Research speed is mostly the speed at which you discover you're wrong — which makes tooling a first-class research activity

Vivek locates the real game in throughput of disconfirmation: the stories about Alec Radford “involve volume. more runs per day, more wrong ideas discarded per week, a model of reality that updated faster than anyone else’s… research speed is mostly the speed at which you discover you’re wrong.” The consequence is that “tooling is a first-class research activity” — “launching a run should be one command… comparing two runs should take seconds, not an afternoon of archaeology.” Karpathy’s “overfit a single batch before training at scale” is the canonical cheap-first move: “thirty seconds, half your bugs, gone.”

This is the research-craft root of Autonomous coding loops need small stories and fast feedback to work — fast disconfirmation requires small, cheap, reproducible iterations, the same loop that lets agents ship overnight. It’s why A mediocre agent inside a strong harness outperforms a stronger agent inside a messy one generalizes from agents to researchers: the machinery that turns ideas into tested results is the multiplier, not raw capability. And it carries the warning of Speed without feedback amplifies errors — agents lack the self-correction mechanism that constrains human mistakes in reverse — speed is only an edge when each run actually tells you you were wrong, which is what overfitting one batch and comparing runs in seconds buys you. Building that machinery yourself is the move in Engineering is no longer the junior partner — at the frontier, research and engineering have fused.