The Human Is the Product, Not the GPU
Published
In most stories about AI, the hero is the hardware: bigger GPUs, bigger data centers. Tendril is built on the opposite belief. The scarce, valuable input is human judgment, the everyday sense of which answer is actually better. This essay explains why the human is the product, not the GPU, and what that means for how AI gets built.
In most stories about AI, the hero is the hardware. The headlines are about bigger chips, bigger clusters, and bigger data centers. The unstated assumption is that intelligence is something you buy by the rack: pile up enough GPUs and a smart machine falls out the other end. Tendril starts from the opposite belief. The scarce input is not compute. It is human judgment, the everyday sense of which of two answers is actually better.
What the machine cannot supply
A model can generate a hundred answers to a question in a second. What it cannot do, on its own, is know which of those answers a real person would find helpful, honest, and clear. That judgment has to come from people. When a model learns to be useful rather than merely fluent, it is learning from human preferences: this answer over that one, this tone over that tone, this explanation a parent could follow over a wall of jargon.
Those preferences are not lying around for free. They have to be collected, one comparison at a time, from people who actually use the language and live in the context the question comes from. A warehouse full of chips does not produce a single one of them. People do. That is the sense in which the human is the product. The compute is a tool. The judgment is the thing of value.
Why this flips the usual picture
If compute were the whole game, AI would belong, permanently, to whoever owns the most hardware. But if the rare ingredient is human judgment, then the people who supply it are not bystanders to AI. They are upstream of it. The quality of a model is capped by the quality and breadth of the human feedback that shaped it.
That has a practical consequence. A model trained mostly on the judgments of one narrow group will be good for that group and worse for everyone else. To build AI that serves more people, you need judgments from more people: more places, more languages, more kinds of daily life. You cannot buy that with a bigger budget for chips. You earn it by making it easy for ordinary people to contribute.
How anyone contributes
This is exactly what Tendril is for. The work is small and concrete. You see two AI answers to the same question and you tap the one that is better. It takes seconds, and it does not require any technical skill. Each tap is a single human-preference judgment, the precise signal that teaches a model what better actually means to a person like you. We call this surface Tap to Train.
You do not have to be an engineer, and you do not have to understand how a model works inside. You only have to know, as a human, which answer you would rather have received. Multiply that across enough people and enough taps and you get something no hardware budget can produce on its own: a broad, honest record of human preference that open AI can be built on. For TEND Points, you earn recognition for the contribution. TEND Points are not money.
The machines get the attention. The judgment is what matters. If that idea appeals to you, the next piece is why this data should belong to everyone: read Open AI as a Public Good, or just open the app and start tapping.