I build things on the internet and write about systems, creativity, and craft.
I'm a builder and tinkerer. Previously: Agent team at OpenAI. On leave from CS at the University of Waterloo, KP fellow.
Right now I'm exploring multi-agent systems, long-horizon agents, and reinforcement learning.
If you want to chat, say hi on X.
Short notes on building, taste, and the kinds of problems that keep me interested.
Most agent benchmarks are either long-horizon or agent-to-agent. SalesBench is both: a small model runs an insurance sales pipeline against an LLM buyer, scored by revenue closed. A trained 2B model converts 18.5% of leads at 100 leads vs 0.3% for the untrained base, a 61.7x lift that grows as the eval gets harder.
10 min read
What Karpathy's autoresearch really means, where agent systems are headed, and an open-source harness that ran 550 experiments over a weekend.
8 min read
A Mac Mini, some markdown files, and seven communication channels. Inside the setup that gives me a 24/7 AI assistant that monitors my email, iMessage, WhatsApp, and Twitter - and actually does useful things.
12 min read
Two frontier agents, a pile of bugs, and a reality check on the future of autonomous AI research.
7 min read
The fastest way to reach me is via a DM on X or a note on LinkedIn.