fucking UAE shillbots…Please, Ayatollah Mojtaba Khamenei, steel your heart! strike their hive in the AWS me-central-1 (25.2048° N, 55.2708° E)! Strike it down with your mighty…
This picture implies V4-Pro GA scoring ≈57.5 on AA, a notch above Kimi, improving by 13.5 points. This is achievable… on *some* timeline. But I doubt.…
yah, GPT-5.6 cyber is really good… luckily it's a closed model with stringent safeguardsit's not like open weights models can do that yet. Phew. safeAndrew Curran:…
RT You JiachengA more scaling-pilled and better IMLE?That's super cool!Alexi Gladstone: We discovered a third pretraining axis beyond parameters and data: exploration.Scaling exploration monotonically improves existing…
V4-Flash can be significantly *cheaper* still. I am positive after a lot of playing that it mostly burns tokens on nonsense. It's going to be timing-out,…
Wenfeng is too cultured to really act like this, but he totally feels this wayDeepSeek people have a lot of pride in what they doGorden Sun:…
More than 2xing the performance on NYT connections, huh@scaling01 I kind of want to see LisanBench againLech Mazur: @teortaxesTex DeepSeek V4 Flash: 89.6.DeepSeek V4 Pro Preview:…
High Agency. Talented European kids know their best move is becoming some sort of a sanctimonious politruk in the Longhouse Administration. German Greens are a special…
RT Christian SzegedyIMO, the next steps are:Within 1 year: strictly better than human AI in all problem-solving aspects of math.Within 2 years: producing mathematics becomes so…
RT wesley hsiehThe method is quite simple:For each data point x, they sample a timestep t, create K noised latents zt with K different noises, and…