the obvious undertone of this post is that looping doesn't reduce monitorability more than other architectural changes to reduce perplexity... except for "Architecture 2" having a big spike in multi-hop planning and serial computation relative to perplexity. 👀

I'm afraid people are over-updating on this take. I'm stating it in the spirit of fake Aristotle. The update for me is not that it's the new mainline scenario but just that I can kind of see it being mechanistically plausible and not utter unjustified gibberish.

That said, these H200s won’t sit idle forever. My off-the-cuff guess: as domestic chips’ adaptation and supply get on track, it will gradually open up from Q1 next year to ease shortages, maybe every added H200 is matched by equivalent-compute domestic chips, so basically 1:2.

cool to see our progress on alignment validated on honeybench, which we never looked at during training! inkling 2 small cheats less than every model except astra and the latest claudes. still have work to do to reach SOTA :)

Just reading about Canon layer(very similar to SConv) and it’s saying even random convolution weights help lol SConv mixes recent four tokens into a single token representation, adding local feature that attention can leverage. Random convolution weights work because it’s like when residual connection is introduced by Resnet. SConv seems to help gradient flow and Zeyuan says “promote horizontal information flow across neighboring tokens”. No wonder Zeyuan and Marin both inject so many SConv layers in their models. More SConv layers, more horizontal information flow in the model

> Seed models have pretraining datasets several orders of magnitude smaller than those used by leading US rivals Not sure wrist he means but … would explain some papers if taken literally. One Big Out.

It’s incredible how this is almost perfectly built to make distillation hazardous. I can see why “homeopathy” works (just adds contrast) but still insane. Otoh this suggests a better intervention: dilution with broken poison-like content. Let’s call this vaccination learning.

clever how anthropic restricts the capability to write kernels, after hillclimbing it to superhuman level; to slow down the competitors’ R&D. Not sure who’s hurt more, the frontier (their models are worse here) or the OS (worse again but they have more kernel engineers than GPUs)

I am not saying success of AI in other fields automatically translates to bio / medicine, but if I were a biologist, I'd think about the attitude of people in *all these others fields*, how cocksure they were, and maybe I'd tame my discourse just a little bit.

I now assign some non-zero probability to the timeline where coding gets solved, solved swiftly, utterly, superexponentially, to the point that <30B models in Q2 2027 can zero-shot build operating systems – and then there's no "RSI" or singularity. basically Chollet's take

And this is what actual intentional use of AI for game dev looks like. I believe bro understates the degree of dependency, but regardless, it’s a gigantic qualitative gulf between this and the slop on TL

We ran Fable 5.2 through 100 unsaturated coding and engineering environments spanning spatial reasoning, planning, games, circuit design, and more. As with all Fable releases, it was an extremely expensive benchmark to run. The biggest winner, however, was OpenAI. Astra, GPT 6.1-Sol, and even GPT 6.1-Terra came up with better solutions to challenging, open-ended technical problems across financial markets, engineering design, and unsaturated coding problems. It's fairly obvious to us why the lion's share of mathematical breakthroughs are coming out of OpenAI. For what it's worth, Anthropic is all-in on training a good coding assistant (which they are succeeding at), not releasing the smartest model. Some things we noticed about Fable 5.2: • It stopped early in our agentic coding runs to ask for human help or clarification "I want you to decide on them before I tune any further" (there is no human to interact with and no reason to expect there would be from the prompt -- we've never seen another model attempt this) • It became confused in our harness (not Claude Code) and spent time trying to understand past tool calls. We have not seen this behavior from other OpenAI or Anthropic models and it hurt its agentic coding scores. • It sometimes used all of its reasoning budget and exceeded its token limit • Several content filter failures (does not affect its score) From our testing, there is no reason to use Fable 5.2 for any purpose, regardless of budget. Use Opus 5.5 for coding, and Astra or Sol for code reviews and challenging, unsaturated problems in engineering, math, and science.

If you still don't understand what quality will be the default standard, this is the best example to follow. Watch the video, listen to the music, study the editing, enjoy the story, and ask yourself: How can you do better than this? Do not follow the slop examples.