In Q3 2025, Bot Auto made a groundbreaking achievement with its “driver-out” run, where a truck drove autonomously on public roads without any humans in the vehicle. This milestone was reached with notably low costs in training data annotation — only $212,552 — a fraction of what is typically expected in AI development. This cost discrepancy reveals a fundamental shift in AI: transitioning from a data-driven model that depends heavily on human-labeled data to a compute-driven model where machines generate supervision autonomously. Traditionally, AI development resembled a workshop, heavily reliant on human effort to label training data. However, the emerging factory model replaces this manual labor with computational power, significantly scaling intelligence production. This transition is rare but has been validated by advances such as Meta’s Segment Anything and transformative AI systems like ChatGPT and AlphaZero, which learned and improved by leveraging large-scale compute and self-generated data rather than human-labeled examples. The critical question now for any AI company is where its labeled data originates — human labeling signals an old paradigm, whereas compute-generated supervision marks the dawn of a new industrial AI revolution.
Back
