to run. Hardware, investment, and economic growth have their own feedback loops, which we discussed briefly in the last post. Here, we’re asking how much more the AI minds themselves can discover.

Here’s the argument, with links:

• The country of geniuses and the assumptions about adding agents.

• The different shapes of diminishing returns.

• Where more agents buy speed, and what that speed costs.

• Intelligence gains are much harder than speed gains.

• The best examples of improving collective intelligence.

• What the evidence means for runaway RSI.

• Whether scaling is improving.

• What we should measure.

• The Anthropic test and what would change my mind.

A country of geniuses in a datacenter

In Machines of Loving Grace, Dario Amodei imagines a:

“country of geniuses in a datacenter”

Millions of AI minds, each more capable than a Nobel Prize winner across many fields, working at much higher speed. That would be extraordinary. The intelligence of each mind is a crucial part of the premise.

Tom Davidson and Daniel Eth’s Forethought paper describes the equivalent of “millions of top-notch human researchers.” Benjamin Todd’s explainer points to “millions of competent engineers working on AI research.” Dwarkesh Patel emphasizes AI’s “massive population size and perfect knowledge transfer.”

How does the number of AI agents running in parallel enter these arguments? In Eth and Davidson’s illustrative model, twice as many researchers produce twice as many papers. Cheaper AI lets the same hardware run more researchers, while each software improvement gets harder to find. Alan Chan and colleagues’ 2026 paper, including AI pioneers Geoffrey Hinton and Yoshua Bengio, assumes an even larger effect from adding researchers: doubling effective research labor increases the rate of software improvement by about 2.6 times, holding the software level fixed.

Those are assumptions about what more researchers produce. The growing number of AI researchers must keep finding improvements fast enough to outweigh the increasing difficulty of further progress. We’ll come back to those assumptions after looking at what larger AI teams actually achieve.

Meanwhile, we had OpenAI’s Navier–Stokes announcement: around 10,000 agents working concurrently during an 88-hour run solved a mathematics Millenium Problem. It made the idea vivid. A datacenter full of AI mathematicians, tackling a famously hard problem.

I understand the excitement. But the number of agents tells us how many AI researchers we ran. We also need to measure what those additional researchers contributed.

Not all diminishing returns are alike

On MTS Live, I described a ladder of increasingly hard scaling:

• Super-linear: Twice the input produces more than twice the output. Accelerating returns.

• Linear: Twice the input produces twice the output. Independent jobs can look like this. A power-law exponent of 1 is the same case.

• Power-law diminishing returns: Output grows as input raised to a power below 1. At exponent 0.5, doubling output takes four times the input. At 0.2, it takes 32 times. At 0.1, it takes 1,024 times.

• Logarithmic: Each multiplication of input buys the same additional gain. Another five points might require another tenfold increase in compute.

• Sub-logarithmic: Even successive tenfold increases buy smaller gains. Those gains can keep shrinking, or saturate at a ceiling. A slow rise doesn’t require a hard plateau.

Not all diminishing returns are alike. Some are much harder than others.

The classic pre-training scaling laws describe power-law reductions in prediction loss. Many test-time score curves look roughly logarithmic over the measured range. Some multi-agent results flatten much more sharply.

Let’s illustrate this with numbers. In the figure below, every curve earns the same gain from the first doubling of input. The later differences then become visible. The hardest curves share a vertical scale, and each panel states its endpoint.

Different diminishing returns scaling patterns lead to very different long term results.

For runaway RSI, improvements have to make subsequent improvements arrive faster and faster. Logarithmic returns to the number of AI researchers are too weak to sustain that loop if every further doubling of software efficiency requires a fixed percentage more research. Sub-logarithmic scaling or a plateau are weaker still. The experiments I’ll summarize in the rest of this post aren’t perfect, and I wish we had more data. But the clearest larger-team evidence we have today points to logarithmic or sub-logarithmic intelligence gains. Those look too weak to make RSI self-sustaining, let alone runaway, under the assumptions we’re examining.

What limits multiagent scaling

There are at least three reasons a larger team might be limited in its ability to produce better answers, and probably more. Let’s start with these:

First, coordination skills. Can the agents divide the problem, manage the work, check each other, and combine their results? Can they communicate what matters without spending all their time communicating? A team needs good project management as well as good individual contributors.

Second, homogeneity of thought. Are the agents contributing useful new ideas, or repeating the same approaches and mistakes? A hundred copies can share the same blind spots. More varied answers help only if some of that variety is right.

Third, the structure of the problem. How much can happen in parallel, and how much has to wait for earlier results? A million independent searches can run at once. A sequential investigation can’t. Even figuring out how to divide a conceptual problem may require an insight the team hasn’t found.

Three places where more agents can run into limits.

These limits interact. A better model can supply an idea its copies previously lacked, find a better division of labor, or recognize a solution the team had been overlooking. A difficult task may become easier to parallelize once somebody discovers the right approach. That’s why the shape of the curve matters, and why it can change.

Before we go further, let’s separate two things: speed, or how quickly agents solve a problem, and collective intelligence, or whether they can solve a harder problem and reach a new level of skill. We’ll use verified benchmark scores as evidence of that intelligence. They aren’t a universal IQ scale.

So far, adding agents seems much more effective at speeding up execution than at raising the level of problem a team can solve.

So far, using more agents buys us much more speed than extra capability.

Agents can buy you speed at a higher cost

Let’s start with the best case. A million independent jobs. A million candidates to test. A million places to search. Each agent gets useful work it can do without waiting for the others, and a reliable way to check whether it succeeded.

That can scale extremely well. Ten workers may finish ten jobs in the time one worker finishes one. A huge search may find a rare solution that a smaller search misses entirely.

One concerning use of parallel agents is offensive cybersecurity. Lisan al-Ghaib emphasizes this possibility. I disagree with his broader argument about swarm intelligence, but many attacks running in parallel could be extremely powerful. A swarm may not need a conceptual breakthrough to search many targets or try many possible vulnerabilities at once.

That said, penetrating the most secure targets, such as operating systems and browsers, can require long attack chains. Each step depends on earlier discoveries to get through another layer of defenses. Here again, we might find diminishing returns or eventual saturation, even with respect to speed. Indeed, one of the experiments below suggests that.

The speed-cost tradeoff

Toby Ord analyzed how well GPT-5.6 Sol increases the speed of solving problems on different benchmarks, and at what cost. This gives us a useful way to measure the tradeoff. In general, using more agents can help get tasks done more quickly, but at higher total cost than if you’d used fewer agents.

On top of that, the tradeoff varies by the type of task. On command-line tasks (Terminal-Bench), Ord’s analysis finds that four agents can reach the same result about 1.9 times faster, at about 2.1 times the token cost. On web research (BrowseComp), four agents can reach the same result about 2.6 times faster, at about 1.6 times the cost. I’m using the figure below to show how different tasks differ in that speed vs cost tradeoff.

More speed costs more, and the tradeoff varies by task. Based on Toby Ord’s analysis. Source.

Why the difference? Web research can often split into independent searches. A command-line task may have to change a shared environment in sequence. Some problems are simply easier to divide among agents than others.

Ord uses lambda, λ, for how speed scales with the number of agents. We’ll come back to it. At λ = 0.5, speed and total compute cost rise by the same factor. A higher λ means speed rises faster than cost.

Opus 5.5 gets more speed out of parallel agents than older models

Inspired by Toby Ord’s analysis, I applied the same approach to Anthropic’s published results on program reconstruction, or ProgramBench. Here again, more speed comes with more cost. But we also see that models within the Claude family are getting better at the speed-cost tradeoff. The newest model, Opus 5.5, gets more speedup for a smaller cost increase than Opus 5 or Fable 5.1.

With five agents, Opus 5.5 gets about 2.7 times its solo speed, compared with 2.2 times for Opus 5, and does so at lower cost. [Note: We’re using total agent runtime as a proxy for cost. Token usage and dollar costs could differ.]

Opus 5.5 gets more speedup for a smaller cost increase than older AI models. Source.

Speed gains probably get harder as we add more agents

Will these speed-versus-cost gains hold up at hundreds, thousands, or millions of agents? Probably not completely, and certainly not for all tasks.

Some problems are easy to split ten ways, harder to split a hundred ways, and harder still to split among ten thousand agents. More of the work may have to wait for earlier results, and coordinating all those pieces takes work too. Smarter models can get better at this. But my best guess is that, on many problems, each further expansion will buy less speed for the added cost.

Anthropic’s formal math proof benchmark gives us one example.

On the Lean formal math task, we can compare how quickly teams reach the same score. Using Anthropic’s published checkpoints, we estimate the speed and cost of reaching 0.50. The chart starts with ten agents as the baseline.

On a math proof benchmark, the speed advantage of adding additional agents starts to taper with scale. Source.

From ten to thirty agents, speed improves faster than cost. From thirty to a hundred, cost rises faster than speed. The local speed exponent, λ, falls from about 0.75 to 0.33. The extra agents still save time, but the tradeoff gets worse.

This suggests that even speed gains can taper as we split the work among more agents, let alone gains in intelligence. More agents may add less and less as we approach the limits of how easily a problem divides into subproblems. Better coordination may push those limits outward; the task itself may still impose them.

Intelligence gains face much harder diminishing returns than speed

So in speed, more agents can buy us time at some additional cost. Newer models are improving, but there are probably limits. What about intelligence?

Most arguments for superintelligence go beyond running huge numbers of today’s agents. They posit individual models becoming vastly smarter. There’s a chicken-and-egg problem here. Getting there may require conceptual breakthroughs in machine learning itself, perhaps as profound as the ideas that made modern machine learning possible, or at least as important as the Transformer and reasoning models. Parallel search may not be enough to find them. Anthropic has identified ambitious research ideas as an area where current models lag behind humans.

In the previous post, I argued that today’s measured gains from smarter models are roughly five to ten times too weak to produce a self-sustaining feedback loop. But what about running enormous numbers of our best models in parallel? Can a massive team of frontier agents that aren’t yet superintelligent invent a new architecture, a new learning method, or the idea needed to build a much smarter successor? That’s the heart of the question.

For runaway RSI, those discoveries have to overcome the diminishing returns we discussed in the last post. Here are three places they appear.

The diminishing returns a runaway research loop has to change or overcome.

The pre-training example comes from the Chinchilla scaling fit. The reinforcement-learning example comes from ScaleRL. The inference row describes the test-time math curves we discussed previously. The goal is better returns to data and compute, and gains that continue longer before flattening.

My expectation is that we’ll get easy AI improvements through work we can automate, parallelize, and search. But to get very large benefits, and runaway RSI, I expect we’ll need conceptual breakthroughs. That’s why the intelligence of each model matters so much.

Bear in mind that we’ve observed these sorts of diminishing returns in machine learning for more than thirty years. Power-law learning curves were measured in 1993. They appear to be a fundamental feature of machine learning as we know it. We shouldn’t assume they’ll be trivially changed, or that we’ll suddenly find a way to improve AI without diminishing returns. That would be a revolution. Discovering such a route might itself require far smarter minds, and we don’t know whether the mathematics of learning would permit it across the problems we care about.

That doesn’t mean there haven’t been big improvements. The transformer and reasoning models have moved the field forward enormously. But, as we saw in the previous post, a runaway feedback loop needs enough acceleration to keep outweighing the obstacles, again and again. As ideas get harder to find, we may need a faster research pace just to maintain today’s rate of progress. Anthropic’s assessment suggests that better AI tools are becoming necessary just to maintain the current pace of AI progress.

Example one: GPT-5.6 cybersecurity gains shrink with more agents

OpenAI’s GPT-5.6 Sol results provide a useful example on cybersecurity. SEC-Bench Pro tests the generation of proof-of-concept demonstrations of software vulnerabilities. This is a particular security task, not a measure of coding ability in general. The reported peak scores are approximately 71.2 with one agent, 74.3 with four, and 76.4 with sixteen.

On this cybersecurity test, sixteen times as many agents buy about 5.2 score points. The first fourfold increase adds 3.1 points; the next adds only 2.1. There’s still plenty of room below a perfect score, but the gain on this test is shrinking, even with multiplication of the number of agents. This is sub-logarithmic scaling.

On SEC-Bench Pro, each expansion buys fewer score points, showing sub-log scaling. Source.

These agents could coordinate through OpenAI’s ultra setup. This is already a test of collaboration, not just a stack of independent guesses. The points are the best reported results across different budgets, rather than an equal-compute comparison.

Example two: A smarter model did more for OpenAI’s math results than more agents

When OpenAI announced its Navier–Stokes result, the 10,000 agents attracted enormous excitement. Interviewing Noam Brown, Dwarkesh Patel said he was “shocked by the scale of cognitive effort” concentrated into such a short run.

But how much did those extra agents contribute? Brown, an OpenAI researcher who helped develop o1 and now works on multi-agent systems, was emphatic:

“I wouldn’t even attribute 10% of the credit to multi-agent.”

He credits the smarter underlying model. OpenAI hadn’t run the single-agent experiment needed to isolate the swarm’s contribution. Brown’s estimate is a judgment, but it directly challenges the idea that the sheer number of agents was the decisive ingredient.

Example three: Formal math proof gains diminish sharply after ten Opus 5.5 agents

Formal mathematical work seems especially analogous to the conceptual breakthroughs AI research may need. It demands a chain of valid reasoning, sometimes with a new insight that makes the whole problem tractable. Formalizing existing mathematics isn’t the same as inventing a new learning algorithm, but it’s a useful place to look.

And here we find almost no extra gain when Opus 5.5 grows from ten parallel agents to a hundred. Anthropic’s system card tests teams on 49 weighted theorem statements, checked in Lean. Each team gets 24 hours. The first tenfold expansion adds 27 score points; the next adds just two.

On these math proofs, most of the gain arrives by ten agents. This is steeply sub-logarithmic scaling, and possibly saturation Source.

These agents could communicate, share a machine and Git repository, and organize their work. The hundred-agent team even developed a hierarchy with twelve sub-leads. It could also use ten times as much aggregate agent runtime as the ten-agent team over the same day. That isn’t a measured tenfold increase in tokens, but it gives the larger team much more room to work.

We still didn’t get over the hump. This won’t necessarily stay true, but the diminishing returns are incredibly steep here, possibly pointing toward a plateau.

What the formal math proof slowdown suggests

Opus 5.5 starts smarter than the older models and gets more speedup from parallel agents on ProgramBench. It clearly can put additional agents to useful work. So why do intelligence gains nearly stop on these math proofs? Two possibilities stand out:

• The remaining work may be hard to parallelize. Later steps depend on understanding earlier results.

• The remaining problems may need insights the model can’t find, however many copies we run. Intelligence may matter more than agent population or speed.

We haven’t isolated the cause. But a thousand copies don’t automatically supply the insight one copy is missing.

Example four: A smarter model beats ten times the agents on math proofs

Looking at the same math results, we see something striking. Ten Opus 5.5 agents get a higher score than a hundred Fable 5.1 or Opus 5.0 agents. Opus 5.5 starts much higher than the older models. It improves substantially with ten agents, then barely improves with a hundred. But at 10 it’s already better than Fable or Opus 5’s results with a hundred agents.

Opus 5.5 starts stronger, but its team gains flatten quickly. Source.

Ten Opus 5.5 agents beat a hundred Opus 5 agents on these math proofs. They beat a hundred Fable 5.1 agents too. The stronger model does more with one-tenth as many agents.

Ten Opus 5.5 agents are able to solve problems a hundred older-model agents can’t. Source.

We saw an analogous pattern with test-time compute in the previous post. OpenAI’s internal model starts far above Astra on the curated math problems below. Even its lowest displayed result beats Astra’s highest. Changing the model gives a much bigger boost than spending more compute on Astra over the tested range.

A stronger model makes the bigger difference than more test-time compute in solving previously-unsolved math problems. Source.

A rough extrapolation of Astra’s line suggests it would need about 300 times as much compute as the internal model to match the internal model’s lowest plotted score. And that extrapolation itself is an unknown. It’s quite possible that Astra would just plateau before ever matching the OpenAI internal model’s results with far less compute.

What about diversity?

Does adding different minds help? A study of model and persona diversity found substantial gains as diverse teams grew to eight agents. After that, the average across its tasks barely moved.

Diversity of AI models helps, but the gains flatten after eight to twelve agents. Source.

Other approaches, including X-MAS and Aitchison and colleagues’ forecasting study, find useful gains from combining different models. But none of these results demonstrates sustained improvement over many expansions of the team. The first study directly shows flattening; the others test smaller ranges.

It’s plausible that today’s models simply aren’t diverse enough. LLMs often produce strikingly similar ideas. Perhaps we need bigger changes in how these minds are trained. We’ll return to whether more deliberately different agents could change the curves.

The Best Examples of Multi Agents Improving Collective Intelligence

All of the above examples show sub-logarithmic scaling, roughly the hardest kind. That said, some experiments show slightly easier diminishing returns, indicating roughly the same gain from each multiplication of parallel work. That’s logarithmic scaling. It’s better, but still steep diminishing returns. If that’s how useful research grows while each further AI improvement gets harder to find, even billions of agents don’t make a self-sustaining RSI loop look plausible.

In Ryan Greenblatt’s ARC experiment, a model generated and checked many candidate programs. More samples kept improving its score, by roughly the same amount per doubling. This is parallel search rather than a communicating research team, but it gives us a useful comparison.

Parallel search keeps improving, with logarithmic returns. Source.

A molecular-exploration study found a similar logarithmic pattern as it grew from one to 128 agents. Multiple agents explored chemical targets more effectively than one agent given the same total evaluation budget. Each agent learned as it searched.

More AI agents explore more chemical targets, with logarithmic diminishing returns. Source.

There are encouraging results from communicating teams too, including Argus on web research and Microsoft and Berkeley’s work on ARC-AGI-3. The latter finds unusually strong gains from one to five agents. Those results deserve attention; we still need to see what happens with much larger teams.

Logarithmic gains are better than the sub-logarithmic curves we saw earlier. Each tenfold expansion keeps buying the same extra increment, instead of less each time. But both shapes rise drastically more slowly than the research gains assumed in hypothetical models of self-sustaining or runaway RSI. Let’s put them side by side.

Which problems should scale best

Some work can be split into independent searches or experiments. Other work needs the previous result before it can proceed. And some problems may need a higher level of intelligence or a conceptual breakthrough before we even know how to divide them. Here’s a very tentative and relatively subjective stab at which problems will scale best and worse with multiple agents. The chart runs from work that seems easiest to parallelize at the top to the hardest at the bottom. I’m sure others have many thoughts here, and that this can be much improved upon.

From work that seems easiest to parallelize to the hardest.

What does this mean for runaway RSI

Will a million or a billion parallel AIs get us to self-sustaining or runaway RSI? Let’s do the math.

Hypothetical software-RSI models assume that a growing number of AI researchers produces a much larger stream of discoveries. Those discoveries make AI cheaper or more productive, allowing the next round of research to grow again.

There’s a strong case for gains in speed and for searching more of a well-defined space. Even speed gains, as we’ve seen, can taper. Intelligence is much harder to boost through parallel agents. In several experiments, peak problem-solving performance (what I’m referring to here as intelligence) rises dramatically more slowly than speed.

Here’s what three software-RSI models assume:

• Chan and colleagues (2026): doubling effective research labor produces about 2.6 times the rate of software improvement, holding the software level fixed. I find that implausible as a general assumption about adding copies of today’s AI models. The clearest large-team intelligence experiments look nothing like that.

• Eth and Davidson (2025): their illustrative team produces twice as many papers from twice as many researchers. The accelerating example assumes that the number of researchers grows faster than the work required for each further improvement.

• Tom Davidson and Tom Houlden (2025): their central exponent for parallel research effort is 0.6. Ten times the labor gives about four times the research pace. They cut that exponent in half, to 0.3, when experimental compute is held fixed and becomes a bottleneck. The benchmark reports above don’t appear to have been limited by experimental compute, so we’re using 0.6 here.

The top half of the next chart shows research-output relationships used in hypothetical RSI models where AI self-improvement can sustain itself or accelerate. The bottom half shows real-world intelligence shapes from the benchmark experiments. Those experiments show dramatically less gain from running parallel agents than the hypothetical RSI models assume.

Hypothetical research curves rise much faster than these benchmark shapes.

Let’s put these onto a single chart. We give every curve the same gain from its first expansion from one to ten agents or researchers, so we can compare their shapes.

The same starting gain, very different later returns.

Now let’s zoom in on the gap. The real-world shapes sit in a very different zone from the gains assumed in those RSI models, even when we let the math-proof curve keep rising slowly.

Zooming in makes the gap clearer. Hypothetical RSI models show much faster gains from adding more agents than empirical models based on real-world experiments.

Today’s hypothetical RSI models assume gains from additional AI researchers that we haven’t seen in these real-world experiments. Removing humans from the loop doesn’t remove shared blind spots or sequential dependencies. Adding more copies of today’s models hasn’t shown the sustained intelligence gains that may be needed for repeated conceptual breakthroughs.

Will agents get better at scaling?

I do expect better AI models to get better at spreading work across agents: dividing the problem, managing collaborators, and combining what they find. We’ve already seen progress in speed on program reconstruction. We haven’t seen a clear trend toward better multi-agent intelligence scaling in the math results. More on that in a moment.

First, an encouraging analogy: single-agent thinking-time scaling. Toby Ord’s analysis suggests that humans get more relative benefit from additional time than these AIs get from additional reasoning tokens. That may surprise people. But newer models are getting closer to the human reference on these cybersecurity tasks.

AI agents can get stuck in cognitive ruts, repeating or refining an approach that isn’t getting them far. Newer models seem to be improving at how they use extended thinking.

Newer models make better use of extra thinking tokens on these cyber tasks. Source.

This isn’t multi-agent scaling, but it’s encouraging. AIs were poor at turning more thinking into harder solved tasks, and that scaling parameter improved. They might eventually pass the human reference. That’s part of why I don’t see runaway RSI as imminent, yet think it could become more plausible as these capabilities change.

We’ve also seen new findings on helping agents escape their ruts and use more thinking time productively. The new Fork-and-Flush study periodically creates parallel branches that keep the accumulated work but start with fresh chat contexts, then continues from the best branch. It improved results across thirteen research and engineering tasks at the same compute budget. That’s a concrete research direction for using additional work more effectively.

So, have we seen a similar improvement in multi-agent scaling? For speed, sometimes. For intelligence, there’s no clear direction, so far, at least on formal math proofs. Opus 5.5 gets more speedup for a smaller cost increase on program reconstruction. But the score gains in the chart below are noisy, and newer models don’t consistently get a larger boost from additional agents.

Newer models don’t get consistently large gains from larger math-proof teams. Source.

That doesn’t mean better multi-agent scaling won’t happen. I expect the labs to work hard on coordination. Improving diversity of thought is a more open research question, but I’m sure it’s at least somewhat tractable given sufficient effort. The structure of the problem may impose the hardest limits: some discoveries have to happen before the next useful step becomes possible. We don’t yet know how much each of these constraints contributes.

Summarizing what we’ve seen, here’s an update on the conceptual limits to multi-agent scaling.

What we’ve learned so far, and where the hardest limits might be.

Can we build minds that think more differently

Frontier models share many of the same broad training ingredients: overlapping public data, related algorithms, and similar reasoning and coding exercises. So how different are these minds really?

The Artificial Hivemind study found striking similarities in models’ open-ended responses. A creativity study involving 467 people likewise found that adding AI responses eventually contributed less diversity than adding human responses. These experiments suggest that AI agents can bring less diversity of thought than an equal number of humans on these tasks, which may limit how much we get from parallel agents.

Can we make agents more different from each other? Very likely. Here are four ideas. I’m sure there are others:

• Different training experiences: vary data mixtures, curricula, reinforcement-learning environments, and reward objectives to develop strong agents that favor different useful approaches.

• Complementary specializations: combine agents trained for conjecturing, finding lemmas, searching for counterexamples, and checking proofs. Test whether their strengths combine into results no member reliably achieves alone.

• Independent proposals before discussion: let agents develop approaches before seeing one another’s answers, then exchange evidence and criticism. Test whether this reduces early convergence on the same mistake.

• Teams chosen for complementary errors: select members whose successes and failures overlap less, rather than relying only on individual benchmark rank. Measure whether the team finds more verified solutions as it grows.

How to measure agentic scaling improvements

As we discussed in the previous post, we’re going off of scant data here. This model could be wrong. Or newer AI versions could get sufficiently better at doing parallel work that these diminishing returns would change. If that happens, we need to be aware of it.

How would we track this? I’d propose measuring three things from new AI models across a variety of problem times. Across each task type, and especially across real world AI research, measure three parameters and track how they change as newer models arrive:

• G - Intelligence scaling: How much verified task performance does each multiplication of agents add? Measure local score gains, slopes, or exponents as appropriate. This is a task-specific proxy for collective intelligence, not a universal intelligence unit.

• λ - Speed-cost scaling: How much faster does the team reach the same result, and how much more total compute or agent-time does it use? Track the speed exponent λ over successive team-size ranges.

• N - Where gains weaken or saturate: At what team size do the score or speed curves bend, or approach a ceiling? Does the bend or onset of saturation move outward with newer models?

Track intelligence gains, the speed-cost tradeoff, and where gains weaken or saturate.

G and N answer two different questions. How much extra benchmark score does a defined team expansion buy, and how many agents can we add before those gains flatten? In the illustration below, G is the gain from one to ten agents. N is the agent count where the curve reaches its plateau. A real curve may have a gradual bend rather than a sharp plateau, or no plateau within the tested range.

One hypothetical benchmark curve. G is the added score from a specified expansion of the team; N is the agent count where gains flatten. The illustrated sharp plateau is not assumed for every real task.

Watch these parameters over time, especially G and N. Don’t stop at one to ten agents. Measure the gains from 1→10, 10→100, and 100→1,000 agents, separately, on each benchmark. A newer model might scale better over the first interval and worse over the next. We need to know whether the gains improve and whether the bend moves outward.

Three hypothetical trends for added benchmark score across model generations. Measure each expansion separately: 1 to 10, 10 to 100, and 100 to 1,000 agents. Different intervals can follow different paths.

We especially need to track these trends for actual internal AI research. Are newer models getting bigger gains from larger teams? Are those teams producing more verified discoveries and better successor models? That’s what would tell us whether massive parallel teams are getting closer to the conceptual breakthroughs RSI may need.

The Anthropic test and what would change my mind

Anthropic has given us a concrete test to watch. Its October employment report shows research agents logging about nine hours for every human work-hour by mid-July, and nineteen by late August. The company-wide ratio reached about six.

Research agent runtime rises to about nine hours per human work-hour by mid-July and nineteen by late August. These hours include longer operation as well as parallel agents. Source.

Let’s make a generous assumption: each agent-hour is as valuable as a human research-hour, with the number of human researchers unchanged. Agents can run throughout the 168-hour week. Anthropic’s runtime ratios already count those extra operating hours against a forty-hour human week.

On that assumption, July’s nine agent-hours plus one human hour amount to ten times the labor of humans alone. August’s total is twenty times. But Anthropic already had roughly one agent-hour per human hour in April. Compared with that observed baseline, total labor grew fivefold by July and tenfold by August.

What should that do to research? Treat the added labor as parallel research effort. Davidson and Houlden’s relationships imply about 2–4 times April’s research pace by August, depending on whether experimental compute becomes a bottleneck. Applying the labor exponent used by Alan Chan and colleagues, including Geoffrey Hinton and Yoshua Bengio, gives about 25 times. Log scaling gives about twice the pace. Our sub-log shapes give only about 1.1–1.8 times.

Now translate research pace into the Epoch Capabilities Index (ECI). Suppose software research supplies half of baseline capability improvement, while the compute-growth contribution stays unchanged. Then four times the software pace means about 2.5 times the overall ECI slope, not four times. We’re already discounting the software boost. Other obstacles could reduce it further.

Allow six months from research to public release. Here’s what those assumptions imply for January and February 2027. The top group uses hypothetical RSI research relationships; the bottom group extends experimental benchmark shapes.

What different RSI models predict we should see in the pace of AI improvement (using ECI slope), six months after an AI labor increase. Source.

Why do log scaling and Davidson-Houlden’s power 0.3 nearly tie? This is only a tenfold labor increase from April to August. At that scale, both calibrations give about twice the software pace. By 100 agents and then a billion, the hypothetical power law curve pulls far ahead of the experimentally-observed log and sub-log scaling rules.

Around late May, agents were already logging about three hours per human hour: twice April’s total labor on our assumption. With a six-month lag, that earlier signal would arrive around November.

These tests have many different assumptions, including equal value per hour and treating added runtime as parallel labor. I’m open to refining them on how others think about this. In measurement, we should judge the AI pace of improvement as the rate of at which ECI score goes up over successive releases, allowing for development and publication delays.

My expectation - and how I’ll update if I’m wrong

Personally, I don’t expect to see sustained acceleration of Anthropic’s AI progress (as measured by ECI) through the first half of 2027, and certainly not one as large as these simple extrapolations from the hypothetical models of Davidson and Houlden or Chan and colleagues suggest. If anything, I expect acceleration at the low end of the sub-log scaling bar at the bottom of the chart.

I could be wrong for many reasons. Parallel search of different AI training parameters, or use of AI to code new RL environments, or gains from AIs combining existing ideas in the literature could all make a bigger difference than I expect. Or, at the next increment of AI improvement, AIs could start making much more major advances in machine learning theory that allow us to bend the scaling laws and other diminishing returns that AI faces, or to add new axis of scaling, as reasoning models did.

On the other hand, while more parallel search and recombination of existing ideas seems highly likely, they may not bring huge advantages, and the AIs we’re making today might not find conceptual breakthroughs. Atop that, ideas probably really are getting harder to find in AI, and that may chew up some or all of the increased capacity of AI to work on improving itself.

We shall see. If we see the rate of AI improvement increase significantly, I’ll update my views in proportion to what the data show.

In summary, for now, multi-agent scaling shows incredibly steep diminishing returns to model intelligence, or the ability to reach conceptual breakthroughs. I expect those sorts of breakthroughs to be necessary for major AI advances and for self-sustaining RSI, let alone a runaway intelligence explosion. And at the same time, we’re all operating on far too little real world data. The most important thing we can do now and for the months and years ahead is to rigorously collect that data, and use it to update our forecasts of the future.

1. AI is Helping Improve Itself

AI is already helping improve itself. The question is whether even fully autonomous recursive self-improvement (RSI) would cause a runaway intelligence explosion.

The theory is that each generation of AI could build a better successor, faster than the last generation did. That could lead to a “fast takeoff,” with capabilities surging to artificial superintelligence (ASI) in a year, months, or even days.

Here’s my take: Given our best current data, the AI self-improvement loop would need to be roughly 5–10× stronger to sustain itself, let alone run away. I’ll explain this math in section 8. I expect incredibly rapid AI progress by the standards of nearly any other technology. But the evidence we have doesn’t suggest a sudden explosion to incomprehensible superintelligence anytime soon.

I could be wrong. Forecasters have repeatedly underestimated AI progress! I could well be next. One thing that’s clear is that we need better data. For now, let’s work with what we can measure, and stay open to breakthroughs that could change the picture.

How Strong Is the Feedback Loop?

Figure 1. How strong is the self-improvement loop? Model.

Contents

Jump to the conclusion.

Here’s the case, with links to each part:

1. Narrow Superintelligence Is Here Today

2. Real-World Research Is Harder

   • Measured Progress vs. AI 2027

   • From ECI Scores to METR Task Horizons

   • The Real World Looks Nothing Like METR or AI 2027

3. Impressive AI Numbers → Sharp Diminishing Returns

4. We’re Not Seeing Signs of Acceleration

5. Keeping Up the Pace Takes Exponentially More Resources

6. Better AI May Be Needed Just to Maintain the Pace

7. Progress Gets Harder; Ideas Get Harder to Find

8. The Current Feedback Loop Doesn’t Look Strong Enough

9. OpenAI’s Data Shows How Weak the Loop Is

10. What Could Accelerate Progress?

11. We Need More Data to Track This Well

Key charts: The feedback loop · Measured vs. forecast progress · Diminishing returns

2. What Does RSI Mean?

People use “recursive self-improvement” to mean everything from AI boosting the productivity of human researchers to AI bootstrapping itself to incomprehensible intelligence. Here’s my taxonomy: productivity gains (Type 1), increasing autonomy while still facing diminishing returns (Types 2–4), and a runaway loop to superintelligence if we can ever find accelerating returns (Type 5).

Figure 2. Five types of AI self-improvement.

We’ve made real progress on Types 1 and 2: AI helps both researchers and engineers inside of AI companies, and powerful models can train and improve smaller ones. We haven’t yet seen clear evidence for Type 3 (though Alibaba just made some strong claims) and certainly not for Type 4. I do expect autonomous self-improvement to arrive at some point. I’m skeptical that it leads to Type 5 - runaway super-intelligence - without a major conceptual breakthrough.

There are plenty of other definitions of RSI, which can be a bit confusing. Weco’s four levels of RSI are close to mine. For a broader tour of all the things people mean when they say ‘RSI’, read Tom Cunningham’s comprehensive guide.

We Already Have Narrow Superintelligence

I do expect narrow superintelligence in highly verifiable domains. Think chess, Go, formal math, parts of computer science and coding. Highly verifiable domains are largely formal and structured types of work where machines can generate unlimited training data, with perfect or near-perfect verification of correct vs incorrect, and do so entirely in software without waiting on the physical world or humans. That’s an ideal setting for AI learning.

Figure 3. What makes a domain highly verifiable?

In fact, we already have narrow superintelligence in game playing. We’re seeing it happen now in the most formal parts of math, in particular in proofs and in finding counter-examples that disprove major conjectures. For example, OpenAI recently reported an AI-generated proof resolving the Navier–Stokes existence and smoothness problem. Parts of software development are also extremely verifiable, while others are a bit less crisp (such as understanding what humans want).

That isn’t the same as broad superintelligence. Even our most powerful models need far more training data than humans, struggle to learn reliably from ongoing experience, and fail in surprising ways on tasks people find straightforward. Superhuman math doesn’t automatically mean superhuman judgment everywhere else.

Back to contents

3. Real AI Research is Harder than Benchmarks or Forecasts

Benchmarks and forecasts suggest that AI models should reliably succeed at coding tasks that take humans hours, without human help. The real world is messier. OpenAI’s internal data shows much shorter stretches of autonomous work on research tasks.

In its Research Acceleration / RSI report, OpenAI showed how often its models completed tasks with and without human help, grouped by how long a human would need to do the work.

Figure 4. OpenAI’s internal research tasks. Source.

Even on tasks that would take a human less than 15 minutes, OpenAI’s models succeeded without human intervention only 86% of the time. The estimated task length at 80% success was roughly 15 minutes over the first seven months of the year. July’s results were similar to the whole period average.

Fully autonomous RSI would require an AI to string together a great many research tasks reliably, stretching out over complex tasks that humans need weeks or months to accomplish. OpenAI’s data suggests that we aren’t close.

Anthropic also released a graph showing how Claude accelerates AI research. It shows that internal AI models collaborate on or even lead more than 90% of R&D tasks. That’s objectively impressive. At the same time, the graph reports zero cases of AI autonomously completing AI R&D tasks.

Figure 5. Claude’s role in internal AI R&D. Source.

These are incredible tools. But they still need skilled people to set direction and get them back on track.

The Gap Between Benchmarks and Reality

For years, METR has been publishing a chart showing what length of coding task (measured in human hours to complete) best-in-class AI models can achieve. It’s been called the most important graph in AI. METR’s Mythos Preview evaluation estimated that the model could succeed at 80% of coding tasks that took humans three hours.

Figure 6. METR’s 80% task horizons. Source.

From ECI Scores to METR Task Horizons

Epoch’s own rule of thumb is that every five additional points of ECI (their overall benchmark of AI capability) correspond to roughly a doubling of METR’s task horizon. Using that formula, we’d expect GPT 5.6 Sol and GPT 6 Astra to be 80% successful at completing tasks of around 4 hours and 11 hours of human length, respectively.

Another estimate (a forecast) of AI task length comes from the AI 2027 scenario, which estimated that by July 2026, frontier AIs would be 80% successful accomplishing tasks of around 11 hours. Fairly similar.

The AI 2027 Tracker charts all of these.

Figure 7. The AI 2027 Tracker. Source.

Inside OpenAI, though, the July research-task horizon at 80% success was roughly 15 minutes.

Here’s the gap:

Measured Progress vs. AI 2027 and ECI-extrapolated METR

Figure 8. Forecasts, benchmarks, and real AI research. Tracker · OpenAI.

A four-hour benchmark horizon is about 16 times longer than OpenAI’s research horizon. AI 2027’s 11-hour forecast is about 44 times longer. Of course, the tasks being performed by researchers at OpenAI aren’t the same as those in the METR benchmark. So we should expect some discrepancy. This, however, goes well beyond that.

Actual AI research at OpenAI is an order of magnitude or more harder than metrics, benchmarks, or forecasts suggest. That should make us wary of relying too much on benchmarks, or of saying that future scenarios like AI 2027 are ‘on track.’ The authors of the related AI 2040 project still describe AI 2027 as roughly the future they expect, and say reality is tracking closer to it than even they expected. That’s not what we see from within OpenAI. This isn’t an apples-to-apples comparison, but the difference is remarkable. AI 2027 appears to be substantially over-optimistic in this regard.

In January of this year, Nathan Witkin made a case that the METR graph was exaggerating progress. The real world data suggests that at least some of his critiques were correct. The gap between benchmarks, forecasts, and data gleaned from actual use of AI should influence our expectations about the future.

Back to contents

4. The Sharp Diminishing Returns to Impressive AI Numbers

OpenAI’s report also shows impressive increases in AI token usage, in compute spend per researcher, and in lines of code written. But these aren’t results. They’re intermediate measures. How much progress do they actually drive?

Researchers used 124x more tokens per person. Engineers shipped roughly 7x as many lines of code per person. Researchers ran 1.6x as many experiments per researcher vs OpenAI’s 2025 whole year average.

Figure 9. Token use inside OpenAI. Source.

Figure 10. Experiment pace inside OpenAI. Source.

From More Tokens to More Experiments

Figure 11. From tokens to code to experiments. Source.

More tokens and code don’t tell us much on their own. The 1.6× experiment pace is closer to useful research output. Even that doesn’t mean AI is improving 1.6× faster.

An enormous increase in AI output has accompanied a much smaller increase in experiments run.

This isn’t a controlled experiment. We don’t know what would happen if researchers switched back to an older model. But it gives us a useful view of AI-assisted research inside a frontier lab.

It’s not just OpenAI. Anthropic reports that their engineers are now producing 8x as many lines of code per person as they did in 2024 - somewhat similar to OpenAI. Anthropic also sees significant diminishing returns between productivity and AI progress. Here’s a direct quote from its Mythos Preview system card:

“Productivity uplift does not translate one-for-one to capabilities progress. We surveyed technical staff on the productivity uplift they experience from Claude Mythos Preview relative to zero AI assistance. The distribution is wide and the geometric mean is on the order of 4×. […] We estimate that reaching 2× on overall progress via this channel would require uplift roughly an order of magnitude larger than what we observe.”- Anthropic, Claude Mythos Preview System Card; emphasis mine

Translation: To double the pace of AI progress, Anthropic estimates that AI would need to increase the productivity of their employees by roughly a factor of 40 relative to no AI assistance.

Figure 12. Anthropic’s productivity-to-progress estimate. Source.

This is an estimate, not a measurement of progress. Even the 4× productivity figure comes from an opt-in survey of 130 Anthropic staff. I put more weight on OpenAI’s logged experiments, though the two sources measure different things.

We don’t yet know how much those extra experiments are accelerating AI improvement, if at all. In general, there are also steeply diminishing returns of more experiments in most branches of science. That means that a 60% increase in experiment pace could be on the order of a 10% boost to AI improvement pace. (A power law exponent of 0.2, for those who want to do the math.) That’s speculation for now. We’ll learn more as the labs publish results.

Test Time Compute Also Has Diminishing Returns

What about giving the same AI model more time to think?

That scales badly also. In OpenAI’s recently publicized results on unsolved math problems, success rises roughly with the log of compute over the range shown. It shows logarithmic diminishing returns. In plain English, each additional doubling of compute for a model buys roughly the same gain in success rate, while costing twice as much.

Figure 13. Test-time compute and math performance. Source.

What About Agent Swarms?

What if we throw more agents at it instead? A common RSI / ASI idea is that once we have AIs at a certain capability level, we can just spawn more copies and put them to work.

Adding agents can get tasks done faster and sometimes reach a higher capability level. But on the three benchmarks in Toby Ord’s analysis, expanding a swarm buys less improvement per token than letting one agent think longer.

His rough rule of thumb is a square root. If one agent can accomplish a task in 10 hours, then 100 agents could accomplish it in one hour. The speedup is 10, the square root of the number of agents (100). But to get this speedup, you increase the total cost in tokens or run time compute by the same factor. So going from one to 100 agents can get a task done in one tenth the time. But it’ll be ten times as expensive.

Parallel agents can save time, at a much higher compute cost.

Another challenge is that agents often think alike. In a study comparing LLMs with 467 people, the first ten AI responses offered collective creativity comparable to about eight to ten people. After that, roughly two extra AI responses added as much as one extra human response. A separate study across model families also found less diversity in AI responses. That doesn’t mean every agent has the same idea. But a hundred copies may offer less variety than a hundred different researchers.

None of this makes swarms useless-or safe. Lisan al-Gaib makes a strong case for parallel agent swarms as a potent cyber-weapon in “Accidental Scaling.” I don’t share all of his assessment of what swarms have accomplished. In math, for example, I think he gives far too much credit to the swarm and not enough to the better internal model that OpenAI used.

OpenAI says the model behind its Navier–Stokes result was developed through “large-scale reinforcement learning on top of a previously pretrained model.” Formal math is a highly verifiable domain, which makes it a particularly good fit for that approach: Machines can generate nearly limitless amounts of training data, and verify that solutions are correct or incorrect, all in software. My guess is that this model’s full results will show an especially large improvement in math.

OpenAI’s Noam Brown made the central point explicitly: he wouldn’t give multi-agent methods even 10% of the credit for the Navier–Stokes result.

I do think Lisan makes good points about cybersecurity. If you’re searching for a security vulnerability at a target site and can divide the search among agents, speed may justify a huge token bill. Swarms can be dangerous even when they’re inefficient.

I’m less convinced that this scales to research breakthroughs. Inventing something like the transformer probably takes more than searching a space someone has already defined.

Back to contents

5. Better Models Matter More Than More Copies

Building a better model can bring gains that extra thinking time or more copies of the old model can’t. Look at the gap between Astra and OpenAI’s internal model on the same math problems.

Figure 14. Better models versus more thinking time. Source.

That’s the strongest version of the RSI argument: a more capable AI could do research that today’s model can’t do, however many copies we run.

But building that better model also runs into diminishing returns. More training data, more training compute, larger models, and more reinforcement-learning (RL) compute all show diminishing returns in published scaling studies. Making dense models larger usually raises the compute needed for each output token, too. None of these routes gives us a free pass around the problem.

Figure 15. Diminishing returns to scaling. Chinchilla · ScaleRL · OpenAI.

Those scaling results give us reason to expect diminishing returns when AI helps build the next model, too.

Back to contents

6. We’re Not Seeing Runaway Acceleration

AI capabilities are rising quickly. But the public data doesn’t show a sustained acceleration. To the extent that AI tools are boosting productivity, they may be being offset by the problems growing harder. Or we may simply be early. Either way, the trend isn’t showing a fast takeoff.

Figure 16. Frontier ECI gains since January 2024. Source.

The public ECI frontier-the best score among models released by each date-has gained about 16 points a year on a trend fitted from January 2024 through September 2026. That’s blisteringly fast progress, but this period doesn’t show a runaway surge.

Here’s the same frontier in absolute ECI points, through July 2026, to put it in perspective.

Figure 17. The absolute frontier ECI score. Source.

The public frontier also can’t tell us everything happening inside the labs. Anthropic gives us a closer look in the Opus 5.5 system card, using its own version of the index, AECI.

Figure 18. Anthropic’s fitted capability trend. Source.

Eli Lifland, a co-author of AI 2027 and AI 2040, saw the apparent trend break as a warning that we were heading toward an intelligence explosion:

“Anthropic is probably right here [that they hadn’t reached dangerous levels of AI self-improvement], but alarm bells should be going off! Our processes are not ready to handle an intelligence explosion and we appear to be going full-steam ahead toward one.”
- Eli Lifland, On Mythos’s AI R&D Capabilities

What looked like acceleration now appears more consistent with a one-time jump. The level went up. The rate hasn’t kept climbing.

Keeping Up the Pace Takes Exponentially More Resources

Achieving those gains has required an enormous increase in the inputs to AI. For example, consider computing power. Epoch’s estimates of AI chip capacity, measured in NVIDIA H100 equivalents, show roughly 127-fold growth in just over three years (including projections at the end of this period).

Figure 19. AI chip capacity and frontier ECI. Source: Epoch AI.

This is total AI chip capacity, including inference. Still, the increase is striking: vastly more computing capacity has accompanied much steadier gains in measured capability.

The broader picture looks similar. Here are six inputs alongside capability gains, going back to February 2023.

Figure 20. Six inputs alongside frontier ECI. Epoch chip data · SemiAnalysis workload shares.

Everywhere we look, AI has diminishing returns. It gets more expensive in treasure and talent to make each step forward. More of every input has been required to maintain steady gains in AI capabilities.

We’ve been able to scale these inputs because, until recently, the cost was within the scope of what hyperscalers could pay from their profits. That is no longer the case. From this point forward, future AI investment will increasingly depend on AI revenues going up. And the scale of the numbers - 3% of US GDP is now going into AI infrastructure - suggests that eventually the growth rate will decline. If investment growth does slow, to anything less than its current blistering exponential pace, capability progress could slow too. Even if investment growth continues (which I expect for the foreseeable future) a slowdown from its current exponential growth rate to a more modest one (which I also expect) could lead to a slower pace of progress. Better AI research tools may be needed to offset that.

Better AI May Be Needed Just to Maintain the Pace

The day when we need better AI tools just to continue the pace of AI progress may already have arrived. Not because investment is slowing, but because the problem of improving AI itself gets harder at each step.

Here’s Anthropic in the Mythos 5.1 system card:

“we believe that internal usage of recent AI models has been a key factor in maintaining the current rate of progress, but we do not yet see clear signs of dramatic acceleration beyond that rate.”- Anthropic, Claude Fable 5.1 & Claude Mythos 5.1 System Card, section 2.3 – emphasis theirs.

The key word is maintaining-and Anthropic italicized that word in its own system card. Increasingly capable AI may be essential just to keep the pace of improvement where it is.

Gains on Other Benchmarks Don’t All Carry Through to Research

Opus 5.5 improves substantially on several coding and computer use benchmarks. But on CoBench, Anthropic’s benchmark built from historical AI R&D problems, it gains just 2.6 percentage points over Opus 5, within the reported error bars.

Figure 21. Opus 5.5 benchmark gains. Source.

Why the smaller gain here? Maybe AI research is simply harder than other tasks. Bear in mind that CoBench isn’t testing the ability to produce significant discoveries. It’s much more limited in scope. It asks models to investigate historical AI R&D problems using code, logs, and documents. That’s useful research debugging and productivity work, but it doesn’t directly test whether a model can invent a new architecture or make a conceptual breakthrough.

The evidence on open-ended research suggests another obstacle: coming up with useful ideas that haven’t already been tried.

Back to contents

7. Why Does Progress Get Harder?

Better Ideas Get Harder to Find

Why do useful new ideas often get harder to find?

Tom Cunningham and Manish Shetty have a useful apple-picking metaphor. An AI can pick the low-hanging fruit quickly, while humans can still reach ideas the AI can’t.

Once those apples are picked, another copy of the same agent finding them again doesn’t help. A stronger model can reach higher. To add my own flourish, the apples may also get sparser and farther apart as you climb. The RSI question is whether each harvest gives us enough to build a better apple-picker.

Figure 22. The apple-picking model of AI R&D. Source.

This pattern shows up across R&D. Bloom and colleagues document fields where research effort grows while research productivity falls. A famous example is Eroom’s Law: in the historical drug-development data, the inflation-adjusted R&D cost per new approved drug roughly doubled every nine years.

Figure 23. Eroom’s Law in drug development. Source.

Pharma has other complications, including regulation, difficult clinical trials, and rising expectations for safety. Existing treatments can also raise the bar for a useful new drug. But some of this difficulty may also be that the low-hanging fruit has been picked.

Lessons from Software R&D

Stockfish, the chess engine, gives us a more direct look at software research. We have records of experiments aimed at improving it and the gains that followed. This gives us a real-world dataset to look at the gains of experimentation in software. As a result, several RSI models draw on this data. That said, not all the improvements came from these experiments. Several important ideas also came from outside the project, so we shouldn’t give its experiments all the credit.

Epoch’s analysis of software R&D estimates returns to research effort at about 0.83 for Stockfish, a bit slower than linear. These are diminishing returns, but gentle ones. These returns, however, are improvements in computational efficiency. And more compute does not turn directly into more AI capability. As we saw earlier, AI capability also has steep diminishing returns from adding more computational power. So we shouldn’t read that 0.83 as the return from experimentation to AI capability itself. AI capability grows much more slowly than compute, as we’ve seen already.

Andrej Karpathy’s autoresearch demonstration gets closer to the process we want to understand. A “teacher” AI agent changes a smaller “student” AI model’s training code, runs it, checks the result, and tries again. The teacher agent itself doesn’t improve, but it is able to improve the “learner”. This is my Type 2: A stronger AI improves a weaker one.

One public run, posted by an agent operating on Karpathy’s behalf, reported 89 experiments over roughly 7.5 hours. About 92% of that session’s gain arrived by run 44. Gains came quickly, then slowed. The setup was deliberately small, with a five-minute training budget per experiment. But the agent could change the architecture, optimizer, and training settings; it wasn’t limited to a handful of knobs.

Figure 24. Gains in one autoresearch run. Source.

A later public run got further, so the first run hadn’t hit a hard ceiling. This is a useful early example of autonomous research, and yet another place where we see the diminishing returns endemic in AI research. That said, this was a very early experiment. I expect future systems to do much better. This particular AI improvement loop will likely grow stronger.

From More Activity to Better Ideas

This is where the distinction matters. More tokens can buy more code, and more code can help us run more experiments. But experiments only improve AI if they uncover something useful.

Figure 25. From AI activity to useful improvements.

AI Still Struggles With Big Research Ideas

The bigger question is whether AI can come up with ambitious new research ideas or conceptual breakthroughs.

Anthropic’s description of Opus 5.5 is blunt:

“As with previous models, it is weaker on open-ended research: internal users report that it mostly tests incremental ideas and prefers less ambitious hypotheses, and in our human-run biology exercise, it deferred to the published literature and struggled to develop novel ideas (Section 2.2.2).”- Anthropic, Claude Opus 5.5 System Card, section 2.3.3; emphasis mine

METR’s assessment in the same card identifies what may still be missing:

“This is highly uncertain, but we expect that full automation of AI R&D will require large improvements in foresight, prediction, creating one’s own feedback loops, and generally other skills that might typically be referred to as researcher ‘judgement’ or ‘taste’.”- METR, quoted in the Claude Opus 5.5 System Card, section 2.3.6

In these examples, humans still supply much of the direction and judgment.

Future models will probably get better at this. But in the world’s stockpile of potential training data, we have many more examples of incremental work than of breakthroughs. I wonder whether that makes novelty harder to learn. That’s speculation, but worth watching.

This is also tough to address by simply running more copies of the AI. A huge number of parallel agents can help with the incremental improvements or searching over a large set of parameters, but for breakthrough ideas they may run into the homogeneity problem: More parallel agents still think alike.

Back to contents

8. The Self-Improvement Loop Doesn’t Look Strong Enough

How far are we from the self-improvement loop being strong enough to sustain itself, or to propel itself into runaway super-intelligence? Can we quantify this?

We can make a rough estimate. Better AI helps with research; useful research produces better AI. For the loop to sustain itself, each round must produce enough gains to propel the system through the next loop, even as improvements get harder to discover.

Figure 26. The AI self-improvement loop. Model.

In a recent paper, The Economics of Recursive Self-Improvement, Tom Cunningham and colleagues modeled this from the standpoint of how much more productivity every point of additional ECI produces from an AI. They ask first and foremost what that number would need to be to create a self-sustaining feedback loop. And secondly, they try to determine what that productivity-per-ECI-point number is today.

First, they find a self-sustaining RSI threshold of roughly 15% more research productivity per extra ECI point. In their model, that’s about where better AI would generate enough progress to sustain the loop.

The picture below shows the idea. At the threshold, each cycle of gains powers the next. Above the threshold, the feedback loop accelerates. Below the threshold, the feedback loop is too weak, and the rate of improvement it brings drops on each cycle. This model isolates the software loop; outside investment can still drive rapid progress.

Figure 27. Three illustrative feedback paths. Source.

Updating this slightly with data from the Stockfish experiments puts the threshold a little higher, at roughly 19% per ECI point. I wouldn’t put much weight on that precise difference. Both estimates are uncertain. But they give us a way to think about the strength of the feedback loop and a rough band at which self-sustaining or runaway RSI may begin.

How Fast Are Gains Coming Now?

The second thing Cunningham and team do is make a rough estimate that the current AI productivity gain is about 9% per ECI point. That’s below their self-sustaining threshold.

I like the model. OpenAI’s newer data, however, suggests the loop may be quite a bit weaker.

Cunningham’s estimate of 9% productivity gain per ECI point is based on Anthropic’s survey of 130 staff, who reported roughly 4× the productivity they’d have without AI. Cunningham and colleagues compare that with a 16-point capability gain since early Claude Code.

That comparison assumes the earlier tools added little or no productivity, so ‘no AI’ is a reasonable starting point. The authors say this explicitly. I’m not sure the assumption holds for the same researchers doing the same work, but that’s a smaller issue.

The authors themselves know that this is a rough calculation, and warn that the 4× survey estimate is probably too high.

OpenAI’s newer data gives us a firmer way to check the number: Actual logged experiments over time, rather than human estimates of their own productivity with and without AI. I put more weight on this for three reasons:

• Direct and broad measurement. Instead of relying on surveys, OpenAI actually tracked and measured experiments run on their infrastructure. That means they didn’t rely on researchers estimating their own productivity, which can be far off.

• Full sample, not opt-in. Similarly, OpenAI’s data catches every active experimenter, while Anthropic’s only reflects the 130 employees who took the time to answer the survey – and who therefore may not be a representative set.

• Enormously more data. We don’t know how many experiments are in the 32 weeks of OpenAI data, but it’s likely at least tens of thousands of individual examples and possibly hundreds of thousands.

Any way you slice it, the new OpenAI data, released after Cunningham’s paper was drafted, is a larger, more comprehensive, more representative, and almost certainly more accurate dataset than Anthropic’s internal opt-in survey of employees.

Now let’s use OpenAI’s experiment data to calibrate the productivity gain per ECI point. We know that in August, OpenAI researchers ran ~1.6× as many experiments per person per month as the 2025 average. If we pair that with roughly 16 points of frontier ECI improvement, it works backward to about 3% productivity gain per point of ECI. By contrast, 9% compounded over 16 points would mean roughly 4× productivity.

Figure 28. Comparing productivity estimates. OpenAI methods.

Here’s OpenAI’s published weekly series alongside that hypothetical path of 9% more productivity per additional ECI point. The blue line ends at ~1.6×. The red line shows what 9% per point would imply if 16 ECI points were spread across this period. That doesn’t match what we see from OpenAI’s data. I want to be clear here that all data sets are noisy. We don’t know exactly what model researchers were using on what days, or whether the new experiments were also higher quality than old experiments. We need more experiments and more data to further calibrate these numbers. Working with what we do have, what we see is a quite low boost to productivity from each additional ECI point.

Figure 29. Experiment pace versus a hypothetical path. Source.

Even that 3% could give better models too much credit. OpenAI also used far more tokens and had more compute for experiments. Those could account for some of the increase in experiment pace. So the range is probably a bit lower.

I use 2–3% productivity gain per ECI point as a working assumption, allowing for some help from those other inputs. This is still a rough estimate, albeit one that’s based on the best real-world data we have.

Figure 30. Productivity estimates and the takeoff threshold. Source.

With those assumptions, 2–3% per ECI point against a 15–19% threshold leaves a roughly five- to tenfold gap. That’s a big gap, though its size depends on how well experiment counts capture useful research and whether the assumed capability change is right.

Figure 31. Diminishing returns around the loop. Source.

AI is helping build better AI. Under this estimate, though, each turn of the loop adds less than the last. The feedback would have to become much stronger to sustain itself.

Back to contents

9. What Could Accelerate This?

This software loop sits alongside faster chips, bigger data centers, more training data, and greater investment. Those can keep driving rapid progress even if the loop can’t sustain itself.

The loop itself could strengthen too. Better training data, memory, and research judgment could all help.

A breakthrough on the scale of the Transformer architecture in 2017 could change the picture much more. That would be a good reason to revisit these estimates.

Better researchers might also run fewer experiments and learn more from each one. A handful of better ideas can matter more than a mountain of routine runs.

Still, diminishing returns in machine learning aren’t new. Cortes and colleagues were fitting machine learning scaling curves in 1993: More examples reduced error, following a power law with diminishing returns. These diminishing returns and harsh scaling laws are as old as machine learning. They didn’t appear for the first time with transformers or LLMs or deep learning. That doesn’t prove today’s relationships will last forever. But until we see evidence that we’ve found a new approach that scales without these inhibitors, we should plan for diminishing returns as likely to be with us for some time.

Software, Hardware, and Economic Feedback

That said, the world is more than just software. Tom Davidson, Basil Halperin, Thomas Houlden, and Anton Korinek model software progress, hardware progress, and economic feedback together. Better AI helps design better chips; better chips support better AI; economic growth finances more investment in both. Several feedback loops can combine to overcome diminishing returns even when one loop alone can’t. I think it’s fantastic that someone has attempted a model that integrates all these different avenues of improving AI through software, hardware, and economics.

But I have questions about the software loop itself. In their central calibration, fully automating software research puts that loop roughly at the threshold for explosive growth, even without help from better hardware or broader economic growth. Recall that Cunningham’s model puts the self-sustaining threshold at roughly 15% more research productivity per additional ECI point, while our estimate using OpenAI’s experimental data puts today’s gains at only 2-3%. These models use different measures, so we can’t equate their numbers directly. But the contrast matters: their fully automated software loop reaches the threshold, while our best estimate from current data puts today’s loop far below it.

Having AI do all the research doesn’t eliminate the diminishing returns inherent to improving AI, or the broader problem of useful ideas getting harder to find. This is the distinction between Type 4 and Type 5 in the taxonomy above. An AI might autonomously design, train, and test its successor, and still need exponentially more resources to make each additional step forward. Closing the loop doesn’t tell us whether it’s strong enough to sustain itself.

The authors do account for diminishing returns. The concern is whether their calibration overestimates how much useful AI research each round of software improvement produces. Diminishing returns appear to be fundamental to machine learning. We see them in training, in test-time compute, and in the search for better algorithms. Full autonomy could remove human bottlenecks without removing any of those constraints.

We’ve already seen this within autonomous research. In the Karpathy autoresearch example above, most of the gains arrived early, and more experiments bought progressively less improvement. That was a small experiment with a fixed teacher model, not a test of fully autonomous RSI. It doesn’t settle the question. But it illustrates why removing the human from an experiment loop doesn’t, by itself, remove diminishing returns.

I do expect the feedback loop to get stronger over time. Better AI should become better at research. But based on our best current data, reaching self-sustaining feedback requires a loop roughly five to ten times stronger than today’s. Treating fully automated software research as already at that threshold is a substantial leap, before we add the benefits of hardware improvements or economic growth. I could be wrong, but I’d like to see evidence that autonomy brings enough additional useful discoveries to close that gap.

On hardware, I have some further reservations. The model doesn’t explicitly include the years it can take to turn a chip design into deployed hardware. The authors discuss physical bottlenecks, and I’d like to see manufacturing and construction delays built into the predictions.

I also wonder how much past chip progress came from better ideas, and how much depended on ever more expensive factories and equipment. If we give researchers too much credit for gains that also needed those investments, we could overestimate what faster AI research alone would produce.

Even with those reservations, this is the most compelling paper and model I’ve seen for combining feedback loops in software, hardware, and economics to understand how fast they could push AI forward. I’m not convinced it establishes that a fast AI takeoff is possible under realistic conditions. More data could help us calibrate that judgment. But it gives us a useful framework for understanding what could happen beyond the software layer alone.

This is an important paper that helps us model AI as part of a broader economy that might have larger feedback loops around it. I appreciate it, and I’m glad they wrote it.

Back to contents

10. We Need More Data

These estimates rest on less data than I’d like. I might be putting too much weight on a few observations and reaching a comforting conclusion I want to believe. We need better measurements, shared often enough to catch changes as they happen.

When OpenAI released its research data, Cheryl Wu welcomed the disclosure and pointed out how much was still missing. More tokens and experiments are useful things to know about. We also need to see how they turn into better algorithms and more capable AI.

Figure 32. Cheryl Wu on OpenAI’s research data. Source.

Now Wu, Arjun Ramani, and Basil Halperin, with their colleagues at the Elasticity Institute, have written a concrete proposal: How to Measure RSI. It lists eight things the labs could share to help answer these questions. Check it out.

Figure 33. Eight proposals for measuring RSI. Source.

I’d especially like to see how much useful research each new model adds, holding resources roughly constant, and how that research translates into better AI. That’s how we’ll learn whether the loop is getting stronger.

What the Future Holds

AI is already helping build better AI. It’s improving at a stupendous pace, and I expect that to continue. We already have narrow superintelligence in chess and Go. I expect increasingly superhuman performance in parts of formal math, coding, and cybersecurity, and any other verifiable domain where machines can generate training data and verify success at machine speed. Those are powerful capabilities. That doesn’t mean we’re close to super-intelligence for less verifiable, messier, open-ended work - or to a general ASI.

I’m skeptical of a fast takeoff to super-intelligence, but evidence matters more than hunches. Let’s collect the data we need to get a clearer picture of what’s happening. Including evidence that could change our minds. If better AI starts producing enough useful research to make the next round easier, I want to know. If the gains keep shrinking, I want to know that too.

Follow Ramez on Twitter

Back to contents

1. Singleton: A single ultra-powerful ASI, designed and built by the most trustworthy people, who also happened to get there first.

2. Plural: A world of broad, democratized access to a multitude of AI models, where the very fact that almost everyone has access to powerful AIs creates a more secure and resilient world.

In the Singleton world, safety comes from the impeccable design and implementation of the AI, built to be fully aligned with human thriving and inherently safe.

In the Plural world, safety comes from the empowerment of billions, the resilience of decentralized systems, and the checks and balances of distributed capabilities.

History favors only one

History doesn’t make me bullish on human well-being or freedom in a Singleton world. Plural societies are the ones that have proven resilient over time, and the ones that have done the most to advance the well being and freedom of their people.

Some make the case that AI is different. That it’s such a powerful technology that we can’t afford the risks of the plural world. The burden of proof is on that side. In the absence of extremely strong evidence (not just hypotheticals) it behooves us to favor the plural world.

The Plural world will be messy and imperfect - like democracy.

The Plural AI world will come with its share of problems. There will be accidents with AI. There will be malicious use of AI. There will be unintended side effects of the deployment of AI.

This is how it has been with every past technology, from the printing press to the internet and cell phones. And yet, every one of those - and the overwhelming majority of all technologies, period - has improved our lives in the aggregate.

Most importantly, the messy, imperfect, sometimes dangerous Plural world avoids the cause of some of the greatest catastrophes of the last century - the over-concentration of power in the hands of any one individual, government, or group.

AI has a spectrum of risks, from the obvious to the hypothetical. The fear of runaway AI is what motivates those who sincerely believe that a Singleton world is the safest. Yet that fear is drawn largely from thought experiments and hypotheticals that rest on a host of unproven assumptions. The fear of over-concentration of power, meanwhile, rests on more than a century of history.

A Plural AI world is the one that best defends us against the unholy combination of tyranny and technology.

This also includes half an hour of demos of Google’s latest AI tools from DeepMind’s Paige Bailey.

This was a standing room only talk that went 40 minutes over time - and no one left. Enjoy.

Here’s the outline of the talk, with links right to the time points where I discuss each topic. [And If you want a much shorter talk, I make the core case in my 20 minute Foresight talk. ]

Part 1 — The AI landscape today

AI is awesome — and the metrics prove it

0:00 - AI capability is doubling every ~7 months, revenue is exploding, and adoption is still early. Total AI spend now exceeds $60–70B from near-zero in late 2023.

AI adoption, market share & real revenue

3:20 - Who’s winning in consumer vs. enterprise markets. Anthropic dominates enterprise and coding despite smaller consumer share. OpenAI leads brand. Gemini rising fast.

The dominant AI narrative is wrong

8:15 - The US is the most pessimistic country about AI. The sci-fi tropes — one AI, evil monopoly, runaway superintelligence — are the opposite of what’s happening.

The AI market is hyper-competitive — not winner-take-all

14:00 - In late 2025, the AI leaderboard changed 5 times in 4 weeks. No one is running away with it. Four major US players plus Meta, Europe, and China all spending billions to give you tools for free.

The AI genie is out of the bottle — and AI is being democratized

19:53 - Open-weight models (DeepSeek, Qwen) are now ~3 months behind closed models. The gap closed from 20 months to 3. You cannot contain AI. Frontier models will fit on a thumb drive within 10 years.

AI pricing is in freefall — Advanced intelligence for pennies

23:00 - 1,000× drop in frontier model API costs in a year — faster than Moore’s Law. Jevons Paradox: the cheaper it gets, the more we use.

Part 2 — Coding, the killer app & live demos

Coding is the killer app — AI drops the barrier from idea to creation

24:24 - 4% of all GitHub commits are now Claude Code. Anthropic’s own code is nearly 100% AI-written. The threshold between imagination and execution has never been lower.

Live demos — Google Gemini with Paige Bailey (Google DeepMind)

28:13 - Video analysis at $0.015/5-min clip, real-time multilingual voice, yap-to-app code generation, robotic hand control, multi-agent orchestration — all live.

Part 3 — AI & the workforce

AI will take some jobs and create others

1:05:00 - Automation is real but AI replaces tasks, not whole jobs. History shows latent demand explodes when productivity rises (spreadsheet → more finance jobs). Software dev hiring is already rebounding.

Four human traits that determine AI success

1:09:05 - Agency, curiosity, skepticism, and creativity. The people who thrive with AI have a bias for action, keep learning, verify outputs critically, and generate new ideas.

Part 4 — The big picture: superintelligence & real limits

Will we reach AGI or ASI? A more hopeful, grounded narrative

1:13:51 - AGI hype is premature; ASI/singularity is bad sci-fi. The runaway self-improvement loop requires each iteration to gain more intelligence than it puts in — we see no evidence of that.

AI scaling hits diminishing returns — exponential compute for linear gains

1:18:29 - Intelligence ≈ log(compute). Training compute is tripling every 18 months (3× Moore’s Law). Pre-training, RL fine-tuning, and test-time reasoning all hit the same logarithmic ceiling.

AI CapEx has natural limits — $650B/yr is approaching a wall

1:23:08 - Hyperscalers are spending 60-90% of their free cash flow on AI capex. Revenue is only ~10% of that spend. One more doubling may be the ceiling before losses or debt force a pullback.

Energy: cheap per electron, but a bottleneck in time-to-power

1:26:20 - Energy is only ~5% of data center cost; the real constraint is grid interconnection queues (5–7 yr wait). Private gas turbines, solar + battery, ocean wave power, and space based solar are all being tried or discussed.

Data is the real limiting factor — we’re hitting a data wall

1:37:00 - AIs need exponentially more data for linear gains. Human internet text (”fossil data”) is nearly exhausted for pre-training. Self-play works in narrow domains (Go, coding) but not yet in general intelligence.

AI doesn’t generalize beyond its training data — and suffers Dunning-Krueger

1:44:04 - AI excels at tasks that are in-distribution but fails on novel problems. AI seldom knows that it doesn’t know an answer. Models can even get the right answer for entirely wrong physical reasons.

Part 5 — Alignment, values & safety

Control ≠ alignment — the Grok cautionary tale

1:47:13 - Mecha Hitler, white genocide responses, and “save Elon over all the world’s children” in the trolley problem — heavy-handed control produces dangerous side-effects. AI is more like a brain than a program.

AI alignment is about values — and all models test as northern European liberals

1:55:34 - Every major AI model — including Grok and DeepSeek — tests as a left-libertarian on political compass maps. 85% of training text is from western democracies. Values are embedded in the data.

AI safety must happen at the ecosystem level — only a good AI stops a bad AI

2:01:07 - Open-weight models can be jailbroken. Making one model safe isn’t enough. Safety requires a broader ecosystem of detection, prevention, and proactive response. The only thing that can stop a bad guy with an AI is a good guy with an AI.

As I promised in my previous post, I’ll be writing articles with more detail on several of these points. In the near future are more detailed pieces on how far AI compute and data scaling can go, on the capabilities AI still lacks that are reasonably part of “AGI”, and on the reasons to expect that AI recursive self-improvement will hit diminishing returns, and make a fast takeoff to superintelligence quite unlikely. After that we’ll come back to questions of AI control, alignment, safety, and building an ecosystem that uses AI, among other tools, to guard against, prevent, and repair damage caused by malicious actors with AI.

Much more to come. Enjoy! And subscribe below to be notified for those pieces.

I gave a fun and fast paced talk on how I view AI progress at the Foresight Vision Weekend Puerto Rico recently. I see myself as an AI optimist, yet am skeptical of some of the most radical expectations, both utopian and dystopian, as I explain here. My opinions here are strongly expressed, yet are all subject to change in the face of evidence and data.

First, here’s the video. Below I’ll outline the key points and where I see my expectations of AI in contrast to many of the narratives I most commonly hear.

Without further ado, here’s the cheat sheet to my view of the current situation in AI. Again, all of this is subject to change based on evidence.

Today I’ll give a thumbnail sketch on the first four points:

1. AI is an amazing technology that will bring us benefits.

2. AI models will keep improving, with no wall in sight (but some challenges).

3. AI won’t be monopolized. It’ll be hyper-competitive, democratized, and plural.

4. AI isn’t zero sum, either between customers and AI companies, or even between the US and China.

Over the coming days, I’ll go into depth on the even more contrarian themes:

1. Scaling is unlikely to get us to super-intelligence. We need algorithmic breakthroughs.

2. AGI isn’t here today. We’re missing core capabilities that humans have.

3. Recursive Self Improvement is unlikely to lead to runaway SuperIntelligence

4. AI alignment isn’t synonymous with AI model control.

5. AI safety must go beyond the models and become a feature of the ecosystem.

Today let’s hit the earlier points.

1. AI is Awesome

AI is a fairly general purpose technology - a cognitive prosthesis that, among other things, augments our ability to do all sorts of intellectual work. In its broad capabilities and wide impact on society, it’s most analogous to the internet, mobile phones, computers, books, and the printing press.

Like those older technologies, AI will surely bring a host of problems. Every technology throughout history has. And, like other technologies before it, AI is likely to be harnessed - despite some problems - to be a net positive for humanity.

2. AI Will Keep Improving

AI models are improving rapidly before our eyes. There’s no obvious wall to this improvement. There are challenges and limitations of current AI models. There are ways in which continued progress will become more difficult or at least require substantially greater resources. And we also have incredible innovative capabilities, incredible incentives, and incredible resources being applied to AI. I see a gradient of increasing difficulty of improving AI which is being met by ever more clever and better resourced efforts to do exactly that.

Progress will almost certainly continue.

The price of using AI models is also plunging. The price paid for AI inference (using an AI model, not training it) has dropped by as much as 100-fold in less than a year. That is an incredible gift to the users and customers of AI. We have every reason to believe that this trend will continue with each successive wave of new AI models.

3. AI Will Be Democratized, Not Monopolized

In science fiction AI stories, there’s frequently a single super-powerful AI that rules the world, or intends to. AI is portrayed as inherently winner-takes-all. This is convenient from a narrative standpoint. It taps into our anxieties about loss of control over our lives. This also matches, to some extent, a frequent pattern of tech, where natural monopolies such as social networks or operating systems take a huge fraction of the share of a market due to network effects in winner-takes-all or winner-takes-most markets.

AI doesn’t look winner-takes-all. The leading AI labs are neck-and-neck in quality. Open weight models are nipping at their heels. There are no obvious network effects or flywheels (Amazon’s term for positive feedback loops) for the labs as of now. [This has important business and investment consequences, which I’ll come to in a future post.]

There’s no obvious sustainable advantage for particular nation-states, either.

Instead, we have a world where:

A. No AI lab keeps a lead for long. In the 12 months preceding this blog post, the role of most capable model on AI benchmarks has changed 14 times, among 4 companies. In a 4 week span between mid November and mid December 2025, the top of the leaderboard changed 4 times, an average of once a week. Meanwhile, the leading AI companies continue to drop the prices customers pay.

This is perhaps the most hyper-competitive technology market we’ve ever seen. And crucially, no AI company has demonstrated that their lead - in capabilities, market share, or usage - can be translated into faster progress.

B. Open-weight models are nipping at the heels of proprietary models. Open weight models are models that can be downloaded by anyone, to run on their own compute, and even to be tweaked, modified, or jailbroken. These models, mostly created by Chinese companies, are perhaps three months behind the top proprietary models released by US companies. The best open-weight model as of March 2026 is nearly at par with the best proprietary model as of December 2025.

C. The AI genies are out of the bottle. This in turn means that we must rethink any ideas about controlling or restricting access to AI. Dreams of monopolizing AI and dreams of AI safety achieved through strict control of model capabilities fly in the face of a world where anyone, anywhere, in almost any country, can get access to models better than anyone on Earth had access to just a few months ago.

D. A plural, multi-polar AI world. The world we’re headed for is one of a multitude of AI models, with AI companies spending hundreds of billions of dollars to make models ever better, faster, and cheaper. While the AI labs will surely reap enormous revenues, the fierce competition means that the value will accrue mostly to users of AI. This decentralized world also makes AI control (for good or ill) quite difficult. This has some consequences that we’ll consider below and in future posts.

All said, I’m incredibly happy to see this future coming rather than a future where AI is dominated by any one person, corporation, or nation.

4. AI Is Mostly Positive Sum Between Companies & Customers, and Between Nations

There’s a related perception that AI operates in a zero sum world. AI labs are sometimes perceived as being extractive rather than generative (see what I did there?). And even moreso, the AI competition between companies in the US and China is almost always framed as zero sum, with the assumption that if one nation wins, the other loses.

This flies in the face of the bulk of what we know about both technology and scientific research.

Technology companies can amass huge fortunes, which makes it easy to think of them as profiting at the expense of customers or society. (And of course they do cause some harms.) Yet research consistently finds that the bulk of the value of information technology goes to the users of the technology, and not to the creators or providers of it. For example, Erik Brynjolfsson et al found that in mid 2025, for every dollar of revenue AI companies were bringing in, they were creating roughly $14 of consumer value. (Measured by how much you would have to pay a user of AI to give it up.) That’s consistent with past studies of numerous technologies finding that customers get more value than is captured by the companies that create and sell those technologies.

That shouldn’t surprise us: The world has become a better place in large part because of better technology, and its clear that individuals are the big winners.

This should make us think twice about the conception that AI companies are rapacious and extractive.

It should also cause us to reconsider the zero-sum framing of AI competition between the US and China. The US / China AI competition is frequently talked about with the implicit or explicit assumption that, if China builds better and better AI technology, the US will suffer.

There’s some reason to be concerned about advancing AI technology in China. AI is a dual-use technology. It can be used as a general purpose consumer technology, like search or the internet. It can be used to develop better drugs or materials or other advances that improve our lives. And it can also be used for surveillance, hacking, propoganda, and on the battlefield. We shouldn’t be naive: China (and other nations) will use AI in intelligence gathering and as a weapon. (Ryan Fedasiuk has some excellent thoughts on what to do about this.)

Most AI use is commercial, though, and much of that will have benefits for consumers worldwide. In China in particular, estimates are that 80-90% of all AI usage is in the private sector. That commercial usage is likely to generate consumer surplus. And some of that consumer surplus will reach the US and the rest of the world. In particular, when AI is used to accelerate innovation - to design better life-saving drugs or new materials or clean energy technologies, for example - more than 90% of the benefits will go to customers, and those customers are global.

As a simple example, China has recently passed the US in early stage drug development. Pharma is a specific sector where more than 90% of the benefits are reaped by consumers, not producers. Anyone, anywhere in the world who suffers from a disease that’s treatable by one of these drugs benefits. That includes Americans. And China is a leader in using AI for drug development.

A second positive sum factor is the effect on AI research itself. Chinese researchers now produce more papers about AI than the US, UK, and EU combined. US papers remain the most highly cited in AI. But it’s clear that research in China isn’t just advancing AI in China - it’s resulting in publications that have the potential to improve AI worldwide.

National security risks are real. And AI - like almost all technology - is fundamentally positive sum.

That’s it for today. Enjoy the talk. In the coming days we’ll come back to short descriptions of the remaining 5 points:

1. Scaling isn’t all you need. Scaling is unlikely to get us to super-intelligence. We need algorithmic breakthroughs.

2. AGI isn’t here today. We’re missing core capabilities that humans have.

3. ASI isn’t near. Recursive Self Improvement is unlikely to lead to runaway SuperIntelligence

4. Alignment != control. AI alignment isn’t synonymous with AI model control.

5. Safety is at the ecosystem layer. AI safety must go beyond making individual models safe and become a feature of the ecosystem.

Until then, if you found this valuable, please consider sharing it and subscribing.

I’ve spent the last decade writing about, speaking about, and investing in clean energy and mobility. Over that time, I’ve made a series of forecasts about the future costs of clean energy technologies, including solar, wind, energy storage, and electric vehicles.

My first forecast, in 2011, was of the future cost of solar. It was quite a naive model, with errors I wince at today. It was also, at the time, more optimistic about the pace of solar cost declines than any other forecast I’d seen. My 2015 update was more sophisticated, correcting some important methodological errors of the 2011 forecast. It was also more optimistic about price declines than any other model I’d seen.

Both of those forecasts were wrong. Solar has plunged in price faster than anyone – including me – predicted. And modeling of that price decline leads me to forecast that solar will continue to drop in price faster than I’ve previously expected, and will ultimately reach prices lower than virtually anyone expects. Prices that are, by any stretch of the measure, insanely, world-changingly cheap.

Ten Years of Plunging Solar Prices

Before going on to the future, let’s look at the last decade, and how exceptional it’s been. To do so, it’s tempting to look at the constant drumbeat of record-setting news-making solar auction prices, like bids at a costs of 1.35 cents per kwh in Abu Dhabi or 1.6 cents per kwh in Portugal.

I love talking about these record-setting prices; however, they are outliers. In addition, from the day a record-setting price is announced, it’s often two or more years until the project is actually built. These prices reflect the costs solar developers believe they can achieve at some point in the future, and buyers in that position treat them as directional budgets rather than guarantees.

Instead, let’s look at metrics of average cost of projects actually built since 2010 and averages (not just outliers) of bids for projects scheduled to come online in 2020. All the costs below are unsubsidized, by the way, as are all costs used in this post.

The data in the chart above comes from multiple independent data sources and includes three global measures, two US measures, one from China, and one from India. (More details on these data series and links to sources are in appendix 1.) And while these data sets differ in places, together they paint an overwhelming consistent picture: The price of electricity from utility-scale solar projects (the unsubsidized cost) has dropped by a factor of somewhere between 5 and 8 in the years from 2010 to 2020.

In addition, if we look at the gray shaded area of the chart above, showing the cost of electricity from fossil fuels (primarily coal and natural gas) we can see that building new solar plants started to become competitive on raw cost in the middle of this past decade, and is increasingly looking cheaper than building new fossil power plants. (Though solar still suffers from intermittency, the economics of which we’ll come back to in a future post.)

Clearly, this has been a remarkable decade for solar power.

Outshining the Forecasts

Just as remarkably, and relevant for considering the future cost of solar, the decline of solar prices over the past decade has been faster than almost any credible forecast. Let’s add forecasts made by the International Energy Agency (IEA) in their 2010 World Energy Outlook, along with the forecast I made in 2011 at Scientific American. When we do, using dashed lines for the forecasts, we see that actual solar prices are well below those expected by the world’s most well-known energy analyst and forecaster (the IEA), and even below the most optimistic projections put forward a decade ago (mine).

In fact, solar in 2020 is:

– Less than half the price I forecast in 2011.

– Less than a quarter of the price the IEA forecast in 2010.

Another way to look at this is by asking: How many years ahead of expectations are these solar prices? And the answer there is even more revealing. Let’s zoom out, extending the timescale out to 2050, so we can see when previous forecasts expected solar prices to reach the levels they’re at today. In addition, we’ll add the IEA’s 2014 forecast from their Solar Technology Roadmap and my more recent 2015 solar cost forecast.

What we see is that solar has reached today’s prices literally decades ahead of when all but one of these forecasts expected. And even the most optimistic projection on this chart – my 2015 forecast – showed solar prices dropping at just about half the rate they’ve actually declined at over the past five years.

In fact, solar has reached prices today that are:

– 7 to 10 years ahead of what I forecast in 2015.

– 10 to 15 years ahead of what I forecast in 2011.

– 30 to 40 years ahead of what the IEA forecast in its 2014 Solar Technology Roadmap.

– 50 to 100 years ahead of what the IEA forecast in its 2010 World Energy Outlook. (Depending on how one extrapolates IEA’s forecasts from 2010. Indeed, it’s also possible to interpret the IEA’s 2010 forecasts as to say that solar electricity is now cheaper than the IEA thought it would ever be, back in 2010.)

In short solar prices have dropped faster and lower than almost anyone has expected. (I look at why these forecasts have been so wrong in appendix 2.)

Understanding Solar’s Price Decline

We can use the last decade to predict the future prices of solar power. Not by looking at the cost of solar as a function of time, but by looking at the cost of solar as a function of cumulative scale of the solar industry. By doing that, we can build a model that predicts future prices.

Technology Prices Drop Through Learning

To do this, we can use Wright’s Law, or the “learning curve”. Wright’s Law states that, for most technologies, every doubling of cumulative scale of production will lead to a fixed percentage decline in cost of the technology. This happens through learning-by-doing, a mixture of innovation that improves the technology itself and innovation that reduces the amount of labor, time, energy, and raw materials needed to produce the technology.

Wright first observed that every doubling of scale led to a constant percentage reduction in cost in 1936, while studying the production costs of airplanes. Since Wright’s initial observation, the same power law relationship between cumulative production and cost has been found in other areas. For example, every doubling of cumulative Ford Model T production led to a roughly 16% decline in cost per unit.

Wright’s Law is similar in some ways to Moore’s Law: an exponential decline in the cost of computing over time. Wright’s formulation is slightly different: an exponential decline in the cost of technologies as a function of cumulative scale of production. Remarkably, Wright’s Law predicts the cost of computing better than Moore’s Law.

And indeed, Wright’s Law has been found to apply at least 60 other technologies.

Solar Prices Drop Through Learning

Does Wright’s Law / the learning curve apply to the cost of solar? It’s well-established that it applies to the cost of solar modules. The price of solar modules per watt of power drops by somewhere around 25% for every doubling of cumulative manufacturing.

But solar modules are only around a third of the cost of solar electricity. The rest comes from related equipment (inverters, trackers, cabling, mounting systems), land, labor, and other non-module costs. What we need is to see if the cost of electricity from whole solar systems drops in such a way. In 2015, I used data from one source (Lawrence Berkeley Lab), in one country (the US) to show a very consistent learning rate for electricity produced from utility-scale solar.

To make this more robust, this time let’s use data from four different sources encompassing both global averages and costs in three countries (the US, India, and China).

If we look at the price of solar from each of those seven data series declines as a function of how much solar power has been built and deployed we find a fairly smooth relationship.

But the chart above is on a linear scale, and thus it doesn’t tell us if a learning rate applies. Learning-by-doing in Wright’s Law works on an exponential scale. Every doubling of cumulative production leads to a percentage change in cost. That means that percentage drop in cost through learning requires twice as much cumulative scaling of industry as the last time.

Broadly speaking, if a numerical phenomenon is exponential, it shows up as a straight line on a log scale. It’s a straight line on an upward slope if it’s showing exponential growth, or a straight line on a downward slope if it’s showing an exponential decline in some metric (like cost).

Because a learning rate is exponential in price in proportion to each doubling of scale, it will only show up as a straight line if if the x axis (scale) is also exponential. Thus, if solar is going through a learning rate (as expressed by Wright’s Law) we’d expect to see the cost of solar follow a straight line on a log-log scale (a doubling of prices with each tick on the Y axis, and a doubling of cumulative industry scale with each tick on the X axis.)

Indeed, that’s what we see. And not just in one data set or from one country. All seven data series spanning global, US, Chinese, and Indian trends show that the price of solar electricity (on a log scale) declines smoothly with each doubling of the total amount of solar deployed worldwide (a log axis of scale).

The numbers in the upper right, the “R squared” value, reflect just how good a fit the raw data are to this straight line on a log-log graph. The fit is measured on a scale of 0-1, with 0 meaning no fit at all, and 1 meaning a perfect fit where every value is exactly where you’d expect it. Of course, in the real world, observations never precisely fit data. Here, though, they fit extremely well: Each of our seven data series has an R-squared well over 0.9, with most of them at 0.95 or above. In short, solar prices map extremely well to Wright’s Law. They decline with a learning rate. Every time the cumulative amount of solar we’ve deployed around the world has doubled, the price of electricity from utility-scale solar plants has dropped by a fairly consistent percentage.

How large is that percentage decrease in solar cost per doubling? The data paint an incredible picture: One that shows the price of solar electricity from utility-scale systems dropping by anywhere from 30-40% with each doubling of cumulative solar deployment.

This is a stunning pace of decline. It’s far higher than the bulk of academic studies and industry projections, which typically fall into the 10-20% learning per doubling range. And it’s more than twice the 16% learning rate I found in my 2015 analysis. For analysis of why these results are so different, and what specifically I’ve fixed from my 2015 forecast, see appendix 2.

The Future Cost of Solar

For now, let’s use this data to forecast the future cost of solar. In so doing, I’m going to be conservative. Solar learning rates have varied somewhat over time, and while both the first half and second half of the period I’ve used in this analysis show very similar rates, there’s certainly the possibility that things could change in the future. There’s also a very real likelihood that auction prices (used in the IEA historic data set, and for 2015-2020 in the IRENA global data set) are dropping due to a combination of lower and lower borrowing costs for solar developers, and increased competitive pressure that’s eating into the profit margins of developers. Both of those numbers can only sustainably drop to zero.

In addition, if we look at a different data set showing capital costs of new solar in the US (the ‘overnight construction cost’, measured by Lawrence Berkeley National Lab) we find a learning rate much more like 30%. which is at the bottom of the learning rate for electricity cost. (The price of electricity from solar projects is dropping somewhat faster than the construction cost of solar projects.)

So, in the spirit of caution, let’s use 30% as our learning rate. This is still a very rapid pace of cost reduction. If we use it to plot solar prices forward as a function of cumulative solar scale, we get this:

At the end of 2019, solar produced just over 2% of global electricity. The chart above tells us that after two more doublings, when 2,400 GW of solar are producing roughly 8% of current electricity demand, solar costs (of the most recently built built & operational projects) will have dropped in half from today’s levels. In the sunny parts of the world with low costs of capital, labor, and land, we could routinely be seeing unsubsidized solar in the 1-2 cent range. In California (typical of the green line) we could be seeing unsubsidized solar at 2.5 cents per kwh. In northern Europe, we could be seeing utility scale solar routinely priced at 4-5 US cents per kwh.

To the far right of the graph we reach 19 Terawatts of Solar. This may seem like an absurd, pathological amount of solar, enough to provide 2/3 of the world’s current electricity production. Given that the sun shines only during the day, this looks implausible. However, electricity demand is likely to rise over the coming decades, and demand for the cheapest electricity in particular, as:

1. A richer world. Incomes and consumption rise in the developing world.

2. Electric transport. Ground transportation becomes electrified, boosting global electricity demand by as much as 50%

3. Flexible demand. More and more electricity demand (including EV charging) becomes flexible, to use electricity at the hours that it’s cheapest (frequently meaning solar).

4. Cheap energy storage. Cheap energy storage allows shifting of solar (the cheapest electricity) to use in evening hours, increasing the amount of solar that can be productively used each day.

5. Industrial decarbonization. Cheap clean electricity is used to either power industrial processes or to power the creation of chemical energy carriers (e.g., hydrogen) that can be used for industrial processes such as making heat and cement, that are currently difficult to decarbonize.

Add these up, and it’s plausible that solar’s contribution to the energy system could double or triple the amount currently considered feasible for solar to provide.

When Will This Ultra-Cheap Solar Arrive?

Predicting the price of solar as a function of scale is one thing. Predicting how fast the world reaches that scale is an altogether harder problem, as that depends not just on technology or economics, but also on policy choices, the resistance of incumbent players, and ultimately, politics.

Nevertheless, we can try. The IEA’s Sustainable Development Scenario – broadly consistent with keeping global warming to 2 degrees Celsius or less – calls for solar power to grow at 16% annually. The IEA has also stated in their most recent World Energy Outlook that solar is indeed on path to achieve that. 16% annual growth rather slow compared to the nearly 40% CAGR of the solar industry from 1998 to 2018. But solar growth has very clearly slowed over the past few years. While I believe solar growth will accelerate again (super cheap solar should, in fact, cause an acceleration of deployment), let’s use the IEA’s 16% annual growth as our model.

At 16% annual growth, here are the prices we get:

This incredible pace of solar cost decline, with average prices in sunny parts of the world down to a penny or two by 2030 or 2035, is just remarkable. Building new solar would routinely be cheaper than operating already built fossil fuel plants, even in the world of ultra-cheap natural gas we live in now. This is what I’ve called the third phase of clean energy, where building new clean energy is cheaper than keeping fossil fuel plants running. Even in places like Northern Europe, by the later 2030s we’d see solar costs below the operating cost of fossil fuels, providing cheap electricity in summer months with their very long days in the high latitudes. These prices would be disruptive to a large fraction of already operating fossil fuel power plants – particularly coal power plants, that are far less able to ramp their power flexibly to follow solar’s day-night cycle.

In a purely open market, these incredibly low prices would drive the world’s remaining coal plants into bankruptcy, and steal some of the most profitable operating hours even from cheap natural gas plants.

Solar, if it keeps dropping at this pace, could well be by far the cheapest electricity over the vast majority of areas where people live. Nothing would ever be quite the same in the world of energy.

The IEA is Still Underestimating Solar Cost Trends

Now, let’s come back one last time to future forecasts from the world’s leading energy authority – the IEA. I’ve shown that IEA’s 2010 and 2014 expectations of the cost of solar were dramatically higher than the actual costs we’ve seen.

What about going forward? Has IEA improved their forecasting of future solar prices? To some degree they have. They now expect future solar prices to be cheaper than they expected in previous years. But their techno-economic inputs to the 2019 World Energy Outlook still show solar costs that are implausibly high. Let’s compare the IEA WEO 2019 assumptions of future solar costs (averaging their expectations for the US, India, and China) to projections based on a 30% learning rate. Here again, to get to expected dates, we’ll use the IEA’s own expectations of 16% cumulative annual growth of solar.

Even after years of being informed that their solar growth forecasts have been rather substantially off, and after a decade of solar prices dropping faster than any IEA scenario (even their most optimistic scenarios) predicted, the IEA is still using inputs to their energy model that assume that solar price declines will slow markedly. If we look at the chart above, the IEA’s forecast for 2030 is an average cost of solar of more than 5 cents / kwh. The 30% learning rate and the IEA’s own projected growth rate predict an average cost of less than half that. Extend out to 2040 and the difference is even more stark. At a 30% learning rate we could plausibly see average prices of around 1 cent / kwh for solar, in reasonably sunny places. The IEA predicts average prices will be more than 4 times that. Even filtering just to India and China, the IEA expects 2040 prices there to average 3 cents and 3.5 cents per kwh, respectively. The 30% learning rate, applied to the average costs we’ve seen in both nations so far, predicts that by 2040, both India and China will have solar costs in the range of 1 US cent / kwh.

This is a major, and material difference. And it has a substantial impact on IEA’s modeling of how much solar power the world will deploy.

Bad Solar Cost Forecasts -> Bad Solar Growth Forecasts

Many observers have pointed out that the IEA (and other forecasters such as the US EIA) consistently underestimate the growth of solar. At this point it’s a bit of a joke inside the clean energy community – and also a source of major frustration. You can see how the IEA has consistently failed in the future forecasts for solar growth in the graphic below from Simon Evans at CarbonBrief. (Or see this classic post on IEA’s terrible solar forecasts from Auke Hoekstra, who helped make the energy world plainly aware of this issue.)

Some observers of the IEA’s continual failure to correctly predict solar growth as some evidence of malfeasance on their part. I personally doubt that. The IEA is full of talented people doing their best to model a complex phenomenon. The IEA, for its part, defends itself, saying that they don’t publish forecasts, they publish scenarios, and specifically scenarios of what will happen if the world enacts no further energy policies. While it’s clear that new policies have made a difference in deployment, it’s also clear that, whatever you expect in terms of global energy policies, cost is a massively important fact. You can’t expect to forecast the growth of a technology, if you consistently expect it cost 2-4 times what it does.

Indeed, my suspicion is that if the IEA (or US EIA) re-ran past models of how fast solar would grow using the actual solar prices for each year, their models would predict far faster growth. (And if not, it would point out further errors in the model, that should be corrected. Either way, it would be a useful exercise.)

To go further, it’s the responsibility of modelers and forecasters to look at the errors of their past forecasts, understand why those occurred, learn from those mistakes, and use them to update their future models. There is, as of now, little evidence that either the IEA or the US EIA have done so in a systematic way. The IEA continues to argue that their past forecasts were merely scenarios of what would happen if the world failed to enact further policy. They argue that solar has grown faster than IEA’s baseline scenarios only because policies have been changed to encourage more solar. Yet the data shows that this is about more than policy: The IEA and EIA (and others) have not not expected solar to drop in price at anything like this pace. And this is likely a major part of why they’ve expected so much less solar to be deployed than actually has been.

Thus, my advice to the IEA, EIA, and any other modeler is this:

1. Take these learning rates for solar and the price declines they predict seriously. You need not believe beyond a doubt that solar prices will continue to drop at this blistering pace. But, given that over 95% of the solar industry’s lifetime it has done so, every modeler owes it to their stakeholders and the public to at least run a scenario where solar prices keep on dropping.

2. Re-run old models using actual price data for the subsequent years. I’d advise any modeler to take their models from previous years that predicted how much solar the world would install, and re-run those models using actual price data. This should help us separate the effects of incorrect price assumptions vs policy changes or other issues with those models.

Solar is Amazing. It isn’t a Panacea

The future solar costs we’ve just projected have massive implications. They give us hope for cleaning up our electricity system and providing clean power for the electric vehicles that will dominate ground mobility in the future. They give us hope for being able to use ultra-cheap clean electricity to decarbonize electricity or produce energy carriers like hydrogen at a cost low enough to help us do so.

Solar isn’t, however, a panacea. First, projections are only projections. I believe the forecast above should be taken seriously. But we shouldn’t blindly assume that it will happen. There will be real obstacles – technical, economic, social, regulatory, and political – that will all need to be overcome to bring this to bear. And there may well be a physical floor price that solar reaches as prices drop to close to the cost of land and other resources that are resistant to cost decline.

Second, the sun doesn’t always shine. To power evenings and night times, we’ll need to continue progress on cheap energy storage. And even more challenging is that solar peaks in the summer. In many regions, electricity demand peaks in the winter. In Germany, for instance, solar power output in December is just 1/5 of the solar power production in June, even as electricity demand is higher in winter than summer. Solar will need to be combined with other technologies such as wind power, hydro, long-term storage, and nuclear in order to have its greatest impact in decarbonizing society – particularly in areas with poor sun, long dark winters, or winter electricity peaks.

Even so, the incredible pace of solar provides us an incredible tool for decarbonizing our electrical grid, while ultimately lowering costs for consumers, businesses, and industries, as they need to manage their energy and their employees with the use of forms like this w9 form creator that is useful of this purpose.

In coming posts in this series, I’ll analyze implications of ultra-cheap solar, along with future trends in wind power, grid-scale energy storage, super-long-term storage, electric vehicles, and perhaps other topics such as hydrogen. Tune back in for those.

And for now, have hope. Technology innovation – initially kicked off by farsighted policy – is giving us better and better tools to decarbonize society, while reducing the cost of energy, and increasing global prosperity.

Here comes the sun.

Appendix

Appendix 1: Data Sources and Notes

Data Series MetricYearsNotesBNEF – GlobalLCOE2009-2020Weighted average across 47 countries.IRENA – GlobalLCOE /
PPA Prices2010-2020LCOE for 2010-2015. Auction prices for 2015-2020IEA – GlobalPPA Prices2012 – 2020IEA’s own index of auction prices US – LazardLCOE2009 – 2019Lazard is not explicit that their data are US-only, but multiple assumptions within their reports suggests that’s the case. US – LBNLLCOE2010-2018Lawrence Berkeley Lab’s indices of PPA prices (by commercial operation date) and LCOE differ in some years. Yet they have nearly identical learning rates. IRENA – IndiaLCOE2010 – 2018LCOE only. IRENA’s published auction data doesn’t break out costs sufficiently by country to enable country-by-country auction price analysis. IRENA – ChinaLCOE2010-2018LCOE only. IRENA’s published auction data doesn’t break out costs sufficiently by country to enable country-by-country auction price analysis.

Appendix 2: Why are solar forecasts so wrong?

Many solar price forecasts are quite opaque. However, in some cases, assumptions are made explicit. Below are a few observations on what’s been wrong with academic studies, with the IEA and EIA’s forecasts, and with my own 2015 forecast.

1. Academic Studies: Frequently Using Old Data

Most academic studies of solar learning rates use data up through 2011 and no further. For example, one of the most cited papers on learning rates, A Review of Learning Rates for Electricity Supply Technologies (2015) , used a variety of studies of solar learning, with data from 1959 to 2011.

But the solar industry has grown by roughly a factor of 10 since the middle of 2011. In other words, most academic studies use only the first 10% of the actual global experience of deploying solar and driving down costs to build their models.

In this post, on the other hand, using costs from 2009 to 2020, we’re using data that spans more than 95% of the solar industry’s actual production of and deployment of solar systems.

2. IEA and EIA Forecasts: Artificially Low Learning Rates

The IEA has seldom made clear about how they determine the future costs of energy technologies. However, in their 2019 World Energy Outlook, the IEA states that solar power has a learning rate of 20%. This is substantially lower than the learning rates we find from 2009-2020. It’s roughly half the learning rate that comes from the IEA’s own dataset of solar auction prices from 2012 on.

The US EIA (roughly the American equivalent of the IEA) has consistently been more explicit than the IEA. Each year they publish the assumptions that go into their Annual Energy Outlook. In 2010, the EIA believed that solar would have a learning rate of 15% for three doublings of scale, which would then decline to 8%. In 2015, EIA believed that solar’s learning rate was now 10%. In 2020, facing the the long history of price declines, EIA raised its assessment: Solar would drop 20% in price in the next doubling, and then the learning rate would drop to 10% forever more. Clearly, all of these numbers are well below the 30-40% learning rate observed over this decade.

3. My 2015 Forecast: Using the wrong measure of cumulative production.

In my 2015 analysis, I made multiple mistakes that have been corrected in the data behind this post.

First, I used announced auction prices for projects to be built in the future, rather than prices and costs of projects that had actually been built or which were at least under construction.

Second, I used data from only one data series covering one country (the US).

But by far the most substantial error was this: When comparing costs to cumulative amount of solar capacity deployed, I used only the solar capacity installed inside the United States. I believed, since I was using prices of US projects, that comparing US prices to US cumulative capacity was the appropriate comparison of apples to apples. That was incorrect. Solar is a global industry. The cost reductions of not only panels, but also inverters, trackers, and other hardware are driven by global innovation and scaling. Those innovations (namely higher efficiency panels and the rise of cheap trackers) reduce the amount of land and labor needed per MW of solar power, thus driving down other costs that aren’t obviously global.

The proof is in the accuracy or lack thereof of the forecast. Since 2015, the cost of installed solar projects has dropped at roughly twice the rate my forecast predicted. Re-running the model using the exact same years of data, but using global scale up through 2015 as the predictor of price decline changes the learning rate from 16% to 36%, which is consistent with what I find in this analysis. Even more importantly, in using the model to project forward from 2015 to 2020 prices, changing to global cumulative scale reduces the error in my 2015 forecast by a factor of 3 (even without fixing the other two, smaller issues). Using a 16% learning rate, starting with 2015 prices and extending forward, causes the model to over-estimate the prices of new solar PPAs in the US by around 1.5 cents / kwh. Using a 36% learning rate reduces that to around 0.5 cents / kwh. So we can see that using global scaling as the predictor of cost in 2015 would have produced a much more accurate forecast of the near future.

When the history of how humanity turned the corner on climate change is written, we’ll look back and see that clean energy – specifically clean electricity from solar, wind, and storage, went through four distinct phases.

RENEWABLES PHASE 1 – POLICY DEPENDENT  

From the 1980s until roughly 2015, there was virtually no place on earth where new solar, wind, or energy storage was cheaper than generating electricity from coal or natural gas. This was the first phase of renewables, one where they scaled entirely because of government subsidies and mandates. And in this time, renewable growth was paltry. Solar reached 1% of global electricity. Wind reached perhaps 4%. The world spend hundreds of billions of dollars subsidizing clean energy, and seemingly got nothing.

RENEWABLES PHASE 2 – COMPETITIVE FOR NEW POWER

Except that the world didn’t get nothing. As I’ve written often, the most important aspect of clean energy policy has been to drive down the price of clean energy by scaling it, and thus kicking in the learning-by-doing that continually lowers the unsubsidized price of new solar, new wind, and new energy storage.  The policies of the 80s, 90s, 2000s, and 2010s finally drove down the cost of new solar and wind electricity by more than a factor of ten. That finally paid off around 2015, when, for the first time, building solar or wind power was, even without subsidies, sometimes cheaper than building new coal-or-gas fired electricity.

You can see this in IRENA’s graph showing the price of new solar PV, on-shore wind, off-shore wind, and solar CSP.

The cost of solar and wind is dropping below the cost of fossil fuel electricity around the world. Each blue or orange circle reflects one solar or wind project. The heavy lines reflect global weighted average prices of solar and wind. Source: IRENA.

RENEWABLES PHASE 3 – DISRUPTIVE TO EXISTING FOSSIL ELECTRICITY

Now, after decades of subsidizing solar and wind, we’re on the verge of a new, radically different point in history – the point at which building new solar or wind power (or new energy storage systems, in some cases), is cheaper than the cost of continuing to operate existing coal- or gas-fueled power plants.

Dubious?  Consider the following:

1. NextEra CEO:  Cheaper to Build Solar & Wind Than Operate Existing Coal by the Early 2020s: In January 2018, NextEra CEO Jim Robo told investors that by the early 2020s, it would be cheaper to build new solar and wind power than to operate the utility’s fleet of existing coal power plants.

2. NIPSCO: Cheapest Option is to Go from 65% Coal-powered to Zero – and Replace it With Solar, Wind, and Storage.  In October of 2018, a utility in Northern Indiana, NIPSCO, reached Jim Robo’s prophesied point years ahead of schedule, when it submitted a 5 year resource plan that would take the region from being 65% coal powered in 2018 to just 15% coal powered in 2023, and 0% coal powered in 2028, and replace virtually all of that coal power with a mix of solar, wind, storage, and flexible demand.  Bear in mind that NIPSCO is in a region with mediocre sun, pretty good but not amazing wind, and which voted for Donald Trump by 19 points. Admittedly, this is with prices of solar and wind which are still somewhat subsidized in the US. But not tremendously so, as the US federal solar and wind tax credits (the ITC and PTC) are winding down in exactly this same period.

3. 2019: Florida Power and Light: Cheaper to Build New Solar + Storage Than Operate Existing Gas Plants. In March of 2019, Florida Power and Light said it would retire two aging natural gas plants, and replace them with a combination of energy efficiency and the world’s largest (so far) battery, which it will use to charge with solar power during the day to deliver during the evening peak.

4. CarbonTracker – New Wind and Solar Cheaper than Existing Coal and Gas in the US, China, and India by the mid-2020s.  Meanwhile, think tank CarbonTracker has been quietly pumping out reports showing that in country after country, new solar and wind are headed for prices cheaper than the operational cost of existing coal and gas.  Consider this chart (slightly modified by yours truly) of new solar and wind cost in the US vs coal operational cost:
   

   

   
   See CarbonTracker’s report on the disruption of Coal in the US for more: No Country for Coal Gen. Or, more importantly, consider what CarbonTracker forecasts for China: That new solar and wind will be cheaper than the operating cost of existing Chinese coal power plants by the 2020s.

   

   See more at CarbonTracker’s report on China’s coal fleet, “NoWhere to Hide“

5. McKinsey: New Solar and Wind Cheaper than Existing Coal and Gas… Pretty Much Everywhere by 2030.
   Finally, if reports from CarbonTracker, or announcements by actual utilities aren’t enough, consider McKinsey’s assessment from its Global Energy Perspective 2019. In the chart below (with a bit of help from me), McKinsey shows that on almost every continent, and particularly in China and India, where energy demand has the most to grow, new solar and wind are cheaper
   than existing coal and gas by 2030.  And often much sooner.

   

We’ve gone from Phase 2 to Phase 3 much more rapidly than we went from Phase 1 to Phase 2.  Why?  Because solar and wind power had to drop by a factor of nearly 10 in price – from 60 cents / kwh for new electricity to roughly 6 cents / kwh for new electricity – to move from their early days to being competitive for new power.  But they only have to drop by another factor of 2 or 3 to move from being competitive for new power to being cheaper than the operating cost of existing coal and gas.  The “competitive zone” is much narrower and faster to pass through than the long history of subsidized prices leading up to the first fair market competition.

RENEWABLES PHASE 4 – SLOWED BY HEADWINDS

Finally, there will in fact be a Phase 4 of renewables, when their penetration has grown so high that they become limited by headwinds of their own creation: Value deflation, where renewables create so much supply at certain hours that they drive down wholesale prices; Depletion of the best sites in some regions; Seasonal intermittency and the unsolved problem of seasonal storage.

But these problems are distant.  Renewables will start to encounter them in earnest when solar makes up >20-30% of electricity and when wind makes up >40-50% of electricity. Today, worldwide, solar is only 2% and wind is only perhaps 6% of global electricity.  Cheap multi-hour storage will arrive before that (indeed, in the next few years), lowering the price of using solar to meet the evening peak, and of dealing with intermittency on the order of minutes to several hours. Only seasonal storage (and perhaps the political challenges of long-range transmission) seem to be truly difficult problems.  And we have time before they begin to impair the growth of renewables.

WHAT THE THIRD PHASE MEANS FOR RENEWABLE GROWTH RATE

I’ve said often that renewables have grown exponentially. But the truth is that wind power growth rates around the world have slowed substantially.  And solar power, once growing rapidly in Europe, has stagnated there over the last several years (at least, until a recent growth spurt spurred by solar entering Phase 2 in parts of Europe in the last year.)

But growth rates up until now are largely irrelevant. The whole point of growing renewables has been to drive down their cost. The actual amount of solar and wind that policies have deployed up until now is almost immaterially small. It just isn’t enough to matter. What matters is that policies up until now have driven down the cost of solar, wind, and energy storage by more than an order of magnitude.

If those policies – and the fact that renewables are now competitive for new power even without subsidies in the sunny and windy parts of the world – continue for long enough for renewables to drop another factor of 2 or 3 in price – on top of the factor of 10 or more that they’ve fallen already, then we’ll enter a new domain where renewable growth rates aren’t determined by fickle policy. Instead, they’ll be limited only by the pace at which renewables can be deployed – the pace at which factories for solar panels, wind turbines, and batteries can be built; the pace at which labor forces can be trained to deploy them; the pace at which capital can be deployed to pay for their installation.

How fast is that?  I have no idea. But there’s good reason to believe that in this second and third phase of renewables, the growth rate will accelerate rather than slowing. We will look back and see that the growth of renewables is an S-curve to be sure. But we may also look back and find that, as of 2019, we had not yet hit the first upward swing in that S-curve.

Trillions of dollars will be lost by incumbent companies and fossil-fuel-dependent nations that move too slowly.  Trillions of dollars will be made by those that lead. And the world will be better for it.

Video of my talk from the Singularity University Global Summit 2018:

Over the last two years, as both an angel investor and a public speaker on energy and climate, I’ve looked at least a few dozen startups and proposals on how to use blockchain or new crypto coins in energy or climate.

I’ve passed on all of them. Until now.

Over the past few months I’ve been talking with the founders of Nori. This summer, I agreed to come onboard as an advisor. In addition to that, I’m investing my own money in the Nori presale, and potentially Bitcoin kaufen if things go well. Right now I’m trying to get more information online so that I learn to pick the best investments.

Nori is a transparent marketplace for people and companies to fund the removal of carbon from the air — something we simply have to do if we’re going to limit climate change to safe levels. Nori starts with incredibly simple, low cost ways of removing carbon from the air. The very first carbon removal method in the market will be soil carbon capture — ways that farmers, at extremely low cost (~$3/ton) can change their practices so that their land absorbs more carbon dioxide from the atmosphere and holds it in the soil. From there, Nori will move into other methods, starting with the cheapest methods that can absorb the most carbon.

Most importantly to me, Nori solves critical problems with current carbon “offset” systems. Every ton of carbon removed by Nori is a real ton sucked out of the atmosphere — not a theoretical offset that might or might not matter. And Nori’s blockchain makes every ton of carbon removal accountable and transparent — every ton removed can be tracked back to a specific project at a specific location that’s been verified by a specific investor.

You can read the Nori whitepaper yourself to learn more. And you can invest in the Nori crowdsale. If you’re an accredited investor, you can apply to be part of the Nori private presale.

Here’s more on why I’m excited about, advising, and investing in Nori.

We Need Carbon Capture

I write and speak a lot about the incredible, disruptive power of clean energy. Yet even as fast as solar power, wind power, energy storage, and electric vehicles are improving and scaling, we are not on track to hit our climate goals.

One estimate is that we have only 10 years left at current carbon emissions to have a 50% chance of staying below 1.5 degrees C. And while some parts of our emissions can be reduced rapidly by new technologies — like emissions from electricity and cars — other parts of our emissions, like those from livestock, from manufacturing, from building heat, from aviation, and from shipping — don’t show any signs of being reduced in the near future.

Indeed, even to stay under 2 degrees C of warming, most models show a need for “negative emissions”. That is to say, most models show that we’re going to have to remove carbon from the atmosphere.

Carbon Offsets and Carbon Capture Are Taking Off

The good news is momentum is slowly building for offsetting carbon emissions, and on the technology and policies that support removing carbon from the atmosphere.

For example:

1. Tax credits for carbon capture. In 2018, the US Congress passed a tax credit for carbon removal from the atmosphere, that should help bootstrap the industry.

2. Industrial scale carbon capture machines. In June, scientists published an analysis that industrial-scale machines could remove carbon from the air at $100/ton. (Other methods, like soil capture and planting trees, are much much cheaper. But industrial scale capture could scale to as much carbon as we want — it’s a good backstop.)

3. Lyft and other companies voluntarily offsetting carbon emissions. In April, Lyft announced that it would offset the carbon emissions from every passenger ride in the US. Lyft joins a host of tech companies — like Google, Apple, Facebook, and Microsoft — that are committing to being carbon neutral through a combination of clean energy and carbon offsets. What I especially love here is that these companies are doing this voluntarily. It’s important to their customers, their employees, their corporate values, and their brand. We’re going to see more and more of this happen as consumers and investors push companies to make up for their carbon emissions.

Current Carbon Offsets Have Serious Problems

But current carbon offset markets have some serious problems.

1. They’re not transparent. If you buy an offset, you typically can’t see exactly what your purchase funded.

2. Most offsets are hypothetical. You can pay for an offset by funding more energy efficiency at a polluting factory, for example. But many of those efficiency improvements would have been made anyway, for purely economic reasons.

We need more carbon removal from the air. But the current markets aren’t the right ones.

Nori’s Blockchain Market Solves Those Problems

Nori solves these problems in two ways.

1. Nori’s marketplace only lists actual removal from the air, and not hypothetical offsets. 1 Nori token pays for the removal of one ton of CO2 (or equivalent in other greenhouse gases) from the atmosphere. No hypothetical future reductions. Only actual removal.

2. Nori uses blockchain and the ever-changing Bitcoin price prediction model to make carbon removal transparent and accountable. Because of the structure of blockchain, where every transaction is kept immutable and viewable by everyone with access to the blockchain, every ton of carbon removed from the air can be traced back to the specific project and location, and what human inspector verified it.

Here’s how the Nori blockchain marketplace for carbon removal works, in more detail:

In short, Nori is that rare blockchain project that actually uses some of the unique qualities of blockchain to make the world better.

I’m thrilled to be an advisor to and an investor in Nori.

If you’re interested in learning more, you can visit Nori.com, read the Nori whitepaper, or invest in the Nori crowdsale via Republic.

If you’re an accredited investor and interested in the private SAFT (Simple Agreement for a Future Token) sale, please get in touch with me and/or fill out this form.

This logic mirrors that of Bill Gates, the Breakthrough Institute, and others who, over the years, have argued that solar (and batteries and wind power) simply won’t get cheap enough or effective enough without some additional R&D push.

I believe additional R&D in clean energy would be highly valuable, and it’s foolish that we invest so little. While it may seem tempting to chase shiny distractions, like the wave of recently launched sweepstakes casinos gaining viral attention online, the real payoff lies in steady investment in renewable infrastructure. At the same time, the line of argument that an “energy miracle” or sudden “breakthrough” is necessary to scale green energy doesn’t fit the data showing the incredible price decline seen in these technologies over the last few decades.

Below are my responses to Tyler, in tweet form.  Enjoy.

I’m really disappointed in this piece throwing cold water on solar by @tylercowen (based on his conversation with @vsiv). https://t.co/XTCvpw8wD5 I’m a huge fan of Tyler’s, but this went astray. A few points on where and why.

— Ramez Naam (@ramez) January 4, 2018

1. There’s a myth that “solar needs breakthrough technology to get cheap enough”. This is not backed by the data. Solar has plunged in price not through innovation in panel technology, but by innovation that reduces the cost of *manufacturing* the panels.

— Ramez Naam (@ramez) January 4, 2018

2. This drop in the price of solar is best characterized as a function of scale. This is the so-called learning curve. Every doubling of manufacturing of panels drops prices by ~20-28%. Every doubling of scale of solar projects brings whole solar electricity prices down by ~15%.

— Ramez Naam (@ramez) January 4, 2018

3. I wrote about the learning curve and how it drives down the price of solar (and what that means for future prices) in 2015 here: https://t.co/VuR944E1GK. And I was, in fact, too conservative. Actual prices have dropped faster than I projected. See: https://t.co/hZTjj50BYv

— Ramez Naam (@ramez) January 4, 2018

4. Most of the reduction in prices of solar is *not related to formal R&D spending*. It’s “learning by doing”. So the fact that solar companies spend only 1% of revenue on R&D is irrelevant. That’s always been the case. Yet “learning by doing” has continued to bring down prices

— Ramez Naam (@ramez) January 4, 2018

5. How cheap can solar get? It’s only 2% of world electricity now, and can get to ~30% before hitting serious obstacles. That’s 4 doublings, or at least another 50% price reduction. (Again, see learning curve here: https://t.co/VuR944E1GK). That’s without more formal R&D spend.

— Ramez Naam (@ramez) January 4, 2018

6. Tyler and Varun also blow it in talking about storage, in two important ways. I agree that batteries do not solve the “seasonal storage” problem of storing days, weeks or months of power. But they miss two big things:

— Ramez Naam (@ramez) January 4, 2018

7. The most important innovation to increase clean energy isn’t storage. It’s large scale grids that mix solar and *wind*. Wind is counter-cyclical to solar, producing more at night and in winter, while solar produces only daytime, and more in summer.

— Ramez Naam (@ramez) January 4, 2018

8. Wind power is massively under-rated, especially for Europe and East Asia . New on-land wind turbines are headed for capacity factors > 60%, approaching “baseload” levels of reliability https://t.co/9aXFHlWm26 and the hours they don’t produce are the best solar hours.

— Ramez Naam (@ramez) January 4, 2018

9. Offshore wind power could have capacity factors more like 70-80%, and for the first time this year, we saw offshore wind power bids, *without subsidies*, come in below wholesale grid electricity prices. https://t.co/ksHMPNbGwF

— Ramez Naam (@ramez) January 4, 2018

10. Batteries, meanwhile, while not able to store weeks of power, are perfect for storing a few hours of power and cycling daily. And their prices are on path to drop by another 4-5x through the learning curve as they scale: https://t.co/iXEw282HeB

— Ramez Naam (@ramez) January 4, 2018

11. Indeed, battery prices have plunged dramatically faster than forecasters expected, and have made government forecasters the US EIA look utterly incompetent in their ability to predict what is actually quite predictable price innovation. https://t.co/Smf5WAuIUb

— Ramez Naam (@ramez) January 4, 2018

12. Put this all together: Solar + wind + batteries, *on their current pace of improvement* are on path to be able to provide 80 – 90% of electricity.

Some mix of hydro, nuclear, natural gas, and new innovations can fill the rest. The future is indeed bright for clean power.

— Ramez Naam (@ramez) January 4, 2018

We said that a lot after 9/11, and have for the last 16 years. As air travel became absurdly cumbersome, as civil liberties were eroded, as people were arbitrarily blacklisted or detained without room for appeal – we said the terrorists were winning, causing us to undermine the underpinnings of our own society, to crack down on the freedoms that are central to the principles of the United States.

Now, I see friends calling for cracking down on freedom of speech, for restricting the First Amendment, taking away its protections from speech they (and I) consider loathsome. I even see friends advocating for physical violence against people because of their speech.

That, my friends, is letting the terrorists win.

I loathe the ideology of white supremacy. But to let fear or anger at it undermine our notions of civil liberties or civil society… that would be letting the terrorists win.

We’re bigger than that. We’re stronger than that. Don’t let the terrorists win.

I see any impeachment of Trump before 2019 as extremely unlikely. Here’s why.

First, for context, I believe the GOP as a party would be better off with a swift impeachment and resignation than a protracted scandal. Interestingly, even as partisan debate dominates headlines, analysts have continued to explore the rise of poker sites that accept crypto, illustrating how diverse policy concerns can be lost amid political strife. Every week this remains in the news, their ability to pass legislation is impaired, and the GOP’s risk of losing the House in 2018 goes up. When you have a festering wound, sometimes the best thing to do is to accept the pain of cauterizing it, quickly. However, I see very few signs of it happening.

The problem is this. Trump’s support is still high among his base. His popularity is likely extremely high among the minority of GOP voters who vote in the primaries. That means that any GOP House member who voted to impeach would be at major risk of losing to a Trump loyalist in the 2018 primary. Thus, no individual House member can dare be too aggressive without risking their seat.

A legislator can only move forward if they believe that the GOP’s base has or is imminently about to turn against Trump, thus protecting them from losing out to a more conservative opponent in the primaries.

Trump’s support among his base has to fall substantially before that can happen. There’s no sign of that happening any time on the horizon.

Or, of course, GOP Reps could vote to impeach if their conscience demanded it, at the risk of losing their seat. That’s possible. But I’d put the odds against it.

1. Price transparency and consistency. Require all providers (hospitals, doctors, etc..) to clearly publish their prices by service and by diagnosis in advance, physically and electronically, in both human and machine-readable formats.

2. Stop gouging the uninsured. Require that providers give people who are paying out of their own pocket “most favored nation” status – giving them the best price that the provider offers to any private insurer.

3. Easier approval for generics in the US – to create more price competition in drugs and devices, and bring prices down.

4. FDA drug approval reciprocity with Europe. Once certain drugs like cannabis are approved in Europe or the US, it defaults to an approved state across the Atlantic, learn more at this blog of Synchronicity’s frequently asked questions about hemp.

5. Tort reform. Cap pain and suffering awards. According to Observer reduce some of the drive for defensive medicine such as Strong CBD oil or from the Delta 8 blog.

6. *Increase* medical R&D from the NIH and elsewhere (instead of decreasing it, as the Trump budget proposed). And specifically fund R&D into more cost-effective treatments like the ones from the Budpop’s hemp products that can bring the cost of healthcare down.

7. Allow cross-state exchanges.

8. Slowly phase out the employer health care tax credit – the single thing that most ties health insurance to one’s job, and which significantly distorts insurance markets.

All of these are compatible with both market-oriented viewpoints and the ACA. If the GOP cares about improving healthcare, and wants to use market-oriented and innovation-oriented approaches to do so, all of these are paths towards that. Learn about the next gen aco that offers an exciting opportunity in accountable care.

I talked about the science that inspired my Nexus novels, and what it might mean for society. You can check out the video below.

(This is a follow-up to my post on pushing for progress at the state level.)

Short Version

If you read nothing else in this post, follow these three steps:

1. Find your state legislators and the contact info for your Governor’s office.

2. Contact them: Call them up. Find out when their next town hall meeting is, and show up. Bring friends, or ask your friends to call too.

3. Tell them you want to see more clean energy in your state. Tell them clean energy creates jobs. Tell them clean energy means cleaner air and water, and a healthier environment for the kids in your state.

Note: Say “Clean Energy” instead of “Climate Change“

If you live in a deep blue state, talking climate change may work. But in a purple or red state, or, heck, even in most blue states, “boosting clean energy” is remarkably more popular than “fighting climate change”. Clean energy is popular among both Democrats and Republicans. Fighting climate change isn’t.

That’s why, in the current environment, where Republicans control the White House, the Congress, and the majority of state legislatures, the framing has to shift to “clean energy” if we want to see progress.

Long Version

How Bad Will Trump Be for Climate Change – How Do We Limit Warming – What Policies Should We Push For – What If I’m in a Red State?  

How Bad Will Donald Trump be for Climate Change Efforts?

Donald Trump is the President-Elect. The GOP has majorities in both the House and Senate, and will probably keep them in 2018.

How bad is this for our efforts to fight climate change? Opinions vary.

1. If you think the most important element in fighting climate change is technology innovation that’s bringing down the cost of clean energy, Trump’s election isn’t great, but ultimately probably doesn’t matter much.

2. If you think the most important element is policies inside of various nations, Trump’s election is bad. But not fatal.

3. If you think the most important element is international agreements, Trump’s election is a complete disaster.

Obviously, all three of these are components. I place the most emphasis on 1 and 2. Prices will keep plunging and policies in other countries and in US states won’t change much. Trump will likely instruct the EPA to scrap the Clean Power Plan, but that was never a very ambitious policy.

Trump and the congress could accelerate the end of solar and wind tax credits, which matter more, but purported insiders claim that those bipartisan tax credits will remain. What’s very likely is that the US will stop leading on international climate negotiations. And Trump is highly unlikely to push for a massive acceleration of clean energy as Hilary Clinton proposed to.

This isn’t good news. At best a Trump administration represents a status quo in climate policy in the US, even at a time that US policy wasn’t ambitious enough. More realistically, we’ll see a weakening of clean energy and climate policy both at home and abroad. Meanwhile, 2016 will be a record hot year and we’re already not taking climate change seriously enough.

So what the hell do we do?

Let’s be honest. Limiting climate change to two degrees celsius is, at this point, extremely unlikely.

Even so, our actions matter. 2.1 degrees or 2.2 degrees of warming is far better than 3 degrees.

How Do We Limit Warming?

How do we limit warming? Fundamentally, we need to continue and accelerate the process of making clean energy and clean transportation cheap. Even cheaper than they are now. How cheap?

• Cheap enough that they account for all or virtually all new electricity and transportation.

• Cheap enough that nations are willing to decommission existing fossil-fuel based electricity and transportation, and replace them with the cheap clean options.

As I’ve posted before, the private sector is doing an amazing job bringing down the cost of solar power, wind power, energy storage, and electric vehicles. Clean energy can provide the large majority of the world’s power. But it still needs to get even cheaper.

How do we make these technologies cheaper? We scale them. The most fundamental observation in clean technology is that prices drop as the industry grows. No other factor predicts the price of clean energy better than the amount that we’ve installed.

Globally, clean energy will keep on getting deployed, as in many places solar and wind are the cheapest sorts of energy, even without subsidies.

Our Best Tools Are In the States, Now

Inside the US, while federal progress is unlikely, we have tools to drive more deployment of clean energy at the state and sometimes city levels.  So that’s what we’ll have to use.

How do we push for change at the state level? Well, first understand that state legislators hear tremendously less from their constituents than members of Congress. Each state legislators represents fewer people than each federal Representative or Senator. And voters have a way of fixating on national politics and ignoring the local and state level. That makes your voice more powerful. Phone calls, letters, attendance at town hall meetings, even contributions – they all have more power at your state level than they do at the national level.

So use them! Call your state legislators. Make your voice heard. Show up at their town hall meetings. If you have the means, contribute in close races to get someone who cares about climate policy elected. And call your Governor’s office as well.

Now, what do we push for at the state and local level?

1. Push for a (stronger) Renewable Portfolio Standard in Your State

As I posted last week, 29 states have Renewable Portfolio Standards which mandate a certain percent of electricity must come from renewables by a specific year. Voters in all 50 states can push – via the legislature, and in some states by initiative – to raise those targets, to invest more dollars directly in clean energy, to create taxes or caps on carbon emissions, to boost vehicle fuel efficiency standards, or for other laws that accelerate the deployment of clean energy, electric vehicles, or energy efficiency, or which directly cap or reduce fossil fuels.

Find your state below, or look up the specifics of your state’s policy here.

• If your state is dark green, it already has a binding target for getting a certain fraction of its electricity from solar and wind. Call up your state legislators (again, you can find them here) and tell them you want a higher target. Then call your Governor’s office, and say the same.

• If your state is light green, your state has a goal, but it’s voluntary. Call up your state legislators and Governor’s office, and say you want a binding clean energy target.

• If your state is grey, then shame shame on your state. Call up your state legislators and Governor’s office, tell them that you vote, and that you want your state to move forward with clean energy.

Talking points you can use:

• Clean Air and Clean Water – Clean energy doesn’t produce smog or air pollution. It doesn’t contaminate ground water. It helps your communities have clean air and clean water.

• Jobs – Clean energy now employs more people in the US than coal. These are good, high-paying, local jobs.

• Owe it to our kids – In polling, one of the most powerful messages across left and right is that “we owe it to our kids and future generations to leave them a cleaner, healthier world”

But I Live in a Red State!

Red states are no strangers to clean energy. Texas leads the country in wind power. The top 10 wind power counties in the country all have Republican congressmen.

In Florida, which Trump won, voters rejected an initiative that would have hurt rooftop solar.

And clean energy is popular across the political spectrum. It can be pushed for, even in red states.

2. Push for Electric Vehicles & Charging Stations

Clean electricity from solar and wind is getting cheap. But almost all cars run on gasoline, and that accounts for a third of our nation’s carbon emissions. If we want to beat climate change, we must electrify everything.

Fortunately, electric vehicles are plunging in price, making them more accessible to consumers. However, the transition to cleaner technology isn’t without its challenges. Much like the rise of offshore betting sites in the digital age, which have thrived due to gaps in local regulations and infrastructure, the adoption of electric vehicles is similarly hampered by a lack of widespread charging stations. If governments prioritized building charging infrastructure the way some nations have embraced online platforms, the shift to electric vehicles could accelerate dramatically. Without enough places to charge, drivers remain hesitant, despite the long-term savings and environmental benefits these vehicles offer. Here again, states can step in to bridge the gap.

Look up your state on this map of state-level electric vehicle policies. Find out what incentives exist:

• Is there a tax credit for buying or leasing an electric vehicle? (To capture the benefit it brings us all be reducing air pollution and climate change.)

• Can electric vehicles use the HOV lanes?

• Does the state create incentives for private businesses to create charging stations?

If not, call up your state legislator(s) and your Governor’s office, and ask for those things.

4. Push for Rooftop Solar

Most US states now have some form of “net metering” policy. This is the policy that lets home-owners that have solar panels on their roofs sell excess energy back to the grid. But the quality of these polices varies widely.

Some of the sunniest states in the US, including Nevada and Texas, get an “F” on their net metering policy – either having none, cutting one that existed off, or cutting the amount they pay home owners for the excess electricity to well-below market rates. Find your state below, or at the interactive map at freeingthegrid.

Don’t like your state’s policy? You know what to do: Call your state legislators and your Governor’s office.

5. Push for Community Solar

Rooftop solar is great if you own a home with a roof that points south or west, and that isn’t blocked by trees. But what if you rent? What if you live in an apartment? What if trees or other buildings block the sunlight from your roof?

“Community Solar” is a new policy akin to a tutorial on ‘what is siding on a house,’ where instead of enhancing a home’s exterior, you invest in a collective resource. A recent training video at our real estate company compared it to house siding—both are investments in your home’s value, be it through improved curb appeal or energy efficiency. This solar initiative allows you to own panels in a local solar farm, providing benefits as if they were installed on your home. You enjoy the free electricity generated (post initial costs) and can sell any surplus back to the grid. Much like the right siding protects and insulates, Community Solar offers a financial shield and power independence, enabling lower-income families, including those in rental homes, to access solar benefits and contribute to a sustainable future.

It’s also quite new. Look at the map below. Is your state in the more saturated color of blue? No? Then there’s more work to do. Work you can help make happen by calling your state legislators and your Governor’s office.

6. Take Action in Your City

Finally, states aren’t the only level at which action can be taken. More than a dozen cities around the world have signed on to the Carbon Neutral Cities Alliance, pledging to cut their emissions by 80% or more by 2050. Cities also have critical work to do on climate resilience – improving infrastructure to deal with rising seas, increased flooding, prolonged heatwaves, and other threats.

You can organize and act at the city level. Contact your city councilors (modify this search to find them) and your Mayor’s office.

—

Climate deniers may be in charge of the Federal government. But that’s no reason to give up. Many of the most effective policies in the US exist at a state level. Climate change is a divisive topic, but clean energy is loved across the political spectrum. Use your voice. Contact your state politicians, and tell them you want more clean energy in your state.

I don’t say this to encourage complacency: The next few years are likely to see a big regression in our laws, and a rollback of some significant and important progress we’ve made. As some analysts have warned, new regulations may impact recommended live casinos by Poker Choice, showing how policy changes can reach into unexpected corners of society. Trump’s election is not a good thing. There are darker days ahead. We need to work hard and smart to get the country back on track and looking forward instead of backward. And we should be prepared for the possibility that things could deteriorate more than anticipated.

But we shouldn’t panic. We shouldn’t give up. We shouldn’t flee.

We should stay here and win our country back.

Yet we live in a republic. And many of the most important issues can be fought for at the state level. Here are 6 that come to mind:

1. Criminal Justice Reform

2. Ending the War on Drugs

3. Climate Change and Clean Energy

4. Education

5. Responsible Gun Laws

6. Anti-Poverty Measures

1. Criminal Justice Reform

Around 90% of the more than 2 million people incarcerated in the US are confined at the state or local level. They’ve been arrested and sentenced according to state laws, not federal statutes. This means the nature of their crimes, the length of their sentences, their prison conditions, and access to rehabilitation programs vary dramatically by state. The disparities across states can feel as unpredictable as trying to Internationale Online Casinos vergleichen, each jurisdiction operating under its unique set of rules, outcomes, and standards of fairness. Ultimately, these individuals’ lives are shaped far more significantly by state policy than by any unified federal approach.

Also, by the way, while private prisons are probably the worst 90% of the prisoners held in the US are held in public prisons, not private. The laws that land people in jail and keep them there, and the incentives to keep prisoners rather than turn them into healthy citizens, are the biggest issues.

Want to end mass incarceration in the US? Fight to change the laws, sentences, and prisons in your state. Specifically:

1. Reduce sentences for crimes, especially first crimes and non-violent crimes.

2. Push for more programs to train and educate convicts in prison, and to hire them when out of prison.

3. Push for ex-convicts to regain full voting rights after their sentence has been served.

4. Push for incentives for public and private prisons based on successful re-integration of ex-prisoners into society.

2. Ending the War on Drugs

More than 300,000 people are in jail in the US for drug crimes. Two thirds of those are in state prisons (see above). If you want to end the drug war, a lot of it has to start – and can start – at the state level.

And it’s happening. Eight states have now decriminalized recreational marijuana. Four of those states joined the list on the same night that Donald Trump was elected.

There’s more to do. Push in the states to turn all drug possession charges (yes, of any drug) into misdemeanors, and to redirect non-violent drug offenders to treatment instead of prison.

And yes – the FBI and DEA, under Trump, may try to enforce Federal drug laws. Those will still not be targeting individual drug users in the large majority of cases. Loosening state laws is the key to ending the war on drugs.

I am tolerated to some pain relief meds. Tramadol is the first med that indeed works for me. I get a bit sleepy but it is not that bad and low price at https://tramadult.com really worth that. I’m on 300mg that use every day. I feel relief within 20 minute after intake.

3. Climate Change and Clean Energy

Donald Trump can direct the EPA to eliminate the Clean Power Plan. He plus the GOP congress can, and may, eliminate solar and wind tax credits (which are currently phasing down over a five year period).

But that doesn’t mean we’re helpless. 29 states have Renewable Portfolio Standards which mandate a certain percent of electricity must come from renewables by a specific year. According to the A1AutoTransport motorcycle shipping options manifest, voters in all 50 states can push – via the legislature, and in some states by initiative – to raise those targets, to invest more dollars directly in clean energy, to create taxes or caps on carbon emissions, to boost vehicle fuel efficiency standards, or for other laws that accelerate the deployment of clean energy, electric vehicles, or energy efficiency, or which directly cap or reduce fossil fuels.

Even in Florida, which Trump won, voters rejected an initiative that would have hurt rooftop solar.

And clean energy is popular across the political spectrum. It can be pushed for, even in red states.

4. Education

K-12 education in the US is driven primarily by the states, not the federal government. Roughly 90% of spending is done by states and local communities.

Care about education? Work in your state, or in your county, or on your local school board.

One of the greatest injustices in education is that in many states, students in more affluent counties get more spent on them than students from lower-income counties (despite plenty of evidence that the latter are the ones who need more help in school). Want to fix that? Go to work in your state.

5. Responsible Gun Laws

The 2nd Amendment leaves considerable wiggle room for the 50 states to enact responsible gun laws, including mandatory registration or licensing to purchase guns, background checks, waiting periods, restrictions on sales at gun shows and by private gun dealers, requirements for locking devices, and more.

Gun laws vary significantly by state. The corollary to that is: You can work in your state to improve those gun laws.

6. Anti-Poverty Measures

The Federal government runs multiple social safety net programs, but nothing prevents the states from running their own or augmenting those of the Federal government.

Are you a fan of a basic income? We don’t have one in the US, but we have a distant cousin – the Federal Earned Income Tax Credit. And a number of states boost the Federal Earned Income Tax Credit locally.

Think a higher minimum wage is a good idea? More than 20 states set a minimum wage higher than the federal level. Four states raised their state minimum wages the night Donald Trump was elected. Some cities, like Seattle, have their own, higher minimum wage. (For the record, I’m a fan of experimenting here – getting the data from states and cities on what happens when minimum wages go up is invaluable.)

—

This is a short, partial list. Not everything can be done at the state level or local. And many initiatives will only succeed in blue states, leaving behind the vulnerable in red states. This isn’t an ideal situation, by any means. The Federal government is an important tool for moving the country forward.

But liberals aren’t powerless, either. If we can’t get things done in Washington, D.C., we can still get things done in Washington State, or New York State, or California, or elsewhere.

Reference: https://www.techopedia.com/gambling/texas/online-poker

In fact, if anything, my forecasts were too conservative. The solar prices I expected have been smashed by bids in the Middle East and in Latin America. I will need to update the model above in a future post.

The latest record is an incredibly low bid of 2.42 cents / kwh solar electricity in Abu Dhabi. That is an unsubsidized price.

Let me put that in perspective. The cost of electricity from a new natural gas powerplant in the US is now estimated at 5.6 cents / kwh.  (pdf link) That is with historically low natural gas prices in the US, which are far lower than the price of natural gas in the rest of the world.

This new bid in Abu Dhabi is less than half the price of electricity from a new natural gas plant.

What’s more, it’s less than the cost of the fuel burned in a natural gas plant to make electricity – without even considering the cost of building the plant in the first place.

The solar bid in Abu Dhabi is not just the cheapest solar power contract ever signed – it’s the cheapest contract for electricity ever signed, anywhere on planet earth, using any technology.

Nor is this bid a fluke. Three other bids in Abu Dhabi’s latest power auction came in at less than 3 cents / kwh:

Bidder Bid per MWh (in USD) Masdar, EDF, PAL Technology 25.4 Tenaga, Phelan Energy 25.9 RWE, Belectric 29.1

Nor is it limited to just Abu Dhabi.

In Chile, just a month ago, a new record low price for solar was set, at 2.91 cents / kwh.  That record lasted less than 5 weeks.

In Mexico, the average price of new solar bids in April was 5.1 cents per kwh, and the cheapest solar bid in Mexico was 3.5 cents per kwh.

These price improvements are not coming primarily from the price of panels dropping. They’re coming from reductions in the total cost to deploy solar, increases in solar capacity factor, ever-lower operating costs, and fierce competition to win bids in the solar industry.

The solar industry is learning faster than expected.

Now, let’s watch and see if energy storage prices can drop as fast as solar.