Generalist human data is still needed, more than ever

We’d like to define a “generalist” in the world of human data as someone who uses their understanding of the world to create inputs that explain human behavior and help automated systems better replicate how people think and act in society.

Alternatively we have “specialists” and “experts”. These folks serve a similar purpose of using their intelligence to help improve automated systems. Except instead of applying a broad knowledge base, they apply a deep, focused comprehension of a specific domain related to human behavior, enabling highly precise advancements within that narrow area.

Why This Work Looks Simple, and Why That’s Misleading

Generalist work is easy to underestimate because it looks so simple compared to specialized tasks. You see it in homes, offices, warehouses, kitchens; everyday people doing everyday things. There's nothing enthralling about it. You don't need a PhD, or any degree for that matter. And because it looks familiar, people assume it must not be as important.

But this is actually very important to AI. As a matter of fact, it’s essential.

Generalist human data teaches both language models and robots how the world actually works via real environments where humans already live and operate, not simulated ones. This is what draws the difference between systems that look impressive in demos from systems that function when it matters in a real-life scenario. ‍

How Generalist Human Data Built LLMs

We saw this early on with large language models. Before LLMs became the go-to for everyday questions and therapy sessions, progress became stagnant. The problem wasn't the lack of internet data, it was the lack of human ground truth. The real progress came when large volumes of people started interacting with models across everyday tasks, correcting mistakes, ranking outputs, and applying judgment in situations that were never fully specified. That work taught models how humans actually communicate, reason, and express what they want. Without it, LLMs wouldn't feel usable at all.

Early language models had over twenty years of internet text to train on, but it wasn’t just scraping webpages that made them useful. What changed was when humans were asked to review model answers to common questions and say which response was clearer, more helpful, and the most accurate. Generalists wrote ideal responses, flagged misleading or unsafe ones, and showed the model how a human would respond. That dependency models had on generalist judgment is what turned LLMs into reliable tools.

Robotics Makes This Dependency Impossible To Ignore‍

Robots exist in physical space. They will be interacting with not just humans, but sensitive items in your home, your animals, and even children. They touch objects, navigate clutter, use tools, and deal with real constraints. This means not only extreme diligence, but intelligence as well. The real world was not built in a way that optimized humanoid convenience. Every home is different. Every office has its own nooks. Objects vary in ways you can't anticipate. Small inconsistencies pile up until something breaks; except when something breaks, it won’t appear as a message on your screen saying “Something went wrong. Try again.”. The stakes will be much higher. 

The most valuable data in robotics comes from people who can operate naturally across different tasks and contexts, who adapt easily, and who apply judgment when instructions are vague or incomplete. That adaptability is the signal. That signal is different in everyone’s autonomy; there’s no ground truth until you gather that “internet-scale” amount of data so that you find some equilibrium across how the global population moves through their waking hours.

More Autonomy, More Generalists

There's this persistent idea that generalist human data is just a temporary need, something you use early on, then phase out as systems get smarter. It’s interesting, the opposite is actually happening. As models improve, they need more human data. New tasks, new environments, post-training corrections, safety validation, failure recovery; all of these depend on fresh human judgment. But why do we need those new data inputs? The more autonomy you introduce, the more room that allows for humans to tap into their creativity. Once we aren’t spending hours a day on simple tasks, what will we do with that free time? Some will have more time for their family, the outdoors, or the bar; all reasonable assumptions. Many will also spend those extra hours finding ways to make the world a better, more efficient environment. Once we reach those new efficiencies, models need to be taught those new inputs. As long as models are expected to operate in real-world environments, generalist human data will be a permanent requirement.

Even if you ignore physical AI entirely and solely look at LLMs, that trajectory still holds. The difference between a decent response and a genuinely helpful one comes down to tone, intent, and context. These aren't things that emerge from sheer volume alone. They have to be taught. It’s the same way humans are taught by other humans - it’s multifaceted.

And what's interesting is as more people use these systems in diverse situations, preferences don't converge, they fracture. What works for one person feels wrong to another. The only way to navigate that is through ranges of human feedback that captures all those contradictory expectations. So with mature language models, the generalist work evolves.

What Its Worth

This compensation for this work should align with the significance of the work. For many people, generalist human data provides meaningful income and sometimes opens doors to more specialized roles over time. Beyond the pay, the impact compounds. Errors that are caught now prevent catastrophic failures down the line.

That's what makes this work purposeful.

Generalist human data is the foundation that intelligence is built on.

If you're doing this work, you're helping define how intelligent systems understand and operate in the human world. That’s why it matters more than ever.