mostly agree but you shouldn't *only* make decisions based on asymptotes. The constants matter. e.g. Cursor has already created a ridiculous amount of value and that will continue. They'll exit, pivot, or start training their own models before the asymptotic trends kick in

going backpacking with my friends. spreading the gospel. everyone will come out of the forest bitter lesson-pilled

Hmmm 🤔

Why is it important that there exist leading open models built in the west? "We should have open models that reflect western values" can feel vague and hand-wavy. Here's a deepseek screenshot to make it concrete. This is not a model you can do e.g. factuality research on

i really didn't think the gpt-oss was gonna be this good holy moly

got setup in 2 mins. good release.

a second openAI open-weights model has hit hf

holy shit they're gonna release 5 more open models

Does Isomorphic Labs write anything in the public? Where can I learn more abt AI for drug discovery (or knowledge discovery in general)? Is anyone e.g. using alphafold as an RL env to derive rewards from?

uv is good

Time to finally learn what a protein actually is

nooo dw your 7B dense is actually perfect. The big MoE's scare me 😅

yayyy

I intend to acquire cohere after its acquisition of perplexity after their acquisition of TikTok and google chrome

life update: for those who don’t know, i joined Ai2 a few months ago to work on open source AGI. incredibly excited about what we’re building 🚀

Introducing synth-bench -> how good is your LM at generating data for other LM's? Olmo best is weirdly good at this, apparently!

david is one of the most thoughtful people I've met in general, and he's definitely the most thoughtful person I've met when it comes to how to interpret eval scores of language models. ty david + team for helping us make sound decisions!! (P.S. go follow David Heineman)

The space of sequential ops is more expressive, but parallel is more efficient. Beauty of the transformer is it kinda defies this tradeoff