Driving's Open Problem Shapes Tooling
Oscar Beijbom, co-founder and CTO of Nyckel, on the opportunites in the AI/ML tooling market
Autonomous driving created a uniquely deep market for data labeling because city streets keep generating new failure cases faster than models can fully absorb them. A self driving system has to handle rain, glare, dusk, unusual clothing, odd vehicle behavior, and every combination of those conditions, while document AI usually sees a much narrower range of inputs, like blurry scans, wrinkles, stains, and template variation. As models improve with pre training, that narrower document problem needs far fewer newly labeled examples.
-
The key distinction is open world versus bounded workflow. Driving is an open world perception problem where the environment keeps changing in uncontrolled ways. Invoice and bill of lading automation usually happens inside a repeatable business process with a limited set of formats, fields, and edge cases.
-
That is why autonomous vehicle programs were such strong customers for labeling vendors. They kept producing long tails of hard examples that needed human review. In contrast, Nyckel describes a future where product managers or developers label 10 to 100 examples of their own business data, train quickly, and ship without a large outsourced annotation operation.
-
The broader market implication is that value shifts away from raw labeling volume and toward tools that help a domain expert test, fine tune, deploy, and monitor on their own data. That is also why end to end products like Vertex AI are the more relevant comparison than pure annotation shops.
The next phase of AI tooling should look less like giant human labeling factories and more like software that lets the person who owns the workflow teach the model directly. The biggest annotation businesses will remain where the world stays messy and open, like robotics and driving, while most business AI moves toward lighter weight, self serve training on smaller custom datasets.