I run a manipulation research lab at UC Berkeley. We have fourteen robot arms, six dexterous hands, and a budget that would make a language model researcher weep with pity. I am writing this because I need you to understand a number: 120,000.

That is the approximate ratio between the data available to train a large language model and the data available to train a robot to use its hands.

GPT-4 was trained on roughly 13 trillion tokens of text. The largest open robotics dataset—DROID, which my colleagues and I helped build—contains approximately 76,000 demonstration trajectories. Convert those to a comparable unit and the gap is not 10x or 100x. It is 120,000x. Five orders of magnitude. Language models live in an ocean of data. Robot hands live in a puddle.

This is not a funding problem. It is not a compute problem. It is a capture problem. No one has built the infrastructure to record what human hands do, at scale, with sufficient fidelity, across the full range of human dexterity.

Consider what "sufficient fidelity" means. A surgeon performing a bypass graft makes finger adjustments of less than one millimeter, at pressures measured in fractions of a newton, while processing tactile feedback through bone conduction in the fingertips. A piano tuner detects frequency deviations of 0.25 Hz through vibrations transmitted through a metal pin into the pad of a thumb. A sushi chef applies different pressures with each finger simultaneously, calibrated by the specific resistance of different fish textures.

Video captures none of this. Video sees the hand from outside. Dexterity lives inside—in the pressures, the micro-tremors, the proprioceptive feedback loop between finger and material that operates below conscious awareness.

The dexterous robotics hands market was $2.82 billion in 2025. It will be $10.3 billion by 2031. Robot multi-fingered hand production exceeded 51,000 units last year, up 21 percent. Sixty-four percent of newly developed dexterous robotic hands incorporate embedded machine learning. Every single one of them is starving for training data.

Figure AI allocated "much of its recent $1 billion funding" to hiring humans for first-person data collection. Companies are paying $118 per hour for teleoperation data—down from $340 two years ago, but still exorbitant at scale. One startup in Los Angeles pays gig workers to record themselves doing household tasks while wearing camera-equipped headbands. Another in India pays $1 per hour for similar work.

All of this is ad hoc. None of it captures the hands with the fidelity the robots need. None of it compensates the contributors fairly. None of it scales.

The Glove Concern is attempting to build what we in the research community have wanted for a decade: a systematic, high-fidelity, globally distributed capture network for dexterous manipulation data, with economic incentives that align contributor interest with data quality.

I am not a token enthusiast. I have no $GLOVE holdings and no financial relationship with the Concern. I am a roboticist who has spent his career trying to teach machines to use their hands, and I am telling you that the approach is sound. The hardware spec is serious. The data format is compatible with every major robot learning framework. The contributor incentive model—skill royalties tied to data usage—solves the quality problem that plagues gig-worker capture: when your income depends on data quality, you care about data quality.

Whether the token economics work, whether the DUNA structure holds, whether the Council governs wisely—these are questions I cannot answer. I am an engineer, not an economist.

But the data gap is real. 120,000x is real. And someone has to close it.

The Concern is trying. Their hands are open.

☜