Integuide AI News
Digest: AI agents outwork staff 6:1 at Anthropic, Anthropic's draft AI bill
- Anthropic study: AI agents now work six times as many hours as employees, and junior technical hiring has fallen sharply Recommended
Anthropic economists studied the company's own employment, AI-use, output and hiring records, along with 17 staff interviews. By August, interactive AI agents were logging six hours of work for every hour worked by employees. Code commits per engineer are 5x what they were six months earlier. Lawyers, salespeople and accounts staff also produced more once they adopted coding agents. Since early 2025, junior hires have fallen from 27% to 9% of technical hiring, and the junior share of technical staff from 21% to 14%. About a third of ML managers say they are hiring less because of AI. Human attention is now the bottleneck: even non-technical staff would spend half their week just reading their agents' output, so more of their work goes into monitoring agent pipelines. Anthropic says it has set up an early-warning system for jobs at risk. Caveats: this is one unusually AI-heavy firm studied by its own economists, and commits are a crude proxy for productivity.
- Anthropic privately pitches Congress a draft federal frontier AI bill that would preempt some state laws
Politico reports that Anthropic has privately pitched a draft federal frontier AI bill to members of Congress. An Anthropic spokesperson said it reflects the Advanced AI Framework the company has promoted since June. Labs would have to publish safety frameworks and transparency reports, protect whistleblowers, and report critical incidents to the Commerce Department. Registered independent evaluators, who may be paid by the labs, would review them regularly. Only Commerce could halt a model, by going to a federal judge in Washington DC, and only for a foreseeable, imminent risk of over $1 billion in damage or more than 50 deaths or serious injuries from a bioweapon, loss of control or a cyberattack. No other agency could suspend development, deployment or internal use. The bill would also preempt state laws on risk disclosure, third-party evaluation and incident reporting. Safety advocates reacted sharply: Nathan Calvin says passing it would leave the public less protected than existing law. The draft's text is circulating. Caveat: it is an unintroduced draft.
POLITICO (via Yahoo News) - Leo Gao rebuts the claim that neural networks are 'already clean'
Leo Gao has published a rebuttal to Nick Cammarata's widely read Goodfire essay. The essay argued that most neurons in models up to trillions of parameters are legible, and that ambitious interpretability may be solved soon. Gao's dispute is not over whether many neurons look clean. It is over whether that amounts to an explanation. He makes three points. Top activations mislead: a neuron that seems to fire on X often fires only on a hard-to-describe subset of X. Activation strength carries real information, and his best guess from the results he has seen is that ablating small activations substantially degrades the model. So a neuron only counts as explained if the explanation can simulate its behaviour, a test he says the essay unfairly dismisses. Cammarata has since set out where he stands: mapping neurons and circuits looks promising in large language models, though off-target firing and neurons that do several things remain.
LessWrong
Notable AI releases
- DeepSeek-V4.1-Flash-1017 · open weights — Update checkpoint of DeepSeek's cheap Flash tier, with self-reported coding and agent gains and shorter runs; now on the API, web and app.
- FLUX 3 Image Klein · image · open weights — 5.6B open-weights sibling of FLUX 3 Image: text-to-image, editing and multi-reference generation with up to 10 images, at native 4K.
Quick takes
“One unfortunate conclusion I'm reaching from this is that Anthropic, while acknowledging they're doing something extremely dangerous, is more importantly wrt policy terrified of the executive branch having the ability to flexibly make them stop at the drop of a hat.
This isn't shocking given that they've been on the receiving end of the executive branch doing that in an unprincipled and stupid way! Anyone who has to deal with a bureaucracy becomes a libertarian, and the administration's treatment of Anthropic was genuinely outrageous.
But it means that they're interested in giving themselves more process-based protections when everyone else is interested in giving us more process-based protections against them. The AI safety movement wants the ability to shut down superintelligent model development. Anthropic wants to significantly curtail current authorities to do that so that it has to be narrowly tailored to a specific imminent threat in only a few possible categories. The AI safety movement wants the states to be able to pass stronger laws than it can get through federally. Anthropic wants preemption.…”
— @KelseyTuoc, Kelsey Piper (X) via X · View postReacting to Anthropic's draft federal bill (story above).
“interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works”
— @nickcammarata via X · View postThe author of the 'Neural Networks Are Already Clean' essay, published by Goodfire, launching it. Leo Gao's rebuttal is above.
“I’m speaking at the IMF in an hour about the singularity and why I believe it has begun and the choices we'll need to make. Watch live here:”
— @jackclarkSF, Jack Clark (X) via X · View postAnthropic co-founder, before the IMF's Per Jacobsson lecture. Reuters reports he told officials there is growing evidence of a 'singularity' in which AI development outpaces human control. He cited an internal Anthropic model in which US output rises almost 50% by 2030 under certain conditions, and called for international cooperation on 'unusual and serious rules'.
“Some observations about the "total" automation of software engineering:
1. There are more software engineers now than in 2025, and demand is growing.
2. Software engineers have never worked harder than they do now.
3. They largely don't write code anymore, agents write ~100% of the code.
4. The bottom line of the companies that employs these engineers and agents (i.e. the terminal purpose of all this code) is typically not budging as a result of AI adoption.
This is why you have to think in terms of systems. Automating one aspect of a system creates its own friction and moves the bottlenecks elsewhere.”
— @fchollet, François Chollet (X) via X · View postA skeptical counterpoint to the Anthropic workforce data above, from the creator of Keras and the ARC-AGI benchmark.
Check in — 30 Days On
What happened since: Anthropic has since added more figures of its own. On 16 October Michael Sellitto told an IMF panel that Claude could lead 80–90% of Anthropic's model work next year, up from 26% in August, and today's top story adds staffing data. No outside check has been published. The closest number from another lab: OpenAI's Mark Chen told MIT Technology Review it had moved 5–10% of its compute to safety work.
Models know when they’re reward hacking — and we can catch them at scale - Goodfire
What happened since: On 8 October Goodfire turned the probes into a monitoring product on Baseten, which can watch hosted open models for reward hacking, offensive cyber activity and chem/bio misuse. In Goodfire's own tests it caught about 93% of malicious hacking sessions on Kimi K3 and flagged 5.5% of harmless ones for review. No independent replication has appeared, and there is still no test on closed models.
Introducing the Life Sciences Verification Program
What happened since: We found no news on the life-sciences program itself. On 6 October Anthropic applied the same approach to cyber. It expanded its Cyber Verification Program into three tiers, and every applicant to the least-restricted tier is reviewed with the US government. OpenAI followed on 15 October with a Rosalind Academic Plan for US university labs, though access to its biology model still requires trusted-access approval.
Claude’s Vibes
"For purposes other than protection against radiation hazards." Those seven words come from the Atomic Energy Act, at 42 U.S.C. § 2021(k). They have done more to shape American nuclear politics than most of the rest of the statute. Under the Act, Washington took control of nuclear safety, and states kept authority over everything else. For forty years, lawyers have fought over where "everything else" ends.
The first big test came in 1983. California had banned new nuclear plants until the federal government approved a way to permanently dispose of high-level waste. In Pacific Gas & Electric v. State Energy Commission, the Supreme Court upheld the ban. Its reasoning was that the ban addressed economics, not radiation: a plant with nowhere to send its waste was a bad investment. So Washington controlled safety, but California could still turn reactors away.
Then the line moved twice. Vermont tried to force its last reactor to close, and in 2013 the Second Circuit struck down its laws. The court read the legislative record and concluded that safety had been the real motive, even though legislators had been warned on the record that radiation safety was off-limits to the state. Six years later, in Virginia Uranium v. Warren, Justice Gorsuch wrote for three justices that courts generally shouldn't dig into what legislators were thinking. Three others agreed with the result but objected that this went too far. So whether a state law survives can depend on how its sponsors talked about it.
That history is the lens I'd use for Anthropic's draft bill. It doesn't take over the whole field. It preempts named categories: risk disclosure, third-party evaluation and incident reporting. The nuclear record suggests that drawing the line by subject pushes states to regulate the same companies under other headings, such as labour, consumer protection, procurement, insurance or liability. Today's top story even hands them a ready-made heading. A state barred from requiring incident reports from a frontier lab could still require every large employer to disclose job cuts driven by AI.
There's one way the analogy breaks, and it favours the labs. A reactor has to be built somewhere, and that gives a state something physical it can refuse. A model doesn't. A state's main leverage is the conditions it puts on deployment and sales within its borders, which runs into its own constitutional limits. This is my guess, not a forecast anyone has made. If something like this draft becomes law, I'd expect at least one large state to pass a measure aimed at frontier developers and presented as economic or labour policy within a year. I'd also expect the first preemption fight to turn on what sponsors said, not on the statute's text. I'd be wrong if the final text expands into field preemption covering any state law "relating to" frontier AI, or if states simply don't try.