A 2.2B AI firm believes data labeling is over, and a bigger transformation has begun

AI progress moves from scale to experience

For years, we treated AI like a simple math problem: bigger models, more data, more compute. That mindset gave us powerful chatbots and coding assistants, yet now we are hitting diminishing returns.

The next breakthroughs will not just come from size but from giving models experiences. Instead of only predicting text, they will learn by acting inside digital worlds and refining skills.

Inside the AI classrooms, teaching agents to act

Think of new AI training setups as classrooms built from code. Researchers construct interactive worlds where agents can explore, try out ideas, fail safely, and try again.

In these spaces, models observe a state, choose an action, and receive feedback on how well they performed. Over millions of loops, they slowly discover strategies just like a student practicing until the patterns finally click.

Why simple data labeling has reached its limits

Traditional data labeling focused on simple tasks, such as tagging an image, classifying a sentence, or sorting a review. It was essential for the first wave of machine learning, but it is no longer sufficient for agent-style systems.

These models need data that captures real workflows, not just isolated snippets. That means tracking how people actually research, draft, decide, and iterate across complete projects.

How Turing’s CEO sees the new data frontier

Jonathan Siddharth, who leads Turing, an AI training firm valued at around $2.2 billion, says the classic data labeling business is already coming to an end. On a recent podcast, he argued that data needs have shifted.

Early systems lived on tagged images and short classifications. Now, frontier labs want training data that mirrors complex knowledge work and supports reinforcement learning at every stage.

From crowd workers to research partners for labs

In this new landscape, AI training companies cannot just be anonymous outsourcing vendors. Siddharth says they must become proactive research partners sitting alongside frontier labs.

Instead of only supplying labeled datasets, they help design experiments, build an evaluation framework, and assemble expert reviewers.

The value comes from understanding how the models behave, suggesting improvements, and closing the loop between raw data and real capability gains.

Reinforcement learning environments become the new training ground

Reinforcement learning environments sit at the center of this shift. Inside the model, an AI agent observes a situation, chooses an action, and receives a numerical reward.

Over time, it learns which behaviors lead to success. What changes is that training turns interactive. Models move beyond static prompts and begin to test hypotheses, handle surprises, and adapt to new situations as environmental conditions evolve.

Coding sandboxes turn chatty models into real engineers

Code generation is a perfect example. A model that only discusses programming can sound intelligent yet still fail in real-world repositories.

Drop that same model into a coding sandbox where it can read files, run tests, debug errors, and retry, and its behavior transforms. It shifts from sounding like a helpful tutor to acting more like an autonomous engineer that steadily improves.

Teaching agents to survive the messy open internet

The open internet is even messier than a legacy codebase. Pop-ups, login walls, broken links, and contradictory instructions are everywhere. Humans breeze past these annoyances almost without thinking.

For AI, we need training runs that recreate this chaos so agents can practice resilience. They must learn to refresh pages, change strategies, and complete multi-step tasks, even when the browser refuses to cooperate.

Simulated worlds prepare AI for high-stakes decisions

Some of the most essential AI classrooms will never be accessible to the public web. Governments and large companies are already developing secure simulations in which agents can practice sensitive missions.

Imagine disaster response models planning evacuations across ports, roads, and supply chains. In a safe, simulated world, the agent can fail thousands of times while humans observe and determine which strategies emerge before they trust any plan in reality.

The boom and dark side of AI training work

Behind the scenes, a global workforce keeps these systems up to date. Companies such as Scale AI, Surge AI, Mercor, and Turing hire contractors to grade answers, compare responses, and follow complex evaluation rubrics.

Some suppliers and vendors say elite annotators can earn high hourly rates, but reporting shows pay varies widely across the sector and many contractor assignments are irregular, emotionally demanding, and closely supervised.

Data labeling giants face fraud and black markets

Business Insider reported more than 100 social groups advertising verified training platform accounts and said platforms and Meta removed dozens of these groups as part of enforcement actions.

Sellers pitch verified logins, buyers hide behind virtual private networks, and both sides risk losing income or exposing personal data.

Internal documents from leading firms describe long battles with duplicate accounts, spam submissions, and taskers secretly relying on chatbots to complete work.

Why the next bottleneck is rich interactive environments

All of this reveals a more profound truth: the hard part in modern AI is no longer scraping more text from the web. The emerging bottleneck is building rich, interactive training environments that teach models practical workflows while incorporating rigorous evaluation and fraud-detection systems.

That means better simulations, more uncompromising evaluations, and more innovative detection tools. Those who build the best classrooms for AI will likely shape the next decade of progress.

And if you want to see how these shifts are influencing global competition, take a look at China’s growing focus on AI superintelligence, which is sparking new concern in the U.S.

What this shift means for workers and the AI race

For workers and companies, this shift is both unsettling and full of possibilities. Simple labeling tasks become more streamlined as agents learn from complex workflows guided by smaller groups of experts.

Yet new roles emerge, designing environments, auditing behavior, and stress testing models in realistic scenarios.

As I watch this unfold, it feels like we are moving from raw ingredients to fully completed training grounds.

And if you want to see how these evolving AI roles are shaping the broader debate online, check out how woke artificial intelligence sparked a new battle over online free speech.

What do you think about an AI firm owner believing that data labelling is over and turning into something big and new? Please share your thoughts and drop a comment.

Read More From This Brand:

• Trump cracks down on woke artificial intelligence with new orders
• UW scientists use artificial intelligence to open new drug possibilities for proteins
• Microsoft’s AI leader says creating artificial superintelligence is not a goal for the company

Don’t forget to follow us for more exclusive content on MSN.

If you liked this story, you’ll LOVE our FREE emails. Join today and be the first to get stories like this one.

This slideshow was made with AI assistance and human editing.