This is great data, and good but somewhat flawed analysis.

The good part is showing that the drop in thinking tokens persists no matter what grouping you slice across. They make a very persuasive case that there's something systematic going on.

My usual complaint about these "they're nerfing the models, I feel it in my bones!" posts is that they don't account for the workload changing. From working on my own stuff, there are a series of evolutionary/de-evolutionary changes that happen in a heavily AI-written codebase. Initially everything goes great. Then the AI takes on too much technical debt. Improvements slow down and regressions creep up until it becomes a never-ending game of whack-a-mole just to keep up. So you direct some (probably AI) effort towards cleaning things up, reducing duplication, and removing patches for problems that are better fixed with a design change, or workarounds because the harness saw the wrong version or you incorrectly described a problem and it strenuously solved a non-problem. That gets you back up to cruising speed for a while, then the project exceeds some hidden threshold for size in latent space or something, and further progress has to rely on attending to one aspect of the codebase at a time. Once again, the architecture becomes the limiting factor, but in a subtly different way. My sense is that it all boils down to some sort of "attention capacity" -- is your codebase and problem space amenable to looking at one aspect at a time, or is it all snarled together? -- but that's an essay that I'd love to write but really don't have enough experience to do justice.

Anyway, the details don't matter. The point is that not only can you not assume that the difficulty presented to the AI is roughly constant over time, but also there's evidence to believe that it will be normally be increasing. (Unless you're constantly starting new projects instead of continuing old ones.)

That's why I like this writeup. Focusing on thinking trails doesn't eliminate the problem of snowballing difficulty, but it does sidestep the worst of it. In fact, I'd expect the same setup to think more as the complexity/sloppiness creeps up.

The flawed part that bothered me was that it feels like there's a little bit of a predetermined conclusion that thinking is a magic sauce that makes everything taste better if you spread it on everything. I want a high variance on thinking, especially between interactions. A smarter model would have a higher variance, in my opinion. So the accusatory tone (perhaps I should reread it? My first impressions are often wrong) around "look! it doesn't bother to think at all a lot of the time. That can't be right!" seems misguided to me. It should think when it needs to, and if its thinking was clear then it won't need to re-think over and over again; it's all still in the context.

Forgive the anthropomorphization, but consider those studies of chess experts vs novices. Novices have to work way harder, working through all kinds of things from scratch, while the expert instantly and effortlessly recognizes what's going on.

But anyway, the main takeaway fully survives this criticism. The models appear to systematically think less over time. It doesn't matter if a smarter model might be able to think less for the same quality; this is happening over the same model.

This is good feedback. Not sure if you read the long-form article vs. the tweet-thread linked here, but I did attempt to address most of the criticisms you listed in that writeup, if you haven't already read it. It's definitely not written in the same tone as the for-broader-publication thread.

Regarding the amount of thinking as "magic sauce": the main issue is that even in the right tail, the delivery of thinking tokens almost never reaches the levels of published benchmarks. You once could include the word "ultrathink" in any prompt and it would provide a fixed thinking budget of 31,999 tokens. Whereas, in my corpus only 46 out of 36,374 invocations broke 16k thinking tokens, and the P90 was only 2,207.