In Q3 of 2025, Bot Auto achieved its first “driver-out” run on public roads: a trip in which the truck drove itself with no human behind the wheel, and in our case, no humans in the cab at all. This is a milestone reached by only a tiny handful of AV trucking programs. From the founding of the company to that milestone, we spent just $212,552 on one category of work that is usually very expensive in AI: paying people to manually label training data—for example, drawing boxes around cars and pedestrians—so a neural network can learn from them.
To many people that number does not sound like a breakthrough. It sounds like something is missing: a cost not counted, a line item not disclosed, or some clever maneuver hidden just outside the frame.
Such skepticism makes sense, because in AI, annotation is usually not a rounding error; it is a major line item. To see why this matters, consider nuScenes, a well-known dataset in autonomous driving. In total, it contains only about 5.5 hours of driving data. Keyframe by keyframe, human annotators had to identify and precisely label every vehicle, every pedestrian, every object in the scene. That work reportedly took 7,937 hours and cost about $100,000.
The missing piece is not in the accounting. It is in the assumptions.
What many people have missed is not a hidden expense, but a new era—a shift in how intelligence is produced. We are moving from a data-driven regime, where progress depends on the human annotated labels, into a compute-driven regime, where compute itself does that work at scale.
When an industry encounters such a fundamental paradigm shift, the old lens does not merely become incomplete. It becomes misleading.
FROM WORKSHOP TO FACTORY
For most of the last decade, applied AI, especially autonomy, has operated like a workshop. A workshop runs on human effort. If you want your system to learn, you hire people to label data, tag edge cases, and build an entire pipeline around human labor. The scarce resource is not raw data since raw data is everywhere. The scarce resource is labeled signal. Human labor is what keeps the machine learning.
A factory model changes the energy source. In the Industrial Revolution, steam replaced muscle, then electricity replaced steam. In the same way, a factory-era AI company replaces human labeling energy with compute. Instead of paying people to label everything by hand, it uses models to generate supervision at scale. Humans still matter, but their roles move upward, from manually drawing boxes to system verification and quality control. That is the real shift: Labeled signal stops being scarce, and the ceiling on intelligence starts to rise.
This is not something that happens just because a company buys more GPUs. It is a new industrial capability, and it is still rare. Only in the last few years did it become plausible that computers could begin taking over part of the manual work of annotation, not just helping with it. High-profile models such as Meta’s Segment Anything were the first to demonstrate, at scale, that this shift was real. Even now, only a small number of frontier companies can actually do it—and do it well enough to matter.
THE REVOLUTION IN PLAIN SIGHT
We’ve already lived through this factory shift once, but most people didn’t notice the real reason it worked.
The world talks about ChatGPT as an application miracle: It writes, it codes, it controls your home thermostat. But the deeper breakthrough was hiding in plain sight. ChatGPT did not get smart because humans labeled everything it read. It taught itself by reading text at internet scale and learning to predict what word comes next. Once that bottleneck loosened, scale exploded. Models could grow until they hit a new ceiling: compute. And with that shift came a step-change in capability.
We saw the same pattern in another domain. AlphaGo was data-driven: It learned from human games. AlphaZero broke the ceiling by removing that dependency, starting only from the rules, generating its own experience through self-play. Once learning was no longer constrained by a finite archive of human examples, the ceiling moved. As a result, the new system didn’t just improve on the old regime. It rendered it obsolete.
THE ONLY QUESTION THAT MATTERS
Richard Feynman wrote, “For a successful technology, reality must take precedence over public relations, for nature cannot be fooled.” That is the right test for this era of AI. Every marketing team is trying to present its company as AI-native, futuristic, and built for the new era. But none of that tells you what is actually behind the curtain.
The question worth asking of any AI company today is simple: Where does its labeled signal come from? If the answer is still a massive labeling vendor invoice, the workshop is still running, regardless of what the press release says. If compute is manufacturing supervision at scale, the factory has begun. The gap between those two answers is not a detail. It is the whole game.
Xiaodi Hou is founder and CEO of Bot Auto.