Today’s my 21st birthday :) This unlocks a world of new car rental possibilities I’m so excited.

Interesting idea of dropping overlap in eval datasets to decrease costs. As evaluations become harder and longer, costs will become incredibly high. Ensuring tasks don’t become repetitive, yet still rigorously test capabilities, is one way to mitigate this.

So you’re telling me it’s faster and cheaper now 🤯

Diminishing marginal returns on short tasks or simple QA evals can be misleading. In an economic sense, the value of LMs is dependent on the length of their task horizon in “human hours”. Can it do 5 hours of my work accurately or 10? Labs are beginning to prioritize horizon

Got the chance to work out of the Gateway X office today. Two things I learned ⬇️ 1. Code editors aren’t just for code I sat in on a meeting where the CTO went through an entire business first workflow in VSCode using Claude code. The best thing: No code whatsoever. He

(1/2) i felt like no one actually teaches you a good framework for how to read (ML) papers well + fast, so i wrote this 5-minute read tldr: because so many papers suck, here's how to go through them quickly and revisit the good ones

Can we just RL with Karpathy as our reward model?

Such a great read, highly recommend

New favorite account on X

Prompting and evals will be everything Prompting is agency Evals is taste

Gpt-oss-120b and Kimi-k2 provided by Groq Inc are now on the OpenRouter exacto endpoints! Evals are completely reproducible on openbench😁

Powered by Groq 🫡

Shipping speed almost as fast as our LPUs