Anthropic answers critics of its federal bill for the first time, and chip stocks bounce after an eight-day slide — on a day with no major announcements, two stories from last week got answers

There were no major announcements. Instead, two stories that had been running since last week got answers. The first is Anthropic's federal bill. It turned out to be called the "AEGIS Act," and Anthony Cimino, Anthropic's head of federal affairs, answered critics on the record for the first time. A detailed rebuttal came out the same day. The second is the market: an eight-day slide in chip stocks stopped on an intraday basis.

The two stories have something in common. Both come down to the same question: who stops dangerous things, and by what standard? In the bill, the question is how a dangerous model gets stopped. In the market, it is how lending to AI infrastructure gets priced. In both cases, the side trying to calm last week's worries gave an answer. Neither answer went far enough to make the worries go away.

The AEGIS bill: Anthropic's answer and Transformer's rebuttal

Cimino spoke to the Daily Caller. He said the thresholds of 50 deaths and $1 billion in damages were "taken directly from existing California law" and do not shield companies from liability. He admitted that the wording preempting state laws "wasn't precise enough" and said it would be fixed. The draft turns June's Advanced AI Framework into a bill and has been shown to some lawmakers' offices, he said. He did not respond by name to critics such as Heidy Khlaaf or Zvi.

On the other side, Shakeel Hashim of Transformer published a detailed critique. It makes three main points. Auditors can certify themselves and need no license, so companies could end up choosing the "least burdensome" auditor. Stopping a model requires the Commerce Department to go to court, and the court does not weigh the harm caused by not stopping it. And the bill preempts state laws while leaving the federal standards with no way to be updated. Khlaaf also wrote on Bluesky that safety depends on systems and context, and called the bill "safety theater."

Anthropic has promised to fix only the preemption wording. It has said nothing about the audit and shutdown mechanisms. Those mechanisms are where most of the criticism is aimed, so the answer and the criticism still talk past each other. Dean Ball, Bengio and Marcus have not yet responded. The bill's sponsors, its text and the 10^25 FLOP threshold have not been confirmed. OpenAI, SIF, the White House, California and New York have not responded on the record either. For details, see Anthropic says it will fix the AEGIS bill's "state preemption" wording.

On the same day, Jack Clark published the text of his speech at the IMF annual meetings in Import AI 477. It was titled "The singularity has begun." Citing rapid gains in math and on benchmarks as evidence, he called for "unusual, serious rules" on biological research and on AI doing AI research (RSI). He did not mention his own company's AEGIS bill. So in the same week, one company produced both a speech calling for strong rules and a bill that its critics call "weak." Meanwhile, Ramez Naam argued in Noahpinion that adding more agents in parallel does little to increase intelligence. We covered the gap between the two in Jack Clark, who told the IMF "the singularity has begun," and Naam.

Markets and infrastructure: a rebound, but credit is still priced high

In US trading on October 19, the SOX semiconductor index rose about +2.5% on an intraday basis (not the close; the figure comes from secondary sources). Memory stocks led, with Micron up +4% and Sandisk up +7%. The equal-weighted S&P500 fell, though, so the gains were narrow. The 10-year Treasury yield climbed back to 5.13% in the afternoon, leaving little macro support.

One trigger for the rebound was reportedly Goldman Sachs upgrading CoreWeave to "buy." At the same time, it cut its price target from $139 to $106. According to Reuters figures, the yield on the 2031 bonds fell from above 16% to 15.13%. That is still in the 15% range, and Oracle's five-year CDS remains near 300bp. For details, see Goldman upgrades CoreWeave but keeps a lower price target. Stocks have bounced, but debt pricing has not recovered.

Asia was split. The KOSPI opened down more than 3% but closed up +1.76%. The Nikkei average fell −0.62%, and SoftBank Group fell −5.3%. In funding, Bloomberg reported that Tenstorrent raised $1.5 billion in a Series E at a $6.2 billion valuation. South Korea's Elice lists on KOSDAQ on October 20. It will be the first test of demand for AI infrastructure listings since Firmus withdrew its IPO on October 9.

On power, Amazon is reportedly considering a plan with AEP and X-energy to put 12 small reactors totaling 960MW on the site of a former coal plant in Indiana. Amazon would pay the construction and operating costs so they are not passed on to household electricity bills. For details, see Amazon plans 12 small reactors on a former Indiana coal plant site. The structure is also a response to local opposition, which is starting to spread statewide. In Virginia, state Senator Russet Perry has reportedly proposed a three-year statewide pause on approving new data centers. The proposal comes from the home of the largest US market. The Memphis City Council and Loudoun County will vote or pass resolutions on data centers on October 20.

In Europe, German Chancellor Merz and NVIDIA's Jensen Huang presented a letter of intent for an industrial AI partnership (per a government announcement). The federal government will provide "hundreds of millions of euros," most of it for compute on Deutsche Telekom's Industrial AI Cloud. NVIDIA has not issued its own press release, and no exact amount has been given.

US-China lines and autonomous weapons

The US Department of Homeland Security (DHS) has ordered all federal civilian agencies to remove DeepSeek products, Bloomberg and CyberScoop reported. The order reportedly also covers other companies' tools built on DeepSeek models and bans future purchases. It is based on the Federal Acquisition Supply Chain Security Act (FASCSA), the same law used in the designation of Anthropic, which means the law has now also been used against a Chinese open-weight lab. The text of the order has not yet been confirmed and is waiting to appear in the Federal Register.

On the Chinese side, the Ministry of Commerce and other agencies issued Announcement No. 49. It added swarm (cluster) control, jamming-resistant autonomous navigation, and mixed human-machine control in low-connectivity environments to export controls on drones. At the same time, it eased procedures for some items that have become consumer products. The rules take effect on November 18 and do not name the US. The announcement came three days after Announcement No. 48, which added AI technology to the control list, and widens the controls to autonomous technology. The US and China are each drawing lines around AI and autonomous technology on national security grounds.

On weapons, Australia unveiled payloads for Anduril's Ghost Shark uncrewed submarine: one that lays up to 12 mines per mission, and one that launches swarms of Altius loitering munitions. It is the first time weapons payloads have been confirmed on a mass-produced autonomous submarine. There has not yet been a demonstration of launching a swarm from a submerged hull. US Army Cyber Command said all AI agent work is under human supervision and agents report to humans daily. A governance board will decide whether to grant authority without supervision, it said. Inside the military, too, keeping excessive agency in check has become an operational problem.

Labs and research: an anonymous model unmasked, diffusion language models, math

MiniMax released M3.1 Flash (per its blog). "Space Bunny Alpha," which had been appearing anonymously on OpenRouter, was an early version of this model. Output costs $0.60 per million tokens, and the weights will be released "later," the company said. For details, see "Space Bunny" turns out to be MiniMax M3.1 Flash.

Artificial Analysis gave DeepSeek V4.1-Flash-1017 its first third-party score. It scored 45 on the Intelligence Index, tied for second among open-weight models and one point behind the leader, MiMo-V2.6-Pro (46). Cost per task fell from $0.49 to $0.26. Its output is very long, though: the evaluation used 210 million tokens (the median is 140 million). The model is said to have 552B total and 16B active parameters, matching the model card. The DHS order was reported on the same day the model's performance was confirmed by a third party.

Inception released Mercury 3, a diffusion language model (per its own announcement). It runs at 1,068 tokens per second on NVIDIA GPUs, has a 1-million-token context, and costs $0.70 per million output tokens. The company says it beats GPT-6 Luna and Haiku 5.5 on some benchmarks. All figures are the company's own, and there is no third-party evaluation yet.

In math, Terence Tao explained on his blog the outline of the Siegel zero proof that came out of OpenAI. He states that he used SAIR's open math model to work through it. He is still "digesting" the analytic proof, he said. For details, see Tao explains the outline of OpenAI's Siegel zero proof.

In the interpretability debate, Zvi Mowshowitz took a middle position. He called Cammarata's finding that about 90% of neurons in Qwen3-32B have a single main concept "an important update," but rejected claims that interpretability will soon be solved. Leo Gao said that if you cannot simulate a neuron, you have not explained it, and Nanda pointed out that the problem of weak activations remains. Cammarata replied that weak activations are "surprisingly safe to remove." Brangus showed that a "clean" Colorado neuron also responds strongly to Tornado, the web framework. Anthropic's interpretability team has not responded.

On AI text detectors, Emily Bender criticized Vauhini Vara in "Pangram is not an oracle." In a piece for The Atlantic, Vara had used the Pangram detector to suggest that a Caine Prize-winning story may have been written by AI. Bender argues that the scores are probabilities, not evidence, that the tool has not been tested on literary works, and that publicly casting suspicion on marginalized writers on that basis causes harm. Vara, for her part, offered a control: she ran 25 past winning stories and the author's earlier work through Pangram, and all were judged to be human-written. Neither Vara nor The Atlantic has responded further.

Where nothing happened

The silences stood out more than the news. Anthropic, OpenAI and Google DeepMind made no official announcements. Amodei, Altman and Hassabis posted nothing. Day 15 of Codex's 28-day sprint is still "waiting on Tibo." As for OpenAI's "aurora," only secondhand claims that it is an image model have spread, and OpenAI has not confirmed anything.

On the AEGIS bill, Hashim and Khlaaf were the only people in the safety camp to criticize it publicly under their own names. Toner, Kokotajlo and Yudkowsky have said nothing new. On the bullish side, we found no named commentator defending the boom. The only voices were the skeptics, Gary Marcus and Ed Zitron. From defense tech, Palmer Luckey reportedly criticized Anthropic's Public Benefit Fund as something that "helps everyone except Americans" (we have not confirmed when the original remark was made). Anthropic has not responded.

Anthropic's S-1 has not yet appeared on EDGAR. Its IPO valuation and the $16.8 billion figure for July–September have not been confirmed. Microsoft has not commented officially on the Nscale matter, and Nscale has not set a price range for its IPO. None of the three companies has confirmed that the base die for Samsung's NVHBM will be made on TSMC's 3nm process. The White House and Congress (in recess until the November 3 midterm elections) have taken no new AI measures. The Mistral Large 4 weights, Kimi K3.1 and NVIDIA Bolt E4B have not come out either.

Dates to watch this week: the Memphis and Loudoun County votes on October 20; Meta's earnings and Huang's GTC Berlin keynote on the 21st; and whether the S-1 goes public. Whether chip stocks hold their intraday gains through the end of the week depends on whether yields on CoreWeave's bonds keep falling.