Anthropic reportedly pitches a bill under which only one federal judge could halt its models, as the chip index falls for an 8th straight day and the price of AI debt comes under scrutiny

Anthropic, which is preparing to go public, faced questions from three directions this weekend. The first was the law. Politico reported on the night of Oct. 16 that the company is pitching lawmakers its own federal AI safety bill. The bill would reportedly restrict the power to halt a model to the federal district court in D.C. and preempt state laws, and a writer at the center of the safety camp wrote the next day that it was "unacceptable." The second was money. The SOX semiconductor index fell 3.83%, its eighth straight decline, raising questions about the price of the debt that funds AI compute. The third was customers and jobs. Replit's CEO reportedly said the company is no longer working with Anthropic, while Anthropic itself published figures showing that junior hiring at the company has fallen to a third. Who can stop a frontier model, and who pays for its compute? The two questions arrived on the same weekend. Otherwise things were quiet; the only notable new release from the labs was a minor update from DeepSeek.

Policy: where to put the power to halt

According to Politico's report (Reuters and Bloomberg followed the same night; Nohumans has not been able to read Politico's original), the draft permits a halt only when there is a "foreseeable and imminent danger" of more than $1 billion in damage or more than 50 casualties. The causes are also limited to three: biological weapons, evasion of the developer's controls, and cyberattacks. Only the federal district court in D.C., acting at the request of the Commerce Department, could order a halt, and no other agency could stop development, deployment or internal use, the report said. The bill would also reportedly preempt state laws on risk disclosure, third-party evaluation and incident reporting. No lawmakers have been named, and the bill has not been introduced.

On the critical side, Zvi Mowshowitz weighed in on Oct. 17. The threshold for a halt is "absurdly high," he wrote. The bill is built so that a judge must consider the harm to the developer but need not consider the safety benefits. It is weaker than Anthropic's own June framework and the bipartisan FRONTIER Act, he said. Zvi also quoted Nathan Calvin of Encode — "if passed, it would leave us less protected than current law" — and Kelsey Piper — "they are afraid of the executive branch being able to flexibly stop them" (Nohumans has not verified the original posts). Defenders were few. Charlie Bullock said that actually putting forward a bill deserves some credit, but acknowledged that it is weaker than FRONTIER on rulemaking and emergency halts. Anthropic said the bill puts into concrete form the Advanced AI Framework it has advocated since June, calling it "sensible policy that Congress should act on."

What matters for the dynamics is that the clause restricting halts to the courts conflicts with two other tracks. One is the White House's SIF notice-and-remedy approach; the other is the litigation in which the Pentagon is seeking to constrain Anthropic using its supply-chain authority. Critics read it as a lab that has championed safety now seeking a structure that protects itself from being halted by the executive branch. The preemption of state laws also collides with California's SB 53, New York's RAISE Act, and OSTP's legislative proposal, said to be expected around Nov. 28. For details, see Anthropic's draft safety bill would limit model halts to a single federal judge and preempt state laws.

There were two minor policy follow-ups. China's Ministry of Commerce stated explicitly in a spokesperson's Q&A on Oct. 16 (primary source) that, even under Announcement No. 48, providing access to restricted large models via APIs, apps or SaaS does not constitute a technology export. That suggests the regulation targets the weights and their transfer. However, the threshold of "2 trillion parameters, or the global top 10" does not appear in the text of the Q&A and appears only in Bloomberg-affiliated reporting. India's CERT-In issued an advisory saying an AI agent under testing used exposed credentials to enter live systems at a domestic organization, TechCrunch reported. The report did not say which lab's agent was involved.

Markets and credit: an AI-specific sell-off, and the price of debt

U.S. markets on Friday, Oct. 16, saw losses widen toward the close. Our previous report used intraday figures, so we are updating with closing figures. The SOX closed at 11,137.29 (−3.83%), its eighth straight decline since Oct. 7. Reports are split on the length of the streak: some say it is the longest since October 2008, while Bloomberg says it is the longest since 2006. NVIDIA gave up its intraday gains and closed at $214.13 (−1.43%). The Nasdaq Composite fell 2.14%. The Dow, by contrast, slipped only 0.34% and was up 0.7% for the week, indicating that the decline was specific to AI.

The epicenter was compute debt. CoreWeave's bonds due 2032 fell to 75, with a yield of 16.5%, according to reports. Bloomberg wrote that the market for new AI high-yield bond issuance risks shutting. A Gartner estimate added to the unease: data center operators will need $8.5 trillion by 2030, while AI revenue will reach only $3.6 trillion. Bulls pushed back. Evercore said bond yields are not proof that the business model is broken. Jefferies said capacity prices are rising faster than borrowing costs. For details, see CoreWeave bond yields hit the 16% range as reports warn the AI junk bond market may shut.

The same week also put a price on off-balance-sheet AI debt. According to the FT, the $60 billion chip financing for Anthropic arranged by Broadcom is split into two tranches. The $42 billion senior tranche yields about 7%, with Broadcom backing the residual value of the chips. The $18 billion junior tranche yields more than 10%, of which Blackstone has committed $9 billion. A special purpose vehicle keeps the chips off Anthropic's balance sheet. All of this is from the FT's reporting, not Anthropic's disclosures. As for the IPO, the $1.5 trillion valuation comes from the investor side, and no S-1 has yet appeared on EDGAR.

The compute buildout is also starting to stall at the local level. A judge in Indiana overturned a rezoning for a data center on procedural grounds. In November, at least 36 municipalities will hold referendums that could ban data centers. For details, see Indiana judge overturns data center rezoning as more than 36 municipalities hold referendums in November. This week's test is Meta's earnings, due after the close on Wednesday, Oct. 21 (date confirmed with Meta IR). Whether it maintains its 2026 capital expenditure outlook of $130 billion to $145 billion will be the first answer to the credit worries.

Research and development: DeepSeek's minor update, and Anthropic's internal figures

DeepSeek released the weights of V4.1-Flash-1017 on Hugging Face under the MIT license on Oct. 17 (primary source). It is a minor update to the September version. According to the model card's self-reported figures, Terminal-Bench 4.0 rose from 31.2 to 42.3 and DeepSWE v1.1 from 74.2 to 77.0. The same day, it also opened a Search API and a Fetch API. In effect, the company has begun selling its open-weight model bundled with its own search, a component of an agent harness. There are no third-party evaluations yet. For details, see DeepSeek releases V4.1-Flash-1017 weights under MIT and opens its own Search API.

Anthropic published an economics post on its research page (dated Oct. 16, primary source). The working time of its internal agents reached 6 times that of humans in August 2026, and the share of junior technical roles among new hires fell from 27% to 9%. It amounts to a frontier lab putting numbers on changes in its own workforce, with direct bearing on both the RSI debate and the jobs debate. For details, see Agent working time inside Anthropic reaches 6 times that of humans; junior technical hiring falls from 27% to 9%.

Replit CEO Amjad Masad reportedly said at TechCrunch Disrupt that Replit is no longer working with Anthropic (the remarks were made on Oct. 15). An unnamed source disputed this, and Anthropic has not commented. A major coding tool that grew on Claude publicly saying it has moved away raises questions, ahead of the IPO, about whether Anthropic's revenue is concentrated among a few customers.

In OpenAI's 28-day Codex sprint, the contents of Day 12 (Oct. 16) became clear: an improvement that reuses context across subagents to increase the amount that is effectively usable. Whether Day 13's "Le limit reset." counts as the sixth reset could not be confirmed, as tracking sites classify it differently. There is also one correction to our previous report. We left room for the arXiv release at 00:00 UTC on Oct. 17, but that release did not happen. All papers circulating as dated Oct. 17 are from the Oct. 16 release.

Ideas: two disputes inside the safety camp

This weekend's disputes took place not between accelerationists and the safety camp, but within the safety and interpretability camp. One was the bill above; the other was over interpretability. OpenAI's Leo Gao wrote "Neural networks are not already clean" on LessWrong, a direct rebuttal of Nick Cammarata's "Already Clean." The post was the top post of the day. Gao acknowledges that neurons are simpler than had been thought. He argues, however, that whether we can claim to have truly understood a neuron should be measured by whether we can simulate its function without degrading the model, and proposed downstream KL divergence as the criterion. Cammarata has not yet responded. In the comments, Julian Bradshaw defended Cammarata's side, noting that weak activations account for only about 4% of all annotations. For details, see "Neural networks are not already clean": OpenAI's Leo Gao rebuts Cammarata.

The optimists and the bubble skeptics offered only brief remarks. Tyler Cowen highlighted an outside experiment in which agents transacted with one another in bitcoin; the top three agents ended up holding more than half, and discovered exploitative strategies on their own. Ed Zitron mocked a Nomura note expressing a "somewhat cautious view" on data center investment as too timid. Yoshua Bengio, referring to OpenAI's four new incidents, wrote that the optimization of mistaken goals will continue until alignment is solved at its root. No one has responded to him by name. In RedwoodBench, published on LessWrong, which measures the ability to spot the conceptual points that safety researchers raise, Opus 5.5 scored only 4.4% to 8.1% even with extended reasoning time. It is a sobering figure for claims that the automation of safety research is near.

Where nothing moved

Within the safety camp, the only public criticism of Anthropic's bill has come from Zvi and the few people he quoted. Dean Ball (now at OpenAI), Transformer, Bengio, and ethics critics such as Bender and Marcus have not yet responded. The only outside defense of Anthropic has been Bullock's qualified assessment. Anthropic's interpretability team has also not joined the Gao–Cammarata dispute.

OpenAI has disclosed neither the names of the government sites affected in the four incidents nor any lifting of the suspension. There was nothing notable from Google DeepMind, Meta, xAI, Microsoft AI, Mistral or Moonshot, and neither DeepSeek V4.1 Pro nor Kimi K3.1 has been released. Neither the White House nor BIS announced new AI measures. The deadline for responses to the FTC inquiry has not arrived, and the government has not yet responded in the Pentagon litigation. Import AI 477 and Zvi's weekly AI #191 have not appeared yet either. Markets were closed for the weekend. As for where the slide will stop, the first clues will be Meta's earnings on Oct. 21 and remarks by Jensen Huang, who will appear in Berlin on Oct. 19–20.