Data Labeling

3/810 → 6/622; +0.59 percentage points of formation share.

What used to mean armies of contractors tagging bounding boxes has quietly become a specialization contest, and the six companies here show a category splitting along two very different axes: general-purpose scale versus expert scarcity. AfterQuery and Sureform sit on the volume side, building broad data supply lines for every frontier lab and for multimodal and robotics models respectively, while Sciloop occupies the opposite extreme, paying IPhO and IMO medalists to produce math and physics problems that current models simply cannot solve. Besimple AI and Liva AI narrow further still, treating human speech itself as the scarce resource, betting that accents, dialects, and conversational realism matter more than raw hours of audio. Cartpole breaks from labeling entirely, building reinforcement learning environments rather than static datasets, which points toward where the category is heading: this formation's own 2.60x growth trails Reinforcement Learning's 4.17x, suggesting labeling companies are being pulled toward interactive, evaluative data rather than one-shot annotation. That shift is also why the open question about rights, provenance, and consent for human-motion data matters so much right now, since teleoperation and imitation learning turn demonstrations into reusable assets with no standard tracking of permission or revocation. The timing traces directly to frontier labs exhausting easy web-scale text sometime around 2023 and 2024, pushing demand toward harder-to-fake, expert-verified, or physically-grounded data that no scrape can produce.

No comparable outside activity is connected yet.

These share a word with “Data Labeling” but are tracked as separate categories. The real overlap below is the actual share of Data Labeling’s companies also found in that category — not a guess.