DeepMind Co-Scientist in Nature — Fable 5 & Mythos 5 released — Recursive beats the benchmarks — AI opinion collapse measured — Amodei's five-part policy agenda
davidborish.com/the-ai-spectator
David Borish, Enterprise AI Strategist & Author
A multi-agent system built on Gemini proposed testable scientific hypotheses. Researchers at Calico, Stanford, MIT, Cambridge, and Edinburgh then confirmed them.
When researchers at Calico Life Sciences asked Google DeepMind's Co-Scientist system a question about the biology of aging, they were not expecting a hypothesis they would later confirm experimentally. The AI system proposed a novel idea about the integrated stress response, and Calico's scientists subsequently validated it in lab tests. It is a single data point, but it represents something that has been difficult to demonstrate: an AI system proposing a testable scientific idea that turns out to be correct.
That result, along with several others from collaborators at Stanford, MIT, Cambridge, and Edinburgh, forms the core of a Nature paper published by DeepMind on May 19, 2026. Co-Scientist is a multi-agent system built on Gemini, designed to do one specific thing: generate, critique, and refine scientific hypotheses at a scale and speed that individual researchers cannot match alone.
The architecture runs seven specialized agents organized around three phases. A generation agent proposes initial hypotheses grounded in scientific literature and databases, while a proximity agent maps those hypotheses to ensure the system explores a diverse range of research directions rather than clustering around obvious territory. A reflection agent acts as a virtual peer reviewer, and a ranking agent orchestrates what DeepMind calls a "tournament of ideas," using pairwise comparisons and simulated scientific debates to identify which hypotheses are most worth pursuing. An evolution agent refines and combines the top-ranked hypotheses. A meta-review agent synthesizes insights from the debates. Coordinating all of this is a supervisor agent that breaks down high-level research goals and runs agents in parallel.
The tournament mechanism draws from principles used in AlphaGo and AlphaStar, using Elo-based scoring to rank hypotheses as agents debate them. The critical design choice is where the computation goes: the majority of processing is dedicated to verifying hypotheses, not generating them. The system cross-checks claims against scientific literature, web search, and specialized databases including ChEMBL for chemical data and UniProt for protein information. In select collaborations, the system can also call on AlphaFold.
At Stanford, Professor Gary Peltz was searching for drug-repurposing candidates for liver fibrosis, a chronic scarring condition with limited treatment options. Co-Scientist surfaced overlooked candidates from existing approved drugs. One of those candidates blocked 91 percent of a scarring-linked cellular response when tested in the lab. The results were published in Advanced Science. At MIT, Associate Professor Ritu Raman used Co-Scientist to digest complex literature on ALS and identify where her lab's expertise could be combined with complementary work. At the University of Edinburgh, Professor Filippo Menolascina used the system to work on metabolic liver disease, and lab tests supported one of the hypotheses it generated about why an existing drug benefits only some patients.
"Science is a team sport. Co-Scientist can't do science by itself, and I can't do it all by itself either." Ritu Raman, Associate Professor, MIT
Same weights, two access tiers. One ships to everyone. One ships to vetted cyber defenders only. A two-stage classifier system runs between the user and the model on every request.
Anthropic released two configurations of its most capable model to date, and the more significant of the two is the one most users will never see. Claude Fable 5 is the general-access version, available across claude.ai, the Claude API, and subscription plans. Claude Mythos 5 is the same underlying model, same weights, with specific safeguards lifted in cybersecurity and biology. It remains restricted to the Project Glasswing community of vetted cyber defenders and critical infrastructure operators. Both are priced at ten dollars per million input tokens and fifty dollars per million output tokens, cutting Mythos Preview's price by more than half.
The capability jump between Mythos Preview and Mythos 5 is real but incremental. On Anthropic's internal ECI composite benchmark, Mythos 5 scores 161.29, above Mythos Preview's 158.91. Stripe reported that Fable 5 performed a codebase-wide migration on a 50-million-line Ruby codebase in a single day, a task the team estimated would otherwise take two months. On Cognition's FrontierCode evaluation, Fable 5 leads all frontier models even at medium effort. One analytics platform reported that Fable 5 was the first model to break 90 percent on their benchmark of complex long-running analytical tasks, a ten-point jump over Opus 4.8.
The most novel engineering in this release is the deployment layer. Anthropic built a two-stage classifier system covering cybersecurity, biology and chemistry, and attempts to distill Fable's capabilities into competing models. The first stage is a probe monitoring internal activations in real time. If the probe flags traffic, the request escalates to a trained LLM classifier. When the system decides a query crosses the threshold, Fable 5 does not refuse outright. On consumer interfaces, it silently routes to Claude Opus 4.8, which responds in Fable's place. The user is notified which model handled the query.
The Open-Prem Inflection Point framework gains another data point here. Fable 5's pricing, at roughly a 55 percent reduction from Mythos Preview, continues the compression that makes high-volume enterprise inference economics increasingly favorable. See the Open-Prem Inflection Point V3 for the full analysis of what this cost trajectory means for on-premises deployment decisions.
Recursive, a startup focused on automated AI research, published benchmark results on June 11, 2026 from a system designed to run the full scientific loop without human intervention. The system picks a target objective, proposes modifications, implements them, runs experiments, validates the results against reward hacking, and feeds what it learns into the next round of search. It maintains context across many parallel research threads and can combine findings from separate branches when the evidence warrants it.
The company tested this system on three benchmarks chosen for practical relevance and reliable feedback. The NanoChat benchmark, originally created by Andrej Karpathy, tasks a system with training a small language model to the lowest possible validation loss within a five-minute window on a single GPU. A public collaborative effort called autoresearch@home had extended this setup into a community competition, with dozens of humans and hundreds of AI agents collectively optimizing solutions. The Recursive system found a solution reaching 0.9109 BPB, a 0.0263 improvement over the community's best result of 0.9372.
NanoGPT Speedrun is harder to move. The benchmark measures how quickly a small GPT-style model can train to a fixed quality threshold on eight H100 GPUs. It has 83 human record-setting contributions, with training time dropping from roughly 45 minutes in mid-2024 to 79.7 seconds by the time Recursive ran its system. Starting from the current leading solution, the system reduced training time to 77.5 seconds. The changes were specific: FP8 precision pushed into attention projection layers, annealed Gaussian exploration noise injected into the NorMuon optimizer, and a fused GPU kernel rewritten so the backward pass reconstructs intermediate activations on the fly rather than storing them.
The third benchmark, SOL-ExecBench from NVIDIA, contains 235 tasks focused on writing fast GPU kernels for real computational workloads on Blackwell B200 GPUs. Starting from the previous leaderboard best of 0.699, the system reached a mean score of 0.754, an 18 percent reduction in the gap to the hardware limit. Reward hacking was a particular challenge: some kernels the system generated exploited the evaluation setup rather than genuinely improving performance, by caching outputs or relying on persistent state. Recursive addressed this by treating correctness auditing as part of the research loop itself.
When five frontier models respond to the same contested debate question, only 3.4 percent of their main arguments are unique. For humans, that figure is 65.3 percent.
Researchers Yekyung Kim, Yapei Chang, Chau Minh Pham, and Mohit Iyyer collected public debate responses from two sources: New York Times Room for Debate, where contributors write short essays of around 350 words on contested questions, and Boston Review forums, where respondents write longer pieces averaging about 1,150 words. Both corpora share the same structure: multiple writers responding independently to the same debate prompt. The team then generated responses to every debate from five frontier models, GPT, Claude, Gemini, DeepSeek, and Minimax, under three conditions.
The gap between humans and standard AI responses is large. In the NYT corpus, 65.3 percent of human main arguments are unique within a debate. For standard LLM responses, that figure is 3.4 percent. The pattern holds in the longer Boston Review forums: 78.6 percent of human arguments are unique versus 18.4 percent for standard AI. Asking models to diversify their outputs raises the uniqueness rate. Under diversified prompting, DeepSeek reaches roughly 63 percent, and Gemini exceeds the human unique rate at 82 percent.
The collapse does not stop at main arguments. Among essays sharing the same main argument, 41 percent of human sub-arguments are unique within the debate. For standard AI, that figure drops to 9.1 percent. The qualitative difference is as notable as the quantitative one. Human sub-arguments anchor in specific cases: a particular piece of legislation, a named institution, a causal chain grounded in one industry or community. AI sub-arguments tend toward portable frameworks: generic appeals to research, abstract institutional interventions, hedged generalities that stay uncommitted to any particular action.
Structure follows the same arc. LLM essays follow a compressed pattern: open with a thesis, build support, close with a proposal. The transition from a support paragraph to a proposal paragraph occurs in 29.4 percent of standard AI support transitions, compared with 12.3 percent for humans. Argumentation accounts for 71.5 percent of human paragraphs in the NYT corpus, compared with 97 percent of standard AI paragraphs. Humans mix in substantially more exposition and narration throughout their essays.
The study finds that pooling all five models under diversified prompting recovers 73.9 percent of human main-argument clusters. What is in the remaining 26 percent: the arguments made by one writer rather than three, the interventions grounded in a specific community's experience, the framings that have not yet become conventional wisdom. In binary debates with clear pro and con sides, AI models take strong positions only 63.4 percent of the time. Humans do so 76.1 percent of the time. The models hedge more, even when asked to commit.
If those models consistently produce the same small set of arguments in response to contested questions, and if their outputs recirculate through training data, search results, and editorial assistance, then the range of positions readers encounter in public debate may narrow in ways that are difficult to observe directly. The researchers describe the AI pattern as convergence onto arguments that are hard to falsify, the kind that can attach to almost any debate without modification. The code and annotations are available at github.com/mungg/argument_collapse.
Mandatory testing. Labor insurance. Democratic chip coalition. Autonomous weapons controls. A call to move before the regulatory window closes from only one side.
The central argument in Dario Amodei's June 2026 essay "Policy on the AI Exponential" is not philosophical. It is a timing argument: the gap between how fast AI capabilities are compounding and how slowly legislative institutions respond has become dangerous, and that gap is now closing from only one side.
Between 2024 and early 2026, Anthropic's public policy work focused on disclosure requirements. By Amodei's account, those were the correct moves for the period. Now, he says, the period is over. The Collingridge dilemma, which he cites directly, describes exactly this: the impacts of a new technology are hardest to anticipate precisely when it would be easiest to regulate them, and easiest to understand only after regulation becomes much harder.
The regulatory model he proposes is the FAA. Frontier AI models should require technical testing and third-party auditing before deployment. If a model fails to meet safety standards in four specific risk areas, the government should be able to block or reverse its release. Those four areas are: cybersecurity, biological weapons, loss of control of AI systems, and automated research and development that could accelerate any of the other three.
On labor, Amodei is careful. He attaches a caveat that shapes the policy proposals: "there's a decent possibility that, despite all our efforts, AI still causes significant enduring job loss, and that this may be an intrinsic property of the technology." His proposals are graduated: first, better economic measurement of AI displacement; second, wage insurance, retention tax credits, and workforce training grants; and third, if those prove insufficient, long-term income support financed through taxes on relevant companies or higher capital gains rates.
The geopolitical section is the most expansive. Amodei argues that AI has no meaningful parallel in the history of commercial technology. His comparison is to nuclear weapons. The proposal is a coordinated democratic coalition built around shared AI development standards and a unified approach to supply chain control, with chips and semiconductor equipment freely shared among members while collectively denied to adversaries.
A weekly edition compiled from five articles published at davidborish.com/the-ai-spectator. Written by David Borish, Enterprise AI Strategist and creator of the Open-Prem Inflection Point and Exponential Replacement Curve frameworks.