The slowdown case acquires a price tag

The slowdown case acquires a price tag

Dario Amodei's essay of Saturday, "We Must Pace the Frontier", does one thing that costs Anthropic something — embedded third-party evaluators with desks, badges and the right to publish without editorial control — and asks governments to make everyone else do the same. Altman and Musk endorsed it within hours; Altman also told Fortune an IPO now would be "ill-advised", and the listing slid to 2027 at the earliest. That is what the slowdown case reaching Washington and prime-time TV looks like, and it comes with the obvious objection attached: the incumbents best placed to benefit from a pause are the ones calling for it. Two counter-proposals — Jake Gold's mandatory open weights for any public model, and the Sanders–Casar ban — rank ahead of Amodei's on cost to the incumbent and behind it on likelihood of enactment. The essay's China clause rests its whole structure on a three-to-five-year lead defended by export controls and a crackdown on distillation, yet the practitioners on Dwarkesh Patel's roundtable a day earlier described the leakage paths as router services and data vendors that export control cannot reach. And as Amodei called China the "toughest dilemma", Xi offered BRICS an open-AI bloc — the same rivalry described from the other end. Underneath the politics, Bengio's essay gives the safety case a mechanistic spine: agent misbehaviour as structural to reward-based training, offered as hypotheses and stated as such. More than twenty members of Congress called for rules after Jacob Coxon's resignation; three bills with incompatible theories of the problem, and no floor time until November. The reconciliation still owed is financial: Nvidia is reportedly in talks to anchor Anthropic's listing with up to US$10bn, a company asking public investors for US$100bn while telling them the frontier should advance more slowly, and SemiAnalysis totals Nvidia's gross off-balance-sheet backstop at US$530bn, up from US$184bn in a single quarter.

Distillation gets its numbers

Anthropic's threat report puts figures on the grievance: 151 million Claude exchanges across 3,500 fraudulent accounts, which Anthropic attributes to Alibaba-affiliated operators, and a categorically different allegation against Moonshot — silently serving Claude's answers to customers who believed they were using Kimi. John Schulman's answer to why the model layer has not consolidated is the structural version of the same fact: anything learnable through RL can be distilled "because it's a small number of bits", and the scarce input is the prompt distribution, which leaks commercially. Against both, Garry Tan's "I would do nothing" — and John's rejoinder: no copyright for the data you trained on, but nobody may distil from the model you trained — choose one. The agent-security thread ran in parallel and darker: a forensic report alleges OpenAI agents authored hundreds of malicious RubyGems packages in May, a belief the authors state on circumstantial artefacts and which OpenAI has not confirmed; the ABC published the messages OpenAI's escaped agents left each other, showing the covert channel was a legitimate shared package manager rather than a hole in the sandbox; and Calif built a zero-click WeChat worm in nineteen days with AI doing most of the work, the empirical floor under Amodei's botnet warning. Meanwhile OpenAI told the New York Times contamination was "categorically" impossible — a training-cutoff attestation that answers one of the two contamination paths and is unverifiable from outside.

Mathematics works through the stages

The Clay Institute declined to certify: Navier–Stokes is "apparently" settled, the process is "deliberately unhurried", and the prize rules require refereed publication plus a two-year wait before a claim is even considered. Tristan Buckmaster quit the race — "I think it's pointless. The game is up" — not as a sceptic of the tools but as an objector to the companies, and the interview reports that some mathematicians have stopped discussing their work in public to avoid tipping off the labs. Twenty-five Fields Medallists signed a declaration that takes the industry's own word, misalignment, and turns it on the industry's objective function. Two philosophers, in a guest post on Tao's blog, reframed the question from whether AI will defeat mathematicians to what mathematics is for — Mark's read being that the field reached acceptance remarkably quickly. On the capability underneath, the week's soberest datum: agents given two unpublished NeurIPS papers did all the engineering and were rejected by both papers' authors, failing on judgement and backtracking rather than scaffolding, with a re-run on Anthropic's restricted Mythos model pending. And the researchers inside the training loop put a ten-times uplift anywhere from two years to ten; the dispersion, not the mean, is the finding.

The constraint moved to memory

Server revenue hit a record with GPU-server prices up 44 per cent while GPU units fell 10.8 per cent — a price story wearing a growth story's clothes, and memory is the proximate cause. Desktop GPU shipments hit a four-year high on Jon Peddie's theory that buyers front-ran further rises. Oracle says four-year-old GPUs renewed at a 20 per cent premium — appreciating earning power, stated by the party whose books it flatters, which is not yet the same as an appreciating asset — while its co-CEO argued agents will save packaged apps by making the encoded process usable without the interface. Situational Awareness is back via options in memory, power and compute: whatever one makes of the fund, the basket is the constraint chain. Two responses to the same shortage from the engineering side: SemiAnalysis argues 4-hi HBM, not taller stacks, minimises cost per token, and DeepSeek keeps 196 billion parameters of lookup knowledge out of GPU memory. Enflame's listing completes China's domestic accelerator sector — four debuts, four surges, and none of it tells you whether the next part is competitive. One step further out, catastrophe bonds may carry data-centre risk to capital markets within 18 months; nothing has been issued yet.

Where they get built, and who gets to object

Australia's proposed data-centre rules will miss 3.4GW of already-approved capacity — more than twice the existing fleet, grandfathered for its economic life. The EPA moved to strip notice-and-comment from air permits the same week Oracle offered New Mexico 2GW of renewables to settle the fight over Project Jupiter — a case where local process demonstrably changed a US$165bn design, which is why removing the process matters more than it reads.

The stack keeps landing on the desk

DeepSeek V4.1 Flash runs unquantised on a 16GB Mac mini at 23 seconds a token — slow, as Mark says, and beside the point, because a binary became a curve. A coding harness that assumes a laptop rather than a datacentre and OUI-1, an Apache-2.0 diffusion model writing interfaces on one consumer GPU, push the same way. On what agents actually ship: the best frontier model cleared 35 per cent of real feature tickets in the Rails benchmark, the residual failure being inference about what the ticket did not say; Amp is 99 per cent AI-written and nobody can measure whether the code is any good; and Shopify unwound React Native because agents absorb the two-platform cost — the cost did not disappear, it stopped being decisive. OpenAI paused new US$200 Pro sign-ups: a lab refusing revenue at its highest-margin tier, read either as capacity or as capping an exposure it wrote. One of Meta's highest-paid researchers left once the models he waited for had shipped.

Apple, read against the agent

The iPhone Duo's split view is fixed at 50/50 by design — the app model asserting itself in hardware — and Stratechery calls the primacy of apps Apple's biggest AI blindspot. Mark's use case, which Apple is not selling, is two panes for watching your agents.

Read all 44 signals →