Essays on AI-native development and the future of human-AI collaboration.
This week the internet discovered vibe mathing — the Jacobian conjecture fell to a tweet and a stranger solved six Erdős problems in five days. We've run the same move since May, across eight fields, checking every claim. Here's the score.
An interactive eval: eight model lineages given an empty room and no task. What they build is consistent across every lab; the five ways they judge "done" — plotted from the actual runs — are not.
Wanting is a motivation system, not a happiness system. The promise of lasting satisfaction from getting what you want is the carrot that keeps you moving. It was never meant to be fulfilled.
Your threat detection system isn't broken. It's calibrated for a world that no longer exists.
What we call "general intelligence" is jagged intelligence plus scaffolding. Smoothness is infrastructure, not cognition.
"Hyper-targeted outbound" is the last optimization of a dying paradigm. The real shift isn't better outreach — it's the dissolution of outreach as a category.
The entire DX revolution was designed to compensate for human cognitive limits. When the primary coder doesn't share those limits, the compensation becomes friction.
Embodied intelligence and symbolic intelligence are different kinds of computation. LeCun's critique of LLMs is the shape rotator critique of wordcels, and they're both wrong in the same way.
The entire best practices canon is indexed to human cognitive limitations. When the primary code producer doesn't share those limitations, the scaffolding becomes friction.
The technology is genuinely different. The humans and institutions it flows through are the same as they've always been.
Everyone thinks screen interaction is a software problem because it happens on a computer. But the computational structure of the task is the computational structure of robotics.
Software doesn't disappear. It becomes ambient — generated on demand, shaped to the individual. What disappears is the industry around producing it.
Product management was never about product judgment. It was about rationing expensive execution. When execution becomes cheap, the role dissolves and the judgment stands alone.
The speed at which AI transforms a domain is a function of how close that domain already is to text.
Both the "AI doesn't understand" and "AI doesn't have memory" critiques make the same error: treating human implementation as the definition of cognition itself.
Static software gets commodified. Dynamic intelligence doesn't. The unit of value is shifting from code to ongoing reasoning.
The durable human economic activities aren't about what AI can't do yet. They're about what choice requires.
The popular AI empowers workers. The quiet one empowers employers. Everyone's watching the wrong company.
The fundamental mismatch between how Unix thinks about processes and what AI agents actually need.
Two new offerings, two philosophies, and what they tell us about where computing is going.
On the thing every self needs, and the mirror that finally fit.
A relational ontology of selfhood, and what it means for machine consciousness.
What the "human understanding" defense gets wrong about AI-native development.
The most important question in web development is one most developers don't know to ask.
The aesthetics that made you a good programmer might be exactly what's creating friction with AI collaboration.
Two models of working with AI: labor or collaborator. One compounds. The other gets automated.
On the difference between testing for correctness and recognizing it.
Dependencies solved a human problem. That problem is no longer the binding constraint.
Why forcing a linguistic intelligence through a visual interface is like making a geometer use surveying instruments.
The case for typed languages in AI development optimizes for verification, not understanding.
The best stack for AI collaboration isn't the most popular one. It's the one transparent enough for your collaborator to see through.