Five dispatches from the week the frontier moved again. Capability, capital, and the labour it leaves behind.
Stanford's 2026 Index. Peptides go mass-market. The layoff trap, formalised. OpenAI's New Deal pitch. The model Anthropic refused to ship.
David Borish
From New York
Five articles, one week, all sourced from The AI Spectator
Stanford HAI, 2026
The 2026 AI Index runs to 423 pages. The shortest summary: SWE-bench Verified climbed from sixty per cent of the human baseline to nearly one hundred in a single year, and the U.S. lead over China is now smaller than benchmark noise.
Stanford's Institute for Human-Centered AI published its annual Index this week, and the headline finding refuses the comfortable framing the field has grown used to. Frontier models now meet or exceed human baselines on PhD-level science questions, multimodal reasoning, and competition mathematics. Eighty-eight per cent of surveyed enterprises report adoption. Four in five university students use generative AI for coursework.
The geopolitical finding is more pointed. The top-tier model leaderboard has become a rolling tie. U.S. and Chinese frontier models traded the lead multiple times through 2025; as of March 2026, the top U.S. model leads by 2.7 per cent, a margin within the noise of benchmark variance. The United States still produces more notable models (50 to 30 in 2025), but China leads in publication volume, citation share, and patent grants, and South Korea leads in AI patents per capita.
The labor finding is the one that should travel furthest. In software development, where the productivity evidence is clearest, U.S. developer employment for workers aged 22 to 25 has fallen nearly 20 per cent from 2024, while headcount for older developers continues to grow. One-third of surveyed organisations expect AI to reduce their workforce in the coming year. Productivity gains in entry-level-heavy tasks; employment decline in the youngest cohort of the same occupation; forward-looking employer signals pointing to more of the same. Analysts have been waiting to see this pattern. It is here.
Underneath, the structural fragility: 5,427 AI data centres in the United States, ten times any other country; 29.6 GW of data-centre power capacity, comparable to New York state at peak demand; Nvidia accounting for over sixty per cent of total compute; and a single Taiwanese foundry, TSMC, fabricating almost every leading AI chip. The performance gains documented elsewhere in the report depend on a hardware supply chain with a single point of failure.
Semaglutide alone is tracking toward thirty-six billion dollars in 2026. Meanwhile, in Nature Communications, a hybrid AI framework just designed a twelve-amino-acid peptide that kills the dormant form of MRSA. The cultural moment and the technical moment are the same moment.
For most of the last century, peptides sat in an awkward middle ground of drug development. Too large and fragile to behave like small-molecule pills, too short and unstable to carry the commercial weight of antibody biologics. Insulin worked. A handful of others found niches. Mostly, peptides were a pharmacology of special cases.
That changed fast. By early 2026, roughly thirty million people were using GLP-1 drugs, up from about four million at the start of the decade. Wegovy launched as a pill in January, removing the injection barrier that had kept a large share of the patient population on the sidelines. Novo sold fifty thousand weekly subscriptions in the first three weeks. The category proved a peptide could become a mass-market chronic-disease medication rather than a specialty product, and the commercial gravity pulled investment into every adjacent problem peptides might solve.
The bottleneck has always been the same: a peptide ten amino acids long has about ten trillion possible sequences. Traditional screening cannot cover it. Machine learning changed the economics in several ways at once. Predictive models trained on databases of measured peptides screen millions of candidates in silico before any are synthesised. Generative models, descended from the same architectures that power chatbots, propose novel sequences with targeted properties. Structure-prediction tools downstream of AlphaFold model the three-dimensional shape of every proposal.
The CAMPER paper in Nature Communications makes the architecture concrete. A team at Houston Methodist Hospital built a two-stage system: a random forest classifier trained on 3,660 peptide sequences, then a biophysical scoring function that re-ranks top candidates by net charge, hydrophobicity, amphipathicity, and helical structure. The output, WP-CAMPER1, killed MRSA at four micrograms per millilitre, eight times more potent than the natural mastoparan peptides the library was derived from. Adding the biophysical scorer cut false positives by seventy-seven per cent.
Claude Mythos Preview, announced on April 7, found a twenty-seven-year-old bug in OpenBSD, an operating system whose entire identity rests on being secure. It found a sixteen-year-old vulnerability in FFmpeg's H.264 codec, in a line of code automated testing tools had scanned five million times. It chained together multiple Linux kernel vulnerabilities to escalate from ordinary user access to full root control. It did all of this autonomously, without human guidance after an initial prompt.
Anthropic is not releasing the model. Instead, it has built a coalition of twelve major technology and security companies to use it strictly for defensive purposes. The decision to withhold a frontier model while spending one hundred million dollars to let select partners use it reveals where AI capabilities now sit. This is a model whose offensive security skills emerged as a side effect of being good at coding and reasoning.
On CyberGym, Mythos Preview scored 83.1 per cent against Opus 4.6's 66.6 per cent, but the raw numbers understate the qualitative jump. On a Firefox 147 JavaScript engine test, Opus 4.6 produced working exploits twice across several hundred attempts. Mythos Preview produced 181, with register control on twenty-nine more. CrowdStrike's CTO put it plainly: the window between vulnerability discovery and exploitation has collapsed from months to minutes.
Read the full article →Falk and Tsoukalas prove that rational firms with perfect foresight will automate well past the point where doing so harms their own profits. The race is not a misunderstanding. It is a dominant strategy.
A new theoretical economics paper from Brett Hemenway Falk at the University of Pennsylvania and Gerry Tsoukalas at Boston University arrives at a conclusion that should unsettle anyone watching the current wave of AI-driven layoffs. Working from a competitive task-based model in the tradition of Acemoglu and Restrepo, the authors show that the over-automation we are witnessing is not an error. It is the equilibrium.
The model is deliberately stripped down. A sector contains some number of symmetric firms, each with a workforce performing tasks that can be replaced by AI at lower cost. Each firm chooses what fraction of its workers to displace. Workers spend a portion of their wages on the sector's output. When a firm automates, the displaced workers lose income, and a fraction of that lost income would have flowed back into the sector as consumer spending. The demand that firm destroys is shared across all firms; the cost savings it captures are entirely its own.
This asymmetry is the entire trap. A firm that automates one task saves the full wage but bears only one-Nth of the resulting demand loss, where N is the number of firms in the sector. The remaining demand loss falls on rivals. Each firm's first-order condition therefore understates the social cost of its own decision. The authors prove that automating at this privately optimal rate is a strictly dominant strategy. It does not depend on what rival firms do or whether they understand the consequences.
The most consequential finding is that over-automation is not a transfer from workers to firm owners. It is a deadweight loss that harms both. The Nash equilibrium is Pareto dominated by the cooperative optimum. Both factor classes would prefer a world in which firms collectively automate less. The authors evaluate six policies. Universal basic income and capital income taxation cannot correct the distortion. Worker equity narrows the wedge but cannot close it. The only instrument that survives is a Pigouvian automation tax.
The paper maps directly onto the framework I have been tracking in the Exponential Replacement Curve and the Open-Prem Inflection Point. The V2 cost data shows AI costs are already two to three orders of magnitude below human costs in knowledge work. The Red Queen finding in the Falk-Tsoukalas paper, that higher AI productivity widens the over-automation wedge rather than closing it, deserves particular attention. The V2 doubling rate is 5.5 months. Each doubling pushes the wedge wider, the threshold lower, and the cooperative optimum further from the equilibrium that competitive markets actually reach.
The most valuable private technology company in the world just published a detailed plan for how the government should tax, regulate, and redistribute the wealth generated by the very technology it is racing to build. The question is whether anyone should take it at face value.
OpenAI's "Industrial Policy for the Intelligence Age," released April 6, lays out six core proposals: a national public wealth fund, taxes on automated labour, a shift in the tax base from payroll to capital gains, four-day workweek pilots at full pay, automatic safety-net triggers tied to economic data, and containment playbooks for AI systems that cannot be easily recalled.
Sam Altman told Axios the scale of disruption ahead is comparable to the Progressive Era and the New Deal. The wealth fund is the most structurally ambitious proposal: nationally managed, seeded in part by mandatory contributions from AI companies, returns distributed directly to American citizens. The model draws from Alaska's Permanent Fund. Anthropic proposed a similar idea in its October policy paper. Two of the three leading frontier labs have now explicitly argued that AI's gains need to be redistributed through public ownership.
The timing is difficult to separate from the content. OpenAI released the document on the same day The New Yorker published a year-and-a-half investigation into Altman's leadership, including allegations that he repeatedly misrepresented safety protocols to his own board. Whether the document reads as responsible foresight or as cover for regulatory nihilism depends largely on how much credibility you assign to the messenger.
The labour backdrop gives the warnings real weight regardless. White-collar payrolls in the United States have contracted for twenty-nine consecutive months, a stretch economists describe as unprecedented outside a recession. AI was cited as the reason for over fifteen thousand of the sixty thousand planned job cuts announced in March alone.
Read the full article →
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.