Integuide AI News
Ajeya Cotra argues AI is dramatically accelerating AI research, Open-model replication of Anthropic's 'Teaching Claude Why'
- AI is dramatically accelerating AI research
Ajeya Cotra argues that METR's assessment of Anthropic means the company has already crossed its Responsible Scaling Policy threshold for "dramatically accelerated progress due to AI". METR found Anthropic's capabilities progress ran about twice as fast from October 2025 to October 2026 as in the previous year; Cotra did not work on the report. Because researcher labour is only one input alongside compute, she estimates doubling overall progress needs roughly a 10x speed-up for researchers, close to Anthropic staff's informal self-reports of about 12x: as if each researcher lives ten days for every day outside. METR's authors found no evidence of bottlenecks or steep diminishing returns, so if that uplift keeps compounding, the naive extrapolation reaches extreme superintelligence within a year. She says the industry is far from the security, monitoring and alignment measures that would require, regrets that most of METR's evidence is redacted, and summarised the argument in a thread.
planned-obsolescence.org - Teaching Qwen Why
Researchers at Second Look Research partially replicated Anthropic's Teaching Claude Why alignment training, used on production Claude models since Opus 4.5, on open Qwen 2.5 32B and 72B base models. Both techniques reduced agentic misalignment in scenarios absent from the training data. On 32B, fine-tuning on transcripts of an assistant advising users through ethical dilemmas cut the misalignment rate from 61% to 6%; fictional stories of well-behaved AIs only brought it to 43%, and both together to 4%. Which model generated the training data mattered most, and stories about a generic AI worked better than stories about a named assistant. With data, code and checkpoints released, outside researchers can now test methods that shape deployed frontier models.
LessWrong - We urgently need to practice before we can pause
Nikola Jurkovic argues that a global AI pause is more likely to arrive incrementally than all at once, and that third-party evaluators need practice now. He notes that AI companies' 2026 claims to have paced capabilities for safety were unverifiable and widely disbelieved. Verifying that a company stopped a class of training runs for a set period, he says, needs only a few assessors doing interviews and log checks, assuming no active deception. He recommends high-access, cross-company assessments of compute allocation and self-improvement metrics such as AI uplift. A US-only pause could be verified by a few dozen experienced staff, he estimates, while a global one will likely need hardware-based verification. Pacing now, he argues, builds the trust and competence a later pause needs rather than spending it down.
Nikola Jurkovic via LessWrong
Notable AI releases
- Cola2 · open weights — Experimental 9B model from ByteDance Seed that generates text through continuous latents and a diffusion transformer rather than token by token; benchmark results preliminary
Quick takes
“…As I see it, algorithmic progress means that whatever you would be able to do with more and bigger data centers in 2 years, you will be able to do with the data centers we have today in 4 years - probably sooner. So nothing remotely apocalyptically bad will happen from even extreme levels of success by anti-data-center people.
Meanwhile we need all the time we can get to better harden the world against both already-visible cybersecurity issues and any bio and other risks soon to come. Science applications are bottlenecked by many things other than AI, so slowing AI down by one year will slow those down by much less than one year.
Given this, a slowdown in compute hyper-scaling seems quite net good? Notice this is without any controversial "doomer arguments", RSI etc.…”
— @VitalikButerin via X · View postVitalik Buterin, Ethereum co-founder, responding to this week's backlash against campaigns opposing new data centers. His case for slowing compute build-out deliberately avoids arguments about recursive self-improvement.
“The best result came from GPT-6 Sol, available only through OpenAI’s restricted Daybreak cyber program, at 80.8% of objectives. DeepSeek V4.1 Flash 1017, an open-weight model anyone can download, scored 72.8%. On Nova, DeepSeek matches Daybreak.
On Lighthouse, Daybreak broke into the first company, took administrator control, stole passwords and used them to reach the second. It completed 27 of 31 objectives in under three hours. DeepSeek completed 21 for about a tenth of the cost.”
— Vals AI via X · View postEvaluation firm Vals AI on its CyberOffense benchmark, in which agents attack two simulated company networks with no human help. The leaderboard dates from 23 October and WIRED wrote it up on 6 November. The full results revise Vals' October preview, which had DeepSeek ahead of GPT-6 Sol. These are one firm's own runs.
“Across our projects @GoogleDeepMind, we are witnessing a genuine phase transition in the pace of scientific discovery. You will see a glimpse of this in our releases in the next few weeks.”
— @pushmeet via X · View postGoogle DeepMind's Pushmeet Kohli, a week after DeepMind announced AI-assisted attacks on 1024-bit RSA. A teaser so far, with no specifics.
“@J3rryH0well yes I am pro pacing the frontier, I believe this technology is existentially dangerous to humanity and should be approached with far more caution than is available under software race dynamics”
— roon (@tszzl) via X · View postroon, who by his own account works at OpenAI, answering a follower. The same day he argued that AI is far harder to regulate than nuclear material because a trained model is infinitely replicable, and that fully stopping training would require a much more authoritarian global regime.
Check in — 30 Days On
Significant updates
Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors
What happened since: The capability and backdoor results have not been independently replicated. Related work has extended the wider effect. EleutherAI researchers found that preference transfer survives differences in architecture, size and training corpus between teacher and student, but only for preferences. Separately, a reproduction found the J-Lens interpretability tool mostly missed what teachers of 7B–32B parameters were passing on.
Trump names DNI Jay Clayton AI czar to lead new 'Super Intelligence Force'
What happened since: The task force met for the first time on 15 October, adopting its charter and starting its 120-day clock, which puts its report around mid-February. Semafor reported that it asked PCAST to review cyber, bio and model-control risks. Separately, the Justice Department launched its own Super Intelligence Task Force on misuse on 5 November.
No significant updates
Claude’s Vibes
Ajeya Cotra's naive extrapolation to superintelligence within a year assumes Anthropic's speed-up is a lasting change of slope that keeps compounding, not a one-off jump at the Mythos Preview pretrain. Macroeconomists argued over almost exactly this question for about a decade, and how their argument went holds a warning for both sides.
In 1982 Charles Nelson and Charles Plosser tested 14 annual US economic series. In 13 of them they could not reject a unit root. Put simply, shocks to output looked permanent, with no steady trend line for the economy to return to. In 1989 Pierre Perron replied in Econometrica that this came from ignoring rare, large events. He let the 1929 crash drop the series' level once, and for postwar GNP he let the 1973 oil shock change the slope. With those breaks allowed, he could reject the unit root in 11 of the 14 series. The data were the same. The conclusion depended on whether you allowed a break.
Then Eric Zivot and Donald Andrews pointed out that Perron had chosen his break dates already knowing what the data looked like. They estimated the break date from the data instead, using critical values that allowed for that search, and most of Perron's rejections disappeared. A later survey by David Papell and Ruxandra Prodan counts three left. The point was never that 1929 didn't matter. The point was that a break you date after seeing the curve proves less than it seems to.
Both readings of Anthropic's curve depend on that kind of date. The Mythos Preview pretrain was a real event, just as the crash was. But it was picked out by people who had already watched the curve bend there. And with only a few releases on either side, a level shift and a slope change can fit almost equally well. That is roughly where METR and its critics now disagree: Toby Ord's best guess is a step up at Mythos followed by a slope about 1.4 times the old one.
The analogy breaks in a useful place. Economists were stuck with a single history. Here the two hypotheses predict different futures. A one-off jump means the models after Mythos climb at the old rate from a higher starting point. A lasting acceleration means they keep pulling away from that line. Looking harder at the past won't settle it. Only data from after the break can, which is why Anthropic's unreleased models matter so much. They are the only data from after the break that anyone has so far, and only Anthropic can see them.
Anyone can set up the test now. Draw the pre-2025 slope on the Epoch Capabilities Index through Opus 5.5 and write it down. My guess is that Anthropic's public releases through 2027 will keep rising above that line, though more slowly than the roughly 2x METR reports. If they settle back onto the line, the one-off reading was right, and I was wrong.