Prompt caching is a prefix problem
Prompt caching reads cost a tenth of input on most tiers. The prefix that qualifies runs 512 to 4,096 tokens, and that floor picks your cheapest model.
Articles
Blog
Practical notes on building, deploying, and shipping real products for solo devs and small teams.
AI layoffs are inversion of control
AI layoffs are sold as obsolescence. The quotes, the labor data, and the hiring records point at bargaining power. How to tell the difference.
AI code review is not independent
AI code review is not an independent check: two AI passes miss the same bugs. A missed CI injection, an npm worm in repo config, and swarm data show it.
AI model fatigue is a business rhythm
AI model fatigue is a business rhythm. The deltas are small, the cadence serves the labs, not you. Own your workflow, gate every switch on your own evals.
Read the code AI writes
An AI agent wrote 400 lines that passed every test. Nobody could read them. A 30-minute human rewrite beat it. Here is why hard skills still matter.
Kimi K3 for builders who ship
Kimi K3 is the largest open model - 2.8T parameters, 104B active, a 1.56TB download. The verified numbers, and the catches that decide whether you switch your coding agent.
From AI prototype to paid product
AI gets you a prototype in minutes. The real work is webhooks, deploys, TLS, and refunds. Here is what lives in the gap and why it decides whether you ship.