The Solved and the Unsolved
AI agents are extraordinarily good at solved problems and extraordinarily bad at unsolved ones, and knowing which you're facing is now a core engineering skill.
Field notes
Notes on technical leadership, data systems, applied AI, and the work of building organizations that can take on difficult problems.
16 entries
AI agents are extraordinarily good at solved problems and extraordinarily bad at unsolved ones, and knowing which you're facing is now a core engineering skill.
Thoughts on where a moat for the frontier labs might come from, and why I think the next capability shift will be subtle.
Frontier labs are training AI agents to work like long-running consultants, favoring general-purpose autonomy over deeper software specialization.
What good software looks like is determined by the technological environment in which it is written
Heads: I develop a novel exploit, Tails: I hack into Hugging Face
They're Great!
When AI makes implementation cheap, software engineering shifts toward domain modeling, repeatable patterns, and systems designed to be dependable by construction.
A more systematic engineering discipline can replace the apprenticeship in accumulated coding taste with a faster, clearer path from foundations to independent judgment.
AI makes implementation cheaper, shifting the constraint in software engineering toward judgment, validation, and delivery.
A useful loop for building with agents when you do not know enough to specify the right solution up front.
The principles organizing my thinking about how software delivery changes when implementation becomes cheap.
Thoughts on well-designed software with ubiquitous coding agents
Don't live your life on cruise control
Import Amps scoped documentation functionality into Pi
Different mental models for thinking about working with Coding Agents
Building this portfolio site with Astro, Tailwind, and a little help from AI.