AI Is Going to College. The Generalist Era Is Over.

There's a thing I've noticed in the community lately. Not one big announcement. Not one platform sending an email that says "the rules have changed." It's quieter than that. It's the texture of the posts.

Six months ago, the recurring question was: "Why did my queue go empty?"

Now the recurring question is: "Am I qualified enough for this work?"

That shift is not small. That shift is everything.

The First Phase Is Ending

The first phase of AI training was, in hindsight, simple. Companies needed enormous volumes of human feedback to teach their models the basics. Does this response make sense? Is this helpful? Is this offensive? Rate A versus B. Write something better. Flag the hallucination.

That work didn't require a PhD. It required literacy, judgment, and patience. And for a few years, it paid remarkably well for what it was.

That phase is not over. But it is compressing.

The volume is still there — for now. But the rates on generalist tasks are quietly sliding. The competition for those tasks has increased. And the platforms are less interested in your ability to show up and do the work than they are in your ability to do this specific kind of work, which requires this specific background, which we will now verify.

The floor is dropping. The ceiling is rising. And there is a widening gap between the two.

What "College" Actually Means

When I say the industry is going to college, I don't mean that you now need a degree to participate. I mean that the work itself has matured.

High school was: can you read a response and tell me if it's good?

College is: can you evaluate whether an AI agent's legal reasoning is sound? Can you break a multi-step coding workflow? Can you tell me not just that the output was wrong, but where in the chain of thought it went wrong, and why?

That's a different job. And it pays like a different job.

Mindrift is listing Data Science and ML Engineer roles at $90 per hour. Not aspirational. Current postings, March 2026. Handshake AI's Fellowship program — which just acquired Cleanlab to automatically audit submitted work — is recruiting physicians, lawyers, and GIS specialists at $30 to $125 per hour depending on domain. RWS's TrainAI division is running active listings for Machine Learning Specialists and physicians at $150 to $250 per hour.

Those are not unicorn numbers on a landing page designed to get your email address. Those are rates for real, active projects happening right now. And the people doing that work are not exceptional. They're just credentialed — and they knew to look.

The Agentic Shift Is Why This Is Happening

Here is the mechanical reason the industry changed, because it didn't happen arbitrarily.

The task type changed.

For years, the dominant task in AI training was what's called chat eval — a single human, evaluating a single AI response, to a single prompt. You read it. You rated it. You moved on. The work was asynchronous, isolated, and could be done effectively on autopilot by anyone above a certain reading level.

That task is being replaced — not eliminated, replaced — by agentic eval.

An AI agent doesn't answer one question. It executes a plan. It opens a browser, runs a search, writes a function, checks the output, adjusts, and continues. Your job is no longer to read the response. Your job is to watch the decision-making process — all fifteen steps of it — and determine whether the reasoning was sound at each step. Where did it make the right call? Where did it drift? Where is the logic flaw you'd never catch if you were just looking at the final answer?

You cannot do that on autopilot. You cannot do it without domain knowledge. And you cannot do it for $15 an hour, because no one with the background to do it correctly will accept $15 an hour when the alternatives exist.

That's not a moral argument. That's a market argument. And the market is adjusting accordingly.

Outlier is calling this RLHF 2.0. Stellar AI has built their entire model around it — the GitHub PR-to-task workflow, the Gold Answers standard, the paid qualifications. Mercor's CEO Brendan Foody has said publicly that the bottleneck for AI is no longer raw data — it's high-quality evaluations. The APEX benchmarks they just launched are measuring frontier model performance on real-world legal and engineering work, because that's where the gap is.

What This Means for You Right Now

I'm going to be direct here, because I think the community discourse around this tends to be either too optimistic or too fatalistic, and neither is useful.

If you are a generalist, the work is not disappearing. But it is getting more competitive, paying less for the same effort, and the platforms are investing their infrastructure improvements in the expert tiers, not the floor. You should be building your basket — multiple platforms, multiple income streams — and treating generalist gig income as the hedge, not the anchor.

If you have credentials you haven't leveraged, this is the most important thing I can tell you: there is a specific version of this industry where your background commands serious money, and most people in it have no idea it exists. If you are an attorney doing generalist work on Outlier at $25 an hour, I need you to go look at what Handshake AI's Fellowship is currently paying for legal reasoning evaluations. If you are a physician doing the same, I need you to look at what RWS TrainAI is listing for Safety and Accuracy review right now. The gap between what generalists are earning and what verified professionals are earning has never been wider.

If you're technical — Python, data science, ML background — Stellar AI and Mindrift are where you should be focused. The GitHub PR workflow on Stellar is a different kind of work than anything you've done on the mainstream platforms, and it pays to match. The agentic workflows at Mercor require what they're calling "virtual coworker management" — setting up the management processes for AI agents to execute code, not just writing it yourself. It's a new skill. It's learnable. And it sits at the top of the current pay tier.

The Part Nobody Is Saying Out Loud

Here's what I keep thinking about when I look at these numbers.

The workers who are going to do well in the next phase are not necessarily the most technically elite. They're the ones who stop treating this industry as a single, monolithic gig economy and start treating it as a series of adjacent markets — each with different entry requirements, different compensation structures, and different timelines.

The generalist gig market and the expert eval market are running in parallel right now. One is contracting. One is expanding. The question is whether you are looking at both.

I've been in this long enough to have gone through the full cycle — the dopamine hit, the golden handcuffs, the empty queue, the humility. That cycle doesn't go away in the expert tier. The volatility doesn't disappear. Mercor workers report dry spells every few months when budget cycles reset. Handshake has a waitlist problem because the pay attracted too many applicants at once. RWS has a support gap that would frustrate anyone.

The difference is that the ceiling, when the work is there, is substantially higher. And the worker who has positioned themselves for expert-tier work has options that the generalist doesn't.

The industry is not dying. It is growing up.

It spent its early years hiring everyone it could find to do basic tasks at a premium because it needed the volume fast. It is now doing what every maturing industry does: it is getting selective. It is building credential requirements. It is creating quality tiers.

The workers who see that transition coming — who are building toward it instead of waiting to be surprised by it — are the ones who will still be earning from this twelve months from now.

The ones who don't see it coming will keep being surprised every time a project ends.

I'm going to be writing about this in depth over the next few weeks — the specific platforms, the specific roles, and the specific moves that matter right now. If you want to be notified when those go up, the newsletter is the fastest way to find out: breakingeven.online/newsletter.

And if you want to understand the full landscape of what's available right now — from floor to ceiling — the 2026 Platform Tier List is the best place to start.