Writing

Essays and notes on technology, company building, health, and mindset.

News feed

All writing

13 entries

  1. How huge AI models fit, remember, and respondTech, See how attention, compact memory, sparse experts, and GPU teamwork make large models practical to run.
  2. Building with Jev: use cases and implementation architecturesTech, How to put a fast decision model inside search, browser agents, model routing, clipping, smart paste, and graph exploration, with API examples and GLiNER pairings.
  3. Transformer basics: prefill, decode, and the KV cacheTech, Follow a prompt through token embeddings, hidden states, attention, and the KV cache. Then see how a text embedding is a different output.
  4. Jev and System One modelsTech, Why Jev can make decision-heavy software much faster and cheaper, how shared prefill and decision training work, and what people are building with it.
  5. Multi-agent systems scale thought in parallelTech, Notes from Noam Brown's interview with Dwarkesh Patel: several agents can finish a task sooner by splitting the work, at the cost of more total compute.
  6. Durable execution and atomic executionTech, A queue retries the job. A durable workflow remembers its progress. A transaction commits related changes together.
  7. Speculative decodingTech, A small model guesses a few tokens ahead. A large model checks them together, keeps the good guesses, and fixes the first bad one.
  8. Consistent hashingTech, Modulo hashing can remap keys across the cluster. A hash ring limits the change to the token ranges whose owners changed.
  9. Compound Engineering: My Self-Improving Remote Agent SetupTech, How Slack and Hermes give me 24/7 remote agents that learn from me, from others, and from their own work.
  10. How Text Embedding Models Are Trained and ServedTech, A visual, technical guide to how modern text embedding models turn language into vectors: transformer encoders, pooling, contrastive training, hard negatives, backpropagation, dimensionality, and production inference.
  11. Vision-Action Transformers vs. VLA vs. World-Action ModelsTech, The practical difference is what conditions the robot policy and what it is trained to predict: actions, language-conditioned actions, or actions plus future world states.
  12. AI Won't End Software Engineering. It will Make More of Them!Tech, The important effect of coding agents may not be fewer people building software. It may be that anyone with a desire, a question, or a frustration can enter the build, measure, learn loop.
  13. Why Wanting Something Too Badly Can Push It AwayMindset, A practical way to understand why insecurity-based desire can push away the exact thing you want.