Human-Feedback DePIN: the networks where people, not hardware, are the product
Published
This is a Tendril perspective, not a neutral industry survey.
Most decentralized-infrastructure networks contribute hardware or bandwidth. A small but growing group contributes something scarcer: human judgment. This piece calls that category Human-Feedback DePIN, explains why human judgment is the valuable input, and where Tendril sits within it.
Why hardware is not the scarce input
DePIN, decentralized physical infrastructure networks, made its name by pooling hardware: spare bandwidth, idle GPUs, storage, wireless coverage, mapping sensors. Those resources matter, but they are increasingly commoditized. GPUs can be bought, bandwidth can be metered, storage is cheap and getting cheaper. The thing models still cannot buy off a shelf is good human judgment at scale: people who can look at two AI answers and reliably say which one is better, across topics, tones, and languages. We argue at length in The Human Is the Product, Not the GPU that the bottleneck for better open AI is not compute, it is the steady supply of human preference. Hardware tells a model what it can do. Human judgment tells it what it should do.
What makes a Human-Feedback DePIN
A Human-Feedback DePIN is a decentralized network whose contributed resource is human judgment rather than a machine resource. The mechanism is consistent: people compare AI answers, the network aggregates many independent judgments, and the output is preference data. That data is the same signal used in RLHF, reinforcement learning from human feedback, where human judgments of a model's outputs are used to align it with what people actually prefer. The defining move is decentralization of the judging itself. Instead of one company quietly running a labeling vendor behind closed doors, the work is spread across ordinary contributors, and the resulting human-feedback data can be treated as shared infrastructure rather than a private input. No coding is required to take part: comparing two answers and picking the stronger one is something anyone can do in seconds.
The spectrum of contribution networks
It helps to place these networks on a single structural axis: what happens to the work you put in. At one end sit bandwidth-rental networks. You share your unused internet connection, it is resold and routed to whoever buys it, and nothing you helped create stays with the network. In the middle sits paid private labeling, where you annotate data for one client and the dataset disappears into that company's private model. At the far end sits human feedback released as a public good: contributors compare answers, and the aggregated preference data is made openly available for building open AI. The difference is not who pays. It is who uses your work, and whether anyone beyond a single buyer ever benefits from it. A fuller side-by-side lives on the contribution networks comparison.
Where Tendril sits
Within Human-Feedback DePIN, Tendril is the variant that releases the data as a public good. People open a browser tab to lend spare compute, or tap their phone to judge which of two AI answers is better, and the human-feedback data the network produces is meant to be open and shared rather than sold off privately. Tendril also concentrates on a part of the problem the bigger labs underserve: languages most models barely speak, like Swahili, Yoruba, Hausa, Amharic, Thai, Tagalog, and Bahasa Indonesia. That focus is the point of difference we claim, not a ranking. We are describing what Tendril does mechanically, not asserting a position over anyone else in the category. For the full definition of the network, see What is Tendril.
Why this matters for open AI
If human judgment is the scarce input, then who collects it and what happens to it afterward decides who open AI is actually built for. When preference data is locked inside one company, the models it improves answer to that company. When the same data is collected in the open and shared, more of it can flow to the languages and communities that current models handle worst. That is the honest case for a Human-Feedback DePIN: not that it is bigger or faster, but that it changes who the work serves. People, not hardware, are the product, and where the product goes is the whole question.