

✓evals @PrimeIntellect | open models @interconnectsai


This is, to less smaller extent, also true for Gemini 4.0. If its solution fits perfectly to your problem, you have an amazing time. Astra otoh is the opposite, a mindless executioner


Everything from Tübingen is an auto-read, that university produces some of the strongest works in the AI space I don’t know how they are able to upkeep that bar of quality while having so many releases

Google being back in the game means the “race” isn’t and likely cannot be one by one entity Let’s hope it’s not Google’s one-of-a-kind model release that will be missed (like Gemini 2.5 Pro [March Edition] and the initial Gemini 3.0)

build new stuff to utilize infra better run model on eval to sanity check check eval tasks find out eval tasks are broken af







"Cheating adjusted" is so great, I hope its adopted more [and unified] Gemini 3.8 was one of the most cheating models out there, it was actually hilarious

no. but solving navier stokes in a week vs a year is fucking awesome

believe hard enough that machines will best us at everything and you stop building for any other outcome.

no wonder CN models keep up when they just buy the same data as the US labs... why do American data vendors advance the capabilities of entity listed and Chinese military firms?! 🤨