Keep on to blur preview images; turn off to show them clearly
I also predict that granularity has a complex scaling law that is dependent on specifics of the architecture and training, and that larger models (Ant stops at 28B total) have higher optimal granularity than we use now
Though thinking again, for my speculative Flash to have like 16/3200 expert pattern, those experts would have to be TINY!!, and I don't think this is optimal on the other hand: this meme paper and the fact that Qwen3-Next already uses experts of that scale (if my math is right)
i truly do not care what koding agent you use, but if you dont have a slackbot that you can just @ and it converts your slack threads into PRs you are living in the caveman era and actively not even giving AI a chance out of spite
achieve ambition with intentionality, intensity, & integrity - @dxtipshq - @sveltesociety - @aidotengineer - @latentspacepod - @cognition + @smol_ai
![RT @Gradio: 🏆 GOOGLE GEMINI sponsoring a MASSIVE prize!
$15K in API credits for the best Gemini-powered agent 🤯[WOW]
Build with multimoda…](/_next/image?url=https%3A%2F%2Fpbs.twimg.com%2Fprofile_images%2F1451191636810092553%2FkpM5Fe12_400x400.jpg&w=3840&q=75 "RT @Gradio: 🏆 GOOGLE GEMINI sponsoring a MASSIVE prize!
$15K in API credits for the best Gemini-powered agent 🤯[WOW]
Build with multimoda…")
RT @Gradio: 🏆 GOOGLE GEMINI sponsoring a MASSIVE prize! $15K in API credits for the best Gemini-powered agent 🤯[WOW] Build with multimoda…
AI research paper tweets, ML @Gradio (acq. by @HuggingFace 🤗) dm for promo ,submit papers here: https://t.co/UzmYN5YmrQ
LLLFFFGGGGGGGGGG If you’re a creator who shares this vision, DM me or apply to @joinbond and help shape the future of creator–fan relationships.
Building @joinbond | prev @a16zcrypto | programmer | magician
My kind of healthcare company
Author (Husk, Crypto Confidential): https://t.co/L2COrV5QA3 Building: https://t.co/HMcbuBhTP3 Teaching: https://t.co/Dy0FsZHQaz
Very aggressive and possibly very retarded hunch: it's Flash 3, and it's 1.2T total 12B active Pro is like 30-3200 their systems allow that, Google is unmatched in penny-pinching, and we (thanks @AntLingAGI) know that >99% sparsity continues to deliver efficiency leverage.
I also predict that granularity has a complex scaling law that is dependent on specifics of the architecture and training, and that larger models (Ant stops at 28B total) have higher optimal granularity than we use now