All on AI.
Essays on AI models, agents, governance, and what production systems look like after the demo ends.
AI Is a Car, Not the Destination
AI can take us further and faster, but humans still need to choose the destination, steer with judgement, and know when to take back control.
AI will not kill SaaS. It will sort it.
The companies best placed to kill SaaS are not doing it. That observation is doing a lot of work.
Sanding off the messy bits
Humans learn from AI. AI learns from humans. Somewhere in that loop, the texture that made our writing sound like us starts to disappear.
What actually leaked when Claude Code leaked
The headlines said 'Claude source code is leaked'. The accurate version is narrower, more boring, and more useful. The distinction matters for how we think about AI moats.
Prompt injection is the risk we keep under-discussing
The model saying something wrong is a mild failure mode. The model being talked into doing something wrong is the one we're still not pricing properly.
The jagged shape of machine intelligence
AI writes strategy and fails at decimal comparison. The shape of that gap is informative, predictable, and mostly misunderstood.
Prompt engineering is becoming context engineering
The work has moved from writing the perfect instruction to assembling the right context. It's a plumbing job now.
Graceful degradation is the only AI strategy that survives contact with regulators
Why the first design question for any production AI system should be: what does this look like when the model is wrong?
Don't build a business on AI getting worse
Cleanup businesses betting on the persistence of AI 'slop' are building on a shrinking market. The durable bet is the opposite.
The quiet moment when agentic AI starts working
Most new AI tools are incremental. A small number produce the feeling of crossing a capability threshold. The agentic release was one.
Apple Intelligence as a layered bet
Apple isn't choosing an AI partner. It's building an architecture where no partner is critical and every layer is swappable. That's the whole strategy.
AI as collaborator, not mind reader
Users say 'that's not how I'd write it' and blame the model. The missing input is almost always context. Here's what context engineering actually looks like for your own voice.
AI learning from AI: the photocopy problem
As AI-generated content fills the internet, new models are being trained on it. The photocopy analogy is real, but the timeline and impact are less uniform than the headline suggests.
Sora and the quiet economics of generated video
The headline is creativity. The actual impact is a sweeping shift in the economics of stock footage, training data, and visual credibility.
Using ChatGPT as a 'shrink': useful, bounded, complicated
AI chat as low-friction emotional processing has real benefits and real limitations. The framing 'shrink on demand' oversells what it actually does well.
What ChatGPT actually changed in week one
Reading my own week-one post about ChatGPT, three years on. The 'forever changed' framing was right in spirit. The mechanism was less obvious in the moment.