Blog

With 1M-token context windows, agent memory no longer wins by beating full context on short histories. Its clear value today is cost; persistence and learning remain open problems.

A follow-up on agent-system reliability: why checking whether an agent's claims are supported by its own trace, files, and git state is different from checking whether the work is correct.

A black-box measurement of DeepSeek, OpenAI, and Anthropic prompt caches: why prefix identity matters, where provider contracts differ, and how prompt layout becomes a systems interface.

A shared API key, a negative balance, and a harder question for research labs: if frontier-model access becomes part of the cost of doing research, who pays for the tokens?

Auto research is real and getting better fast. But it's not replacing researchers — it's changing where human effort matters. After months of building these systems, here's what I've learned about verification, deskilling, and taste.

Research Copilot is an open-source AI desktop app built for PhD students, postdocs, and faculty — it searches literature across four databases, writes papers with verified citations, runs data analysis, and helps draft grant proposals, all in one place.

RAM Desktop is a macOS research assistant that breaks your question into tasks, collects evidence, and advances the work round by round — pausing at key checkpoints for your review.

Two recurring fears: universities should be industry pipelines, and AI is ending software jobs. I think both framings are wrong — and the reasons why matter.

After years of redirecting to Google Sites, I finally rebuilt my homepage from scratch. Here's why and what's next.