What Everyone Still Gets Wrong About the AI Data Market

+

When you think of AI data, what do you picture? Rooms full of people labeling cat pictures?

That’s so 2022! Let me explain how this market really works…

AI Data Isn’t “Data Labeling” Anymore

“Isn’t data labeling a commodity?” I’ve heard that 100 times.

People picture workers in the developing world labeling pictures as “cat” or “not cat”. This may have been true in 2022, but not anymore.

Today, AI data companies aren’t doing simple labeling. They’re hiring world class experts to generate unique data.

Companies like my investment, Micro1, are hiring top lawyers for $400 an hour to teach AI models about corporate law.

AI labs have already ingested the entire open internet. If they want to keep improving their models, they need unique data.

There Are Many Kinds of AI Data

AI data goes way beyond text.

Different companies specialize in different types. The many niches provide a huge opportunity for new startups:

Egocentric video. This is video shot from the perspective of a person performing a task, like folding laundry. This is very useful for robotics.
Audio recordings. Really helpful for voice models. Many AI data companies are particularly focused on getting recordings of foreign languages.
Tool use. AI models can watch us create spreadsheets or answer Slack messages. This helps us automate white-collar work.

We Will Always Need More Data

“What happens when we reach AGI? Aren’t these data startups toast?”

Hardly. We can always teach AI and robots to do something else useful!

And in order to teach them, we need more data.

Maybe we want to train a robot to make a new machine part. To do that, we need video showing how the manufacturing process works.

Data Labs Have Thousands of Potential Customers

In the early days, data labs sold to a handful of frontier model companies. That presents a risk if those companies cut back.

But now, data labs are moving into another huge market with thousands of potential customers: agent evals.

Companies all over the world are deploying AI agents. But how do you know if your agent actually works?

You probably don’t.

Agent evals measure how well your agent works and help you improve it.

Say you use an agent to approve loan applications. A data lab can send loan experts to make sure your model is handling applications correctly.

Wrap-Up

The AI data market has produced growth like nothing else. If you understand it, you have an advantage on other founders and investors.

Forget commodity data labeling. Today, it’s about finding top experts to produce unique information. And we always need more.

Training the superintelligence never ends.