Garden of Thoughts | SherryAnalytics | July 28, 2026
Sub-title: Why teaching is expensive and copying isn't.
A few people asked me to explain "AI distillation" this week, so here's my attempt.
Strip away the jargon and distillation is a simple idea: take a model that already works well, ask it an enormous number of questions, save the answers — especially the reasoning it shows along the way — throw out the bad ones, and fine-tune a smaller model on what's left. The student never touches the teacher's weights. It just reads text, over HTTPS, like anyone else with an API key.
This isn't hypothetical. In February 2026, Anthropic reported that roughly 24,000 fake accounts generated more than 16 million conversations with Claude over weeks to months, tracing back to DeepSeek, Moonshot, and MiniMax. Around the same time, Google disclosed that Gemini had fielded over 100,000 prompts from an operation trying to reconstruct its reasoning.
Everything that makes the teacher good — years of RLHF, RLVR, expert labelers, sandboxed environments, oceans of compute spent on rollouts — is real money spent on real infrastructure. But all of that investment eventually has to leave the building as plain text, one API response at a time. The student doesn't need to redo any of it. It just needs to collect enough of those responses and fine-tune on them, and years of someone else's R&D budget show up in a few weeks of someone else's training run.
That asymmetry is exactly why distillation is such a sore subject for the labs footing the original bill: the moat is paid for in cash, and it can be copied for the price of electricity and a few GPUs.
Signoff: Distillation itself is an ordinary, well-established technique. The problem was never distilling — it's distilling without asking.
Idea, outline, and argument: mine. I used Claude to help tighten the prose.