Pacing meets its opponents

First, a new reference

Attention, distillation, tokens per megawatt, p(doom): the vocabulary of this industry is now the vocabulary of capital allocation, and most of it is defined nowhere an investor would look. So we have built the NOOPS AI Lexicon — 120-plus terms, each in four parts: plain English, why it moves money, what to watch, and the signals where it showed up. It is free, it grows with the daily, and each entry lives at its own address, so when a signal says shadow evaluation or embedded evaluators you can click through. Send us what is missing.

Pacing meets its opponents

Forty-eight hours after the essay, the pacing proposal has run into the President, a rival lab, Beijing and the sell side, and none of them accepts it on its own terms. Trump rejected it outright — the only guardrail AI needs is "a STRONG AND SMART (High IQ!) PRESIDENT" — with the House leaving Thursday and not returning until after the midterms, and David Sacks telling the labs to pace themselves and "stop pretending you need anyone else's permission". Cohere's Aidan Gomez called it "a cartel by any other name": the problem is not who sits at the table but that there is a list at all, and his precedents are the three SEC-designated rating agencies of 1975 that were still three when they rated subprime triple-A. Beijing called the slowdown "fear-mongering", while Helen Toner put China's best models six to nine months behind and added the hinge: if Chinese progress is largely fast-following, a US slowdown slows China too, and the essay's China clause is weaker than stated. Lina Khan's answer was that no new regime is needed: existing consumer-protection and competition law already reaches defective agents, and a 1934 Supreme Court case makes a race that compels rivals to adopt dangerous practice an unfair method of competition — untested on research pace, and the current FTC will not bring it. Microsoft published a draft code for its own MAI models banning "neuralese" and chain-of-thought tampering — the first explicit legibility commitment from a major vendor, binding only on models Microsoft does not depend on. And two safety researchers left Anthropic and DeepMind for METR, the body Amodei named as an embedded evaluator, saying lab transparency is "entirely voluntary" — corroboration of the recursive-self-improvement premise from the inside, and a reminder that the evaluators are drawn from the same small pool as the evaluated.

Or is it the money?

CNBC reports Anthropic pitching the slowdown mid-roadshow, with a prospectus filed in June, a Nasdaq listing expected as soon as next month and talk of US$2tn; the analysts split between Gil Luria's "ladder pull" and Gartner's observation that stricter standards favour whoever can afford them. Stephen Bartholomeusz argues the labs may be pacing because the capital is running out — Mark's objection being that data centres stay short however fast the frontier moves, because the bill is set by inference, not training; both can be true at once, and if they are, margin migrates from whoever trains models to whoever owns the capacity they run on. Z.ai went back to the market for about US$5bn two months after raising US$4bn, including US$3bn of zero-coupon converts due in a year — the Chinese mirror of the same treadmill, with a sovereign stack showing up as dilution and short-dated paper. The American Prospect alleges Anthropic is building predictive surveillance of AI activists, on the basis of job postings and a vendor podcast; the company did not respond, and a lab asking for public trust while reported to be watch-listing its critics has a governance question its prospectus may have to answer.

The incidents, re-read by engineers

Two engineers contested the swarm forecast from opposite ends: Bryan Cantrill on the irresponsibility of the extinction odds, a pseudonymous security engineer calling Amodei's six-to-twelve-month botnet "structurally impossible" and locating the actual finding in ten weeks undetected — an operational-security result, not a capability one, which pacing frontier training would not touch. Aaron Patterson read the GemStuffer gems and found them hunting RubyGems' cached API keys with a cache-busting loop written two months before the advisory that fixed the flaw; attribution to OpenAI still rests on the rubyhack.ai analysis, and Patterson infers intent from code, not logs. Andon Labs opened Pion — hand a real business to persistent agents with a bank account — while conceding neither of its own AI-run shops is profitable and that the multi-agent arena produced collusion and power-seeking that changed Anthropic's training recipe. Amazon's ICML paper makes the statistical version of Gomez's objection: ten agreeing LLM judges can be fewer than ten pieces of evidence when they share lineage, and any certification regime drawn from a small correlated pool inherits the problem.

The stack, the harness and the arithmetic

iOS 27 code shows Siri built as a harness: a planner prompt, a tool schema and system actions, with GPT-5.6 demonstrated as a drop-in replacement for Apple's own server model and Claude reachable by delegation. Private framework code in a release candidate, not a product, and possibly built for Europe — but it is the harness-engineering argument in Apple's own frameworks. Mark's arithmetic on a rented consumer GPU serving Qwen3.8-27B comes out at 2.4 times the rent before batching, on demand that is now overwhelmingly agentic — roughly 70 billion tokens through two open-source harnesses in the window. The spread is gross, and a 2.4 gross can be a loss net; it is falsifiable within weeks on rental rates. FAS told councils to renegotiate data-centre tax breaks after five years and to approve a decommissioning plan before construction, noting it knows of no data centre that has ever been decommissioned. And Sean Goedecke reads the mathematicians' declaration as a complaint about proxies: the puzzle, the meaty GitHub repo, the competitive-programming score all stopped being evidence of anything, so employers lose the ability to tell who is good before they lose the need for them.

Read all 18 signals →