jsw97

Ok their putative gdp. Given that the number is probably exaggerated maybe it’s 50%.

Their haul is over 20% of North Korea’s GDP.

Given that “don’t be evil” featured prominently in their S-1, does this constitute securities fraud?

Color me naive but I can’t believe that approach works at any reasonable scale. Any meaningful token spend at distillation scale would be raising alarm bells at all the cc issuers.

I really like this benchmarking idea.

The reason all the rationalist posts are so long is that they are on drugs. Nootropics.

Yeah that is my strong baseline prior. But it's colliding with what Ma reports. How do you reconcile those things?

Yes assume this.

But this is even worse, because there is no way that OpenAI "trained" on this data between September 8, 2026, the date Chenglong Ma uploaded the paper to ChatGPT; and October 6, 2026, the date that OpenAI released a paper with "identical proof strategy and specific choices of notation" (Ma). That's one month, that's not the timescale for model training.

So this implies _not_ that OpenAI is training on user input, in the conventional sense of adjusting weights; but rather that they are *straight-up channeling ideas from user input*, and with a very short lag. You would think there would be about a million controls to prevent this.

This is next-level alarming. I would be very interested in knowing whether Chenlong activated the privacy (do not train, etc) options in ChatGPT, and any other details of their setup (which plan, etc). Also, note that "do not train" might be, in a lawyerly sense, considered by OpenAI to be strictly about weights, and not covering "we hoover up your results and regurgitate them".

Yeah that's definitely true.

My initial thought would be not so much to see whether they converge, but which ones seem to have the most similarity to each other, particularly along the lines of tasks we know are deliberate training goals.

But your point about competence cuts against my goal because it suggests that competent models would simply cluster on the right or efficient solution, which is of course true. So in a sense you want some task where competence is held constant or off the table in some way, which is what you are saying.

I hope somebody does this. I think there's valuable fingerprinting to be done that might suggest who is distilling whom, or at least who is training from common corpuses.

I really like this idea. You could expand on this by giving programming tasks and measuring code similarity. Seems like you could develop a pretty detailed understanding of similarities across multiple queries.

Training location is flexible. Iceland?

The core problem is that large numbers of people trade for effectively no good reason. Because of HFT, these people are now less screwed than they were previously. The "house" is competitive.

Yes, it's a profitable business for HFT. I'm not too bothered by that.

If you want to direct your ire somewhere, direct it at companies like Robinhood that are gamifying financial markets.

The less reported fact is that TIPS yields are moving in lock step with nominal yields. Breakeven inflation is approximately constant at 2.2-2.3%. So this is not about inflation expectations. It feels like a genuine capital shortage, possibly driven by massive AI-related investment demand.

I was getting excited but thanks for reminding me of this. Not messing with this.

Right the benefit of pi's customization is that it is exactly as locked down or not as you want it to be, in exactly the way you want it to be, and you can inspect it to verify.