supposedly #DeepSeek v3.1 beats Claude Opus 4 on some coding benchmark or other and is “open-source”. “671B parameters (37B active per token).” Lots of stock footage in the #video. Optimized for "UE8M0 FP8", which I guess is an 8-bit floating-point format supported by domestic (SMIC?) GPUs like the recently-released Xi Yun C600 and a speculated-to-be-near-future Huawei chip. Apparently GPT-4 slightly beats it on coding benchmarks. #AI #news. Video is 2 hours 42 minutes. #toread
#humor #news #video of #Colbert announcing #Mamdani’s election and a bunch of other #USA #politics