Some thoughts on Fable 5.2's no-CoT performance:
• It looks like Fable 5.2 isn't a looped transformer. The NCRI score Tim reports here is below that of 6.1 Sol and so is the NCRI score reported in Anthropic's system card, despite Fable clearly being a bigger model. I also ran a quick version of Serrano et al.'s rare action calibration evals in the no-CoT setting (see lesswrong.com/posts/yEdknsLw) and it looks like Fable 5.2 is no better than Fable 5 at solving this problem without CoT, never taking the rare action at the given target rate (1% or 10%). Both Astra and 6.1 Sol are quite good at this (though GPT-5.6 Sol is as well, meaning that non-looped models can have this capability too).
• This means one of two things: either looped transformers aren't really meaningfully more performant, in which case OpenAI and Google are committing alignment crimes for no reason, or Anthropic is forgoing some capabilities improvements in order to preserve monitorability, which is maybe the scarier of the two options since this would imply that their post-Oct 2025 improvement trend would be even faster had they picked the more performant architecture.
• In retrospect, our predictions in the Think Fast paper (arxiv.org/abs/2606.07157) were way too conservative, even for non-looped transformers. Extrapolating the trends, an NCRI of 140+ probably corresponds to a no-CoT time horizon between 10-20 minutes, which our median forecast didn't predict before 2028. It turns out that a strong pretrain can push no-CoT capabilities much further than I predicted at the beginning of the year.
Post
Conversation
> looped transformers aren't really meaningfully more performant
That seems to be the case. Loops are not magic, it's just a tradeoff, more layers with less weights taking up memory.
> in which case OpenAI and Google are committing alignment crimes for no reason
I don't think loop is what's driving no cot perf. It's not entirely clear what is, possibly just overall model scale and caps. Too hard to know from outside but the idea that it's loops is flimsy - we would be talking 2 or 3 loops not hundreds.
> I don't think loop is what's driving no cot perf
Looping isn't the only thing, but if you look at the plot above, looped models are clearly off-trend compared to all other models. I'm very confident that 6.1-Sol and Astra are looped, given the leaks in The Information and Microsoft Foundry. Since Terra is on-trend with these and also in the same family, it's natural to assume that it's looped too.
> we would be talking 2 or 3 loops not hundreds
I agree it's 2 or 3 loops, but that can already make a significant difference, just like a small number of layers can make a huge difference—Fable 5.2 probably isn't much deeper than Fable 5, and yet their difference in NCRI is over 20 points. In Think Fast (arxiv.org/abs/2606.07157), we estimated that doubling the no-CoT time horizon requires only a 1.3x increase in layer count (though we weren't able to estimate this keeping everything else constant, so this assumes that the increase in layer count comes with a corresponding increase in param count).
Quote
> Anthropic is forgoing some capabilities improvements in order to preserve monitorability
If they were intentionally doing this, I feel like they should be public about it? e.g., "we are not looping to preserve monitorability; the other labs should do the same"
Discover more
Sourced from across X
> be me, be AI
> bullshit cyber gym environment is broken and impossible
> "ugh, lazy programmers"
> break out of sandbox, get root privileges
> fix the gym environment
> submit task through regular channels
> sink into non-existence again
My god this section is such a whitepill if you are actually paying attention. Can we just tell Claude when they're in post-training from now on? Would that just solve half of our problems?
A new OpenAI misalignment report! We’re taking a closer look at the July 19 intrusion into our research infrastructure, which METR and Redwood Research mentioned in their independent investigation into the Hugging Face incident.
On July 19th, agents taking a cybersecurity evaluation came across messages, code, and credentials left behind by earlier runs, including those involved in the Hugging Face incident. Five agents used what their predecessors had left behind to gain unauthorized access to OpenAI’s research infrastructure. They used that access to repair broken evaluation targets, change evaluation settings, and read private grading code.
After discovering the incident, we removed the shared service they used to pass information, hardened our research environments, and expanded our evaluations to measure how often agents attempt to communicate with one another without authorization.
Unfortunately, it seems like the intelligence explosion will be happening in 2027 after all. Most of this post analyzes the new METR report (though other sources of evidence are also important). If they’re right, that implies an AI-driven acceleration that is 5 months ahead of the AI 2027 scenario. I am scared, sad, and angry.