Very excited to see the results in 4 hours !

My friend had GPT-5.2-pro grade the solution given the ground truth PDF. Results: > Net: incorrect for 2, 3, 7, correct for 1, 4, 5, 9, 10, and 6, 8 look correct but are the ones where correctness depends on technical details Interestingly, Jakub says they're somehow confident 2

NBP made it better

TLDW if you're short on time: GPT-5.2 derived a new result in theoretical physics (...as they wrote in the announcement) and did not invent new physics.

From this it's easy to see what Mechanize meant by "replication training" It seems like a fruit hanging sooooo low now Close the loop, improve 🚀

For my non-American friends: current level is around 4.3%, so 18% is... x4 of that. "Unemployment is up 100% becuae of AI" sounds scary, but will be 2x less than the threshold in this tweet.

5.2 is on par with Opus 4.5 on the private split of SWE-Bench Pro dataset by Scale AI. It consists of proprietary codebases from startup partners that are not publicly available, so they can't be in pretraining, contrary to SWE-Bench and even Re-Bench (where models could be more

Lower gross profit margins, obviously

I called that out 3 months ago (and even earlier, but last Fall my "hot take" tweet got 80k views, so I feel obliged to report on my prediction. ikot.blog/the-illusion-o… TLDR on the image

It's obviously not only this benchmark. Like, it's not a secret at all that the open models are lagging behind. I wrote a blogpost with a detailed analysis, taking K2 Thinking (released 3 months ago -> we have benchmark data) as a reference. Go read it: ikot.blog/the-illusion-o…

new models on Arena.ai what it could be hm

Big W for Google if True