Now

What I'm focused on

Last updated July 2026. What's a now page?

• Teaching small models to read the real world. Most of my time goes to curbcheck: a benchmark and a fine-tuned Qwen2.5-VL that reads SF parking signs. It's my excuse to go deep on multimodal post-training and evals, the space I want to work in next.
• Going deep on post-training and evals. Reading, reimplementing, and obsessing over one question: what makes an eval you can actually trust. Small models doing one narrow job extremely well, and the harnesses that prove they do.
• Talking to teams. Figuring out what's next, especially frontier-lab and consumer-tech work on multimodal models and ranking. If that's you, say hi.
• Off the keyboard. Watching too much Arsenal, learning to stop shanking the ball at padel, and defending my seat at poker night.