Writing
- Language Models Are Few-Shot Learners, They Just Can't RememberEssay
GPT-3 promised few-shot learning. What we actually got was few-shot prompting: nothing persists. The machinery for real learning already exists inside every transformer. It's just trapped in volatile memory.
April 21, 2026
- Context Is Software, Weights Are HardwareEssay
Both KV cache and weights modulate activations. Context is software running on frozen hardware. Weight updates redesign the hardware itself. The difference matters more than most people think.
April 18, 2026
- The Irreducibility of Production Experience in SoftwareEssay
Every software system contains two kinds of knowledge. Systems knowledge is replicable; AI has made it near-free. Production knowledge is irreducible; only running in the real world can teach it. That gap is your moat.
April 12, 2026
- LLMs Quietly Cheat on Repos They've Already SeenEssay
SWE-bench numbers look impressive. But a large chunk of those benchmarks run on public repos the model was trained on. Your enterprise codebase is a closed book, and the performance gap is real.
April 2, 2026
- What I Told 100 Founders About Raising VCFounder's Log
I gave a talk at KSUM's VC 101 Bootcamp in Kochi. This is the version I'd write for myself -- the stuff nobody tells you before you're already in the middle of it.
March 15, 2026