so simple.

update: we just went from 6 to 7 people and are hiring more! positions we are hiring for (10k referral bonus): 1. full-time videographer, because our lives are an actual movie and someone needs to document how fast we ship 2. model performance engineer so our model performance is more performant 3. chief of staff. 7 staff feels like a lot 4. founding SDR who has to be 'really hungry' or whatever that means tag someone who would be sensational in one of these roles for a chance to win 10k!

we're not a billion dollar lab. we're just 6 people in a cozy sf office who love distilling models and serving them at internet scale. we just beat claude 4 at video annotation for 17x less cost. couldn't have done it without @grass. incredible team to build with. and we're only getting started.

Introducing ClipTagger-12b. A state-of-the-art video annotation model trained in collaboration with @grass. ClipTagger-12b delivers video annotation capabilities on par with Claude 4 and GPT-4.1 at 17x lower cost. Learn more:

If crypto is about “making the game fairer”, AI is about “playing a bigger game”.

First PR to UwU from @SkillfulElectro merged! UwU now supports models from all top providers, including support for custom endpoints and @ollama

new model in prod. waves of demand 🌊

Introducing UwU, a simple way to generate CLI commands using GPT-5 directly in your terminal Claude Code does too much. UwU is designed to feel like an extension of your terminal, rather than a whole new environment Great for one-off commands you can't remember GitHub below ✨

We trained a model that’s as good as OpenAI’s best models at HTML to JSON schema It’s way cheaper to run and much faster -- cut costs by 50-90% for a handful of orgs doing millions/billions of requests per month The model will be open source soon hmu for early access

Serverless inference is a race to the bottom, so how exactly do these hosting companies make any money? In our latest blog post, @AmarSVS breaks down the economics of hosting OS models 👇