When Microsoft shipped a coding model, it led with the tokens each answer cost, not the score. A small signal that the contest is shifting from beating benchmarks to beating them for less.

AI has mostly solved everyday coding, the kind where answers are cheap to check. Scientific coding is the harder and more consequential frontier, and history says the tools science forces into existence end up belonging to everyone.

The scarce input to AI progress is no longer data. It is a world worth practicing in. But a world is far harder to come by than a dataset, because the property that makes an RL environment valuable pulls against the property that makes it trainable, and that tension is now the rate limit on how fast AI improves.

Today an eval is something a specialist runs: a model scored against a benchmark. As agents do more of the producing, that word escapes the lab. Every answer you get becomes a small eval you are running, and the durable skill of the era turns out to be evaluation, not prompting.

The story about AI and work is automation: the machine takes the job. Here is a job the machine is creating. As progress shifts from training on data to practicing in environments, building those environments is becoming a craft, and the scarce input to AI is no longer data but a world worth practicing in.

Going public is sold as validation. It is first a change of governance, and for a company whose product is an open-ended research bet, it installs exactly the quarterly pressure that long-horizon research is worst at surviving. But the market has learned to reward patience, and the labs built structures to defend the mission, so whether they go the way of Amazon or the way of the quarter is a genuinely open, delicately balanced question.

Claude Code can do more than run a task. It can author a small JavaScript program at runtime that fans work out to many copies of itself, collects structured results, and hands back a synthesis. That program is just a file, so you can commit it and run it again. Here is how it works, what it is good for, and why it can burn a lot of tokens.

The popular story about AI and companies is about headcount: smaller teams, fewer workers, the coming one-person business. That is the surface. Underneath sits a ninety-year-old question about why companies exist at all, and AI is the largest shock to its answer in a generation, though not in the single direction the hype assumes.

Anthropic's biggest developer event of the year shipped no new base model, and was still a major month for AI capability. That isn't a contradiction. It's the clearest sign yet that the frontier has moved from the model's weights to the architecture built around them.

For forty years, the thing that made software the best business in the world was a quirk: serving one more user cost almost nothing. AI inference puts a meter back on every answer, and that quietly unwinds the economics that scale, freemium, and 80% margins were all built on.

AI tools have made writing code faster and cheaper. They have not made software development easier, and by lowering the cost of creating complexity, they may have made it harder.

SWE-bench improvements get read as a developer tools story — better Cursor, better Claude Code. That's real. It's also the smaller of the two effects. Better coding means better agents, and the gap between those is wider than it looks.

Non-profit research organizations occupy a structural position in AI that no commercial institution can replicate: technically serious, independently positioned, and optimized for knowledge rather than product. METR and Transluce are among the clearest examples of what that independence enables.

A note on what separates founders who hire well from those who don't — and why the difference rarely shows up in the interview itself.

Scattered observations from a few years of trying — what worked, what was theater, and what I would do differently.

Enterprise software has always had two distinct layers. AI is collapsing the boundary between them, and the resulting load on your systems of record wasn't in anyone's pricing model.

An engineer's framing of why markets sometimes work better than the people in them — and why that framing has limits.

Most AI applications start with one model and tight coupling to one provider. That's fine for a prototype. It becomes a liability the moment the field moves — and the field is always moving.