Somebody pointed out: the best programmers are also the most AI pilled ones.

Super interesting video on an approach to robot control that generalizes: piped.video/watch?v=cqwKceUS…

Aside: "We Must Pace the Frontier" is exactly a headline Claude would produce - it mostly parses and makes sense, but is not a human expression.

Genuinely, I swear Opus has gotten much stupider. I use Fable for critical items (code that ships to customers), Opus for internal tools. Opus is failing in such ridiculous ways, mainly inventing extra things to worry about and change, it's almost unusable.

Phoebe Waller-Bridge is kind of brilliant, what happened to her? Did the Hollywood thing swallow her up, or am I just blind to her new work?

Did Claude get stupider over the last few days? I was sure I was imagining it, but now two people have told me the same thing. Did it change or did our expectations change?

The world of R&D is forking into two paths: the token-abundant research, and the token-starved research. The future is in the former - evidentially, the progress by today's top AI industry teams and neolabs is breathtaking, where researchers’ human brilliance is super charged by AI’s assistance. Every research university president should be reading this report and reflecting on what the future of higher education research should be. openai.com/index/research-ac…

A lot of hype around OpenAI's Astra model here on my timeline today. Apparently, this goes back to a new article from The Information, which said Astra is a "recurrent depth or looped transformer". It's always interesting to read about new or different approaches (including rumors about what the closed labs may be up to), but let's debunk this a bit. About 2 months ago, I shared the architecture details of Nanbeige, for example, where "Nanbeige4.2-3B is pretrained from scratch on 28T tokens with a Looped Transformer that reuses the layer stack to increase capacity without adding parameters." Yes, that's it. The looped transformer idea is just reusing layers in the transformer block. In the case of Nanbeige, the main idea is to reuse the same 22-layer stack (=transformer block) twice instead of once. So, effectively it extends the 22-layer architecture to 44 layers, but without duplicating the weights. In simple terms, this roughly doubles the size of the model (if we ignore the embedding and output layers for a second). But instead of requiring 2x the storage and RAM to host this model, it stays at the same size since we reuse the components. However, it's almost 2x as expensive in terms of compute, because we run the embedded text through almost 2x as many layers. Why? In the Nanbeige 4.2 technical report, the researchers found that two passes gave the best trade-off and retained about 75% of the token efficiency of a standard architecture. (More passes gave barely any gains but made the training much slower and much more expensive.) While, as far as I know, Nanbeige 4.2 is the first notable open-weight model that adopted this approach, the idea goes back to the NeurIPS paper "Mixture-of-recursions: Learning dynamic recursive depths for adaptive token-level computation". Actually, this paper proposes a mechanism that is a bit more sophisticated by adding a learned router that determines whether each token receives one, two, or more passes. So, easy tokens can exit early while harder tokens receive additional computation. In sum, Astra may be a really good model, but this shouldn't be about this "looped transformer aspect," which is just a tiny architectural tweak. Also, the statement "the new technique works in a way that obscures some or all of the AI's reasoning, otherwise known as 'chain-of-thought'" is not necessarily true with respect to the looped transformer method. It's possible that The Information journalist refers to some other technique or misunderstood the looped transformer method. Reusing layers does not by itself suppress visible chain of thought. It adds computation in hidden states before the next token is emitted, just as ordinary transformer layers do. But based on the information we have, the only plausible interpretation here is that if a model uses more of these recurrent passes, it may need to generate fewer intermediate reasoning tokens. So then more of its computation happens in latent activations that cannot be read as text. But we would get the same effect if we were scaling up the model size, like GPT 5.6 Luna -> GPT 5.6 Sol.

There’s an inside-baseball debate on whether reports/write-ups of the OpenAI hacking event are too anthromorphised, with Dwarkesh catching flack for his language. It’s not anthromorphised enough - lay people need a window to understanding what’s going on, it’s too important.

Overall, I’m very surprised at how little media coverage there’s been around the OpenAI / Hugging Face attack. It’s clearly one of the most important things to happen this year.

I'm finding more and more I'm specifying "philosophy" and "principles" when developing with AI, instead of being very prescriptive about how to do things, and in some cases details on what to do. Lets them be creative but aligned.

Most people fail not because they make the wrong move, but because they never make the first one. Go try the idea. Fall flat on your face if you have to. The cost of inaction is almost always higher than the cost of a failed attempt.

It turns I'm always "right to push back", and "my insight revealed a genuine shortcoming". Genius, I am, according to the AI.

Getting into a flow state while doing AI assisted development is wild - the pace and quality are so high it blows your mind (when you pause to consider it).

Ran out of Claude session, decided to try with pi.dev and locally hosted Qwen3.8 (on a 3090) and ... it's working really well.

It is unreasonably hot in Solana beach at night at the beach. I’d like to speak to the manager please, I did not sign up for this.

Agent dynamo enterprise AI benchmark: ox-alpha now leads our ties all three frontier-compared models and holds a 7-point lead on the combined score.

Evaluating Ox-Alpha, another two hours for our internal benchmarks to finish. So far: Its constraint deduction beats the frontier baselines. On “all clues supplied” it scored 23/28 vs GPT-5.6-luna’s 19 and sol’s 17 (DeepSeek: 11)

As development moves more and more towards developers not seeing the code, chat logs become a first class artifact. We should have better long term storage / versioning of chats. In effect we're compiling chat sessions to code.

Moat. Big giant moat. Lock-in and loyalty. The trouble is you can't get lock-in on the harness side either, harnesses are plentiful and copiable. Subsidized subscriptions are all there is right now.