Simon Willison, creator of Datasette, reviews three shifts in his use of language models in this 20-minute AI Engineer talk.
Coding agents become daily tools
- Willison dates his first inflection point to November 2025, when small model improvements made coding agents reliable enough for his everyday work.
- He replaced his annual resolution to take on fewer projects with an experiment in greater ambition, then found himself juggling too much.
- His pelican-on-a-bicycle drawings illustrate how a memorable informal test can stop distinguishing stronger models.
- “Deep blue” names the professional unease he describes when models perform skills central to a developer’s identity.
Define a goal and provide the tools
- His second inflection point is persistent problem solving: agents can keep working when the objective and available tools fit the problem.
- He describes a beep-locator project for a detector hidden among roughly 40 unpacked boxes.
- Two iPads, two laptops and a phone supplied microphones; he says the generated software found a relevant acoustic-ranging paper and localized the detector within 41 centimeters.
- The story joins problem definition, literature search, tool access and real-world feedback in one project.
A playable game still needs iteration
- His first raccoon-heist prototypes looked like games but held his interest for only about a minute and a half each.
- A later strategy-game experiment changed after he asked the agent to make it more fun.
- He reports a 12-hour run with roughly 1,000 simulated battles, new objectives and additional feedback effects.
- His third inflection point is that sustained iteration can improve the experience, beyond generating something that merely resembles a game.
Faster work raises the ambition bar
- Willison’s security anecdotes warn that persistent problem solving also makes permissions and containment important.
- He asks why his job feels harder as tools improve: easy tasks disappear, leaving more demanding decisions and more possible projects.
- He closes with a cycling quotation that captures the pressure to use the extra capacity.
“It doesn’t get easier, you just get faster.” — Greg LeMond, quoted by Willison at 18:05
Caveat: the projects, model comparisons and security timeline are Willison’s account. His game examples and informal drawing tests are personal demonstrations rather than controlled benchmarks.