Good morning,
This week we cover the return of AI doomerism. A young researcher quit Anthropic, Dario Amodei asked the whole industry to slow down, and we look at who said yes, who said no, and what actually changed. Plus the quick bites, the researcher himself on The Daily, and a model that decides instead of writing.
AI doomerism is back: who wants to slow down, and who does not
Many things happened over the last two weeks, but the most noticeable was the return of AI-doom fears in Silicon Valley. It started when Jacob Coxon, a 27-year-old researcher who worked at OpenAI and Anthropic, resigned from the latter and shared a clear message online: neither company is acting responsibly, and both are "gambling with our lives".
His message set off a series of responses from the top of the frontier labs. Anthropic's CEO, Dario Amodei, published an essay with a three-part plan: outside evaluators with permanent, employee-level access to the labs; coordination between labs in democratic countries on safety standards and on the rate of progress; and, eventually, coordination with China. Amodei's view was supported by Sam Altman (OpenAI), Elon Musk (SpaceXAI) and Demis Hassabis (Google DeepMind).
The "slow-down" movement did not get full support, however. For a start, smaller competitors such as Canada's Cohere were unhappy: its CEO, Aidan Gomez, argued that safety rules should not be written by a few Silicon Valley companies as this would place even more power in the hands of the largest US labs. Predictably, President Donald Trump disagreed as well. He offered a much simpler solution: all AI needs is a "STRONG AND SMART (High IQ!) PRESIDENT".
But perhaps more interestingly, Nvidia's CEO Jensen Huang does not share the doomer view either. Two weeks ago he was asking the G20 not to regulate "theoretical harms". Speaking at Salesforce's yearly conference, Dreamforce, he said that safety is an engineering problem: companies should run as fast as they can and, if a product does not feel safe, "take a pause and make sure you get it right".
China gave the same answer. In Shanghai, Huawei's chairman Eric Xu said Chinese labs are not yet advanced enough to see the risks the Americans describe, and that they therefore "need to speed up". Beijing called the American warnings "fearmongering".
And the money? It did not slow down. In the same week OpenAI was reported to be in early talks to raise at $1.2tn, up from $852bn in March, and Nvidia to be considering an investment in Anthropic's IPO of up to $10bn, at a valuation of ~$2tn. Amodei himself is careful to write that pacing does not mean pausing.
What does this mean? Once again, labs and researchers are flagging their worries, yet once again, they are not taking any real action, and the underlying business continues to boom. Huang's view that if they believe they are too dangerous they should stop might be true; however, it seems oversimplified. Perhaps similarly to oil giants and the climate change problem, it is an everyone-or-no-one problem. If Anthropic decide on their own to slow down, it does not mean that the whole of tech development will. Even if two, three or four companies slow down, as long as one does not, the tech will keep on developing and eventually the one that stays will outpace the ones that slowed and decide how to shape the industry. This raises the question: Can the AI arms race be stopped?
The next episode is on Thursday 24 September, when Trump hosts Xi Jinping in Washington with AI guardrails on the agenda, and Altman, Huang and Cook at the state dinner.
Quick bites
Deals & money
Anthropic signed a $13.7bn, six-year compute deal with Trump-linked Rum Group, on top of the $80bn signed last edition.
New valuations in talks: agent startup Instinct at ~$10bn, defence-drone maker Shield AI at $20bn or more, delivery-drone firm Zipline at ~$20bn.
Apollo is expanding its startup investments to finance AI hardware, while Bridgewater's co-CIO Greg Jensen called for AI-compute giants to be regulated like big banks.
Meta launched Meta One subscriptions, from $2.99 to $499 a month, with 15m subscribers so far.
Models
TypeSafe AI, a stealth startup founded by an ex-OpenAI researcher, launched a new type of model focused on decision-making: Jev (see the Tools to try section for more information).
A new Chinese lab founded by a Tsinghua professor is expected to release a frontier model soon after being valued at $1.4bn.
Salesforce launched Koa, its first reasoning model, built on Nvidia's open Nemotron 3 Super.
Enterprise & security
Nvidia, Palantir and Booz Allen restricted internal use of Anthropic's models over data concerns; zero data retention is not yet available on the new Fable models.
A zero-click flaw called Plugin4Shell affects Claude Code, Codex, GitHub Copilot and Gemini CLI. Claude Code and Codex are patched; Copilot has no complete fix and Gemini CLI is being retired.
Microsoft plans to triple its cloud capacity by 2032, with $50bn of capex this quarter and more than 30m paid Copilot seats.
Insurance corner
RAND's report on the insurability of AI finds the market split three ways: a few carriers cover AI losses (Munich Re; AIUC up to $50m), a growing number exclude them, and most are silent. W.R. Berkley now excludes AI from D&O, E&O and fiduciary cover, and Verisk's optional exclusions have been available since January.
Chips & geopolitics
Huawei announced the Ascend 960DT (Q1 2027) and 960PR (Q3 2027) and said it cannot meet demand in China's ~$50bn AI-chip market.
SK Hynix is in talks with Intel to make memory chips in the US, which would be its first American memory production.
Congressman Ro Khanna wrote to Alibaba, DeepSeek and Moonshot asking them to help pace frontier AI, with an emergency hearing planned for the week of 23 September. According to Pew, roughly two-thirds of US adults think AI is advancing too fast.
Science & society
OpenAI published an AI-generated solution to the Navier–Stokes problem, a $1m Millennium Prize Problem, produced by ~10,000 agents in 88 hours; the Clay Institute has not yet accepted it.
25 Fields medallists, including Terence Tao, declared the goals of AI companies and of mathematics "severely misaligned", and OpenAI withdrew from a Caltech AI-maths event.
ICLR capped authors at 20 papers each after submissions reached 19,525 last year, and submissions for this year passed 60,000; about 20% of them had no author qualified to review.
Entry-level jobs: the Dallas Fed finds young workers' share of employment down from 16.4% to 15.5% since ChatGPT, while the ECB finds intensive AI adopters ~4% more likely to add staff.
Stanford's "virtual biotech" ran thousands of AI agents on drug discovery; they proposed an antibody-drug-conjugate strategy that a pharmaceutical company later validated.
The interview: the researcher who started it, in his own words
This edition's pick was obvious: The Daily from the New York Times, The A.I. Researcher Whose Rebellion Is Changing Everything (14 September, 30 min). Natalie Kitroeff from the New York Times interviews Jacob Coxon, the former OpenAI and Anthropic researcher who ignited the new slow-down wave. He explains how he arrived at his conclusion, including his work at the labs, and expresses his fear for the future.
I would listen less for whether he is right about superintelligence than for the strategy question underneath: what do people inside the labs say in private that their employers do not say in public, and what would a slowdown look like from the inside? He describes colleagues talking about "crunchtime" and "endgame". Worth hearing.
Tools to try: Jev, a model that decides instead of writing
While the labs argued about pace, TypeSafe AI left stealth with $40m of seed funding and a model that generates no text at all. Its founder, Diogo Almeida, worked at OpenAI on RLHF, the training method behind ChatGPT; Jev is his bet in the other direction.
Jev answers closed questions. Give it a case (an email, a claim file) and a question with fixed answers (which team? urgency, 1 to 5? fraud referral, yes or no?) and it returns the answer with a probability. TypeSafe claims up to 193.6x faster and 444.6x cheaper than an LLM, at $42 per billion input tokens, 238x below Fable 5.1.
To try it: (1) join the waitlist, or ask a developer to call it through OpenRouter; (2) pick one high-volume decision, such as claims triage or inbox routing, and write it as a closed question; (3) run a few hundred past cases with known outcomes and check its accuracy against your current model, and whether its 90% answers are right nine times in ten. Trust me, it seems harder than it is. Once whitelisted TypeSafe gives you access to its MCP, and Claude can do all the work for you.
Two caveats. There is no chain of thought to read, and a probability is not an explanation, so keep it away from decisions you must justify to a customer or regulator, such as declining a claim. And every number above is the vendor's own: "zero hallucinations" means it cannot invent an option, not that it cannot pick the wrong one.
Given the novelty, now is the time to test and experiment, not to put it into production.
Have a good week,
