We introduce representative generative benchmarking—custom eval sets built from your own data that reflect real user queries. thank you Weights & Biases for collaborating! link to report in replies

a high utility cli is what the kids want

hosting an event in NYC on July 7 on context engineering thanks to Ramp for hosting lu.ma/euwisvqp

Shooting some sweet sweet content for Chroma

Love Hammad's perspective on working with LLMs. Strongly recommend watching.

You can't just yeet everything in the context window. Check out our research report on the topic.

Using Bun makes me excited to build again, so easy, has everything and just works

The company I work at has this neat property where if a customer gives us one dollar it costs less than that one dollar to provide our service to them.

Go try it out

Agents perform better with ground truth, so we built a tool to help them get that.

Made some ad creative this week for Chroma. Midjourney for stills and video ElevenLabs for voice over and sound effects. Edited together in Davinci.

Sora 2 correctly diagnosing medical imagery when you supply it as the starting frame and the prompt of “diagnose this”.

Working on natural language search over a multimedia corpus. moondream has the best open weight vision model. Don't even look around at other options, just start there and build the rest of what matters.