Ivan Zhao’s sharpest historical lesson is about the steam engine. Early textile factories ran on waterwheels; when the steam engine arrived, “factory owners initially swapped waterwheels for steam engines and kept everything else the same. Productivity gains were modest.” The breakthrough came only when they “decouple[d] from water entirely” and “redesigned their factories around steam engines” — then “productivity exploded.” His diagnosis of the present: “We’re still in the ‘swap out the waterwheel’ phase. AI chatbots bolted onto existing tools.”
The actionable test is whether you’re redesigning the work or decorating it: a chatbot pointed at an unchanged workflow is a waterwheel swap.
Garry Tan adds the asymmetry of who can actually perform the redesign. Describing the bureaucratic friction he lived at Microsoft — he and his PM mentor physically walking a bug across a highway to another team that “wouldn’t reply to our emails, wouldn’t fix our bugs, wouldn’t even mark it as won’t fix” — he imagines the same coordination handled today by “a markdown file and a skill that is run on a cron,” surfacing who is blocked and by whom, so that “you could release something in a day or two that would have taken like six months.” Then the asymmetry: “probably an org like Microsoft can’t. But a startup can. And every startup must.” The redesign is not merely available to small companies; on Tan’s account it is obligatory for them and something an org like Microsoft “probably can’t” do, which is the same inertia that makes Institutional slowness is the runway, not the threat a window rather than a warning. Extended past the company’s own walls, the redesign reaches the interfaces between firms — see AI erases the API line — the side that takes orders can now negotiate. This is the organizational twin of Scaffolding is tech debt against the next model — the bitter lesson applied to product building — both warn against preserving the old structure when the new material’s whole value is that it dissolves the old constraint. It is what AI compresses the distance between idea and execution but not between good and bad judgment looks like operationally (when execution is free, the leverage is in re-imagining the process), and it is the same proactive move as Self-disruption follows the value chain downward — software companies must eat their own agent layer before someone else does — redesign before the old shape is competed away. Mistaking the transition form for the final form is exactly New technology first imitates the medium it replaces — the transition form hides the final form.
Tom Blomfield reaches the same diagnosis from YC’s vantage point and supplies the alternative unit. Companies are “bolting AI on onto the side of the company” — question-answer bots, lightweight agents that can call a tool like web search — and the ceiling of that approach is making each person twenty or thirty percent more productive. His reframe is to ask what happens if the company is reimagined as a series of self-improving AI loops from the ground up, which he says produces “a very very different conclusion.” That is the waterwheel-to-steam redesign made concrete at the level of a component you can actually build: see Without a learning mechanism you have an AI pipeline, not an AI loop for the anatomy, and A human in the loop is a gating mechanism — it caps the loop at human waking hours for why a bolted-on agent stays capped at human speed no matter how good the model gets. Blomfield’s own caveat is worth preserving: he opens by saying nobody knows how to do this yet, and that anyone claiming to have it figured out is probably lying.