Really importantThe pool of actors who can train a "10T Mythos" is much larger than you might have been led to believe. Total param scale is…
> It is a monster model, yet clean in architecture; elegant in optimization, beautiful in MoE designs.I've been sayingthey do think of their project in aesthetic…
Beautiful things on the timelinepretty close! better luck next timeWu Haoning: The more we improve upon building it the more we feel hungry.I am confident that…
The nice thing about this is that it forces roon and his ilk to compete on product harder (while also racing on the model layer). The…
Xi doesn't even have a social media accountThe US is 100 years ahead of China this is what Technocapital looks likeAndrew Curran: The Truth API Data…
This is obviously a toy proof of conceptBut do you understand that the current timelines for "EUV by 2030s" or any other engineering goal are getting…
> Those are between 66-92% *fewer* FLOPs per watt than the B200/B300s that OAI/Anthropic have had for ~1.5 yearslolArya Tschand: I could be completely missing something,…
RT Rudzinski MaciejThis with latest google post on diffusion is converging on things that @Dorialexander and @teortaxesTex sugested about importance of specific kinds of synth dataAnd…
That would be fineKimi doesn't have Flash, and is brutally destroyed by V4-Pro (not to mention Flash) on price. I would like a K3+V4-Pro/Flash-as-subagents stackDeepSeek's pricing…
RT Xinyu ZhouNo. We are focused on intelligence frontier right now. Video generation does not bring much intelligence both theoretically and practically. It is useful though.WK:…