@ptrschmdtnlsn checks out. since we're just coming up on the 4th anniversary of release we won't really know the empirical impact on mean total lifespan for a few more years...enterprise doesn't care past that so my bet is they optimized for that. also they do seem to kill more motherboards
@analogalok lol what a failure. 11% larger than 3.6 which came out 4 months ago, almost 200 points lower on GDPVal-AA and 9% behind on terminalbench. nobody will use this, in 2 days we are getting 3.8
@doodlestein if it's an open-weights model, it's a trick you can only use once and then your reputation is gone. in the future I expect a more robust set of alignment benchmarks which would indicate compromise, but that runs into evaluation awareness, so all of a sudden we see moloch
@antopatrex1 every car company is modeling unit economics per car and nobody's asking who's gonna fix them when they break. there are maybe 200 people on earth qualified to maintain cars at scale. your deployment math is a fantasy if your maintenance workforce literally doesn't exist yet