of course i understand the second order effects. i just don’t agree with the world anthropic wants. i think it is darker, worse, and not nearly as safe as anthropic claims it is. i much prefer a world where progress is slow and diffuse, yet uncontrollable. capital consolidation is one of the biggest factors driving capability speed. if the secrets are out and models are distillable by default, margins get compressed, clusters get smaller, inference is cheap, training runs focus more on specialization than god models. the world gets to build the ai future together. i am unsympathetic to the arguments about china. they’re gonna build their chips eventually. we are racing to out-accelerate them and build an insurmountable lead and let a couple people be lightcone king in the name of western liberal values as a side effect. and then what? it’s dark. i will keep fighting against it.

We consume data we did not create. We inherit tools we did not invent. We run on chips we did not make. But when the commons bears fruit, we fence it.

i feel like if there is 1 benchmark that actually matters nowadays... this is one of them

imagine telling your customers there's a small chance you'll randomly decide they're using your product wrong and you won't tell them but will secretly silently sabotage their work

Its also worth mentioning that anthropic is now storing all logs for fable api calls for “harm reduction and safety” They say they wont train on the data, but does that even matter anymore if they can read the data and build RL environments from it?

GPU MODE has powered much of the public GPU kernel work online, with a permissive license from day one and generous credit from researchers, NVIDIA, AMD, and others. Today we’re moving our datasets to the Researcher Reciprocity License.

An early user (R&D engineer) texted us today: “A few minutes using Altara got me more excited about LLM-based tech than hours of keynotes from OpenAI and Anthropic. It's really refreshing to actually experience real transformation as opposed to endless hype.” Now back to building 🚀

very excited for what is coming next x.com/trekedge/status/206373…

How can i join the NVIDIA Nemotron Coalition?

SITUATION DETECTED: For the first time in internet history, agentic traffic has surpassed human traffic online, per Cloudflare Radar.

Codex somehow hasn't fully permeated MSFT but AWS & OAI announce their partnership and a few weeks later theyve rolled out managed-agents based on the codex harness and NVIDIA rolls out codex to their entire company...

Introducing Self-Distilled Policy Gradient. Token-level rewards, credit assignment, self-distillation. RL and distillation are converging toward the same idea: Policy gradients, it always has been, it always will be. huggingface.co/papers/2606.0…

this is so cool

We are hiring! DMs are open

Align with how @cursor_ai has done its RL stage — Astraflow is a new RL engine that enables asynchronous, heterogeneous, and geo-distributed RL in a native way through dataflow abstraction~ Like @FireworksAI_HQ’s sparse RL transfer design, it syncs only ≤1.1% of model weights — making remote rollout lightweight and efficient. Check it out!!!

i wonder if he had to go through all of the technical interviews lol

The bitter lesson in 26 words: Don’t be distracted by human knowledge, as AI has been historically. Instead focus on methods for creating knowledge that scale with computation, like search and learning.

🚀 BLASST just won Best Paper at #MLSys26! In this paper, we introduce a simple, training-free dynamic sparse attention mechanism that uses a single scalar threshold on online softmax statistics to skip negligible attention blocks. Unfortunately I won’t be there in person, but please say hi to my awesome coauthors! 🙌 Paper: arxiv.org/abs/2512.12087