What this is
British engineer Michael Gomes Vieira published a short post on his personal blog this week, titled "Thinking Is Slow. Copying Is Fast." — surfacing an underappreciated angle: making AI "think" is slow (each inference costs significant compute), making AI "copy" someone else's thinking is fast (one distillation or fine-tuning run can transfer the capability).
He used an analogy: instead of training a new chef from scratch, just photocopy the Michelin chef's recipe book. Distillation itself isn't new, but he placed it inside an "engineering time horizon" framework: the bottleneck for AI projects isn't "can it be done" — it's "how long does one run take."
What's worth noting: this lens explains why cost declines across the LLM industry over the past year have far outpaced chip-level progress — everyone is "copying," nobody is reinventing the wheel.
Industry view
We note that Chinese teams doing model compression have been using this principle for a while — DeepSeek-V3's distillation pathway, Kimi's long-context solution, Zhipu's on-device small-model matrix (models running locally on phones and other edge devices) — all of which are fundamentally "not reinventing the wheel."
But the pushback is direct. A researcher at a London AI safety institute pointed out in the original post's discussion thread: distilled models are systematically weaker than the source on "edge cases" (extreme or rare inputs not seen in training) — copy speed masks the loss of "latent capabilities." A big-tech infrastructure engineer was blunter still: "Copying is fast because you gave up explainability — you don't know when it will break."
In other words, "copying is fast" is an engineering advantage, not capability equivalence. That is the half the principle leaves out.
Impact on regular people
For enterprise IT: When selecting a large-model API, you can confidently use distilled small-model versions — cost may be as low as one-tenth — but don't assume full substitutability for the original. Keep a fallback for business-critical workloads.
For individual careers: When picking AI tools, "is the answer correct" matters ten times more than "is the answer fast" — yet the latter is what buyers and sales teams notice first. Flip your evaluation weights.
For the consumer market: Phone AI assistants getting smarter doesn't necessarily mean models got bigger — more likely they're using distillation. That's good news, meaning on-device AI (running locally, no network required) will get cheaper and more ubiquitous.