1/Pretraining is hitting a data wall; scaling raw web data alone leads to diminishing returns. Today DatologyAI shares BeyondWeb, our synthetic data approach & all the learnings from scaling it to trillions of tokens🧑🏼‍🍳 - 3B LLMs beat 8B models🚀 - Pareto frontier for performance

We are excited to release Nvidia-Nemotron-Nano-V2 model! This is a 9B hybrid SSM model with open base model and training data. This model also supports runtime "thinking" budget control. HF collection with base and post trained models: huggingface.co/collections/nv…

you haven't really experienced america until you've had to deal with at least one complex medical and legal situation each

conference pc emails are weirdly passive aggressive "we are compiling a list" "there will be consequences" "as per your agreement" like chill this is a free service im doing late at night on my own time after my other two actually paid jobs, the world isn't ending in a day

>pro AI super PAC >e/acc lobbyists in DC >offices in New Delhi Yeah

if your policy gradient runs aren't converging it's not a skill issue you're just not pure of heart

turns out the new job offers complimentary flops at home

Oh hey so we did do this a few years ago for grade school science. Would love to see ppl get more of these Text Worlds up and running! arxiv.org/abs/2203.07540

special place in hell for swes who don't understand backward compatibility. bruh wydm you made server side changes that broke my code from last week without updating your api ???

Honored to be back on TIME100 AI for 2025 — alongside my longtime heroes Fei-Fei Li and Regina Barzilay! 😍 The recognition goes to my amazing students and colleagues, who strive to find ways to use AI to better humanity, as opposed to making AI for the sake of making AI better

Our team at Nvidia is hiring for full time! Please contact us if you're interested in working on LLM / DLM post-training and system optimization!

if you've ever used a nemotron model (esp nano v2 or llama nemotron super v1.5), looking to know what kinds of tasks didn't work for you in practice. Can DM or follow up here. we don't make user-facing ai apps, just open models, so I'm polling X in lieu of telemetry :)

compute resellers all the way down

If a PhD student in the US gets an offer to intern at a Chinese frontier lab, we should be paying them to go and come back. Why are we making it so difficult???

An unhackable reward function is the AI equivalent of a perpetual motion machine.

Soon: Her + Minority Report