AI's New Frontier: From Data to Experience

For the past decade, progress in artificial intelligence has been measured by scale: bigger models, larger datasets and more computing power. That approach has produced striking breakthroughs in large language models (LLMs). In just five years, AI went from models like GPT-2, which could barely mimic coherence, to systems like GPT-4, which can reason and hold a meaningful conversation. And now early prototypes of AI agents that can navigate codebases or browse web pages point to an entirely new frontier.

But size alone can only take AI so far. The next leap will not come only from making models bigger. It will come from combining ever higher-quality data with the worlds we build to train models. And the most important question becomes: what do classrooms for AI look like?

Over the past few months, Silicon Valley has placed its bets, and labs are investing billions in building such classrooms, known as reinforcement learning (RL) environments. These environments let machines experiment, fail and improve in realistic digital spaces.

The story of modern AI has unfolded in eras, each defined by the type of data models consumed. First came the era of pretraining on internet-scale datasets. This publicly available data let machines imitate human language by recognizing statistical patterns. Then came data combined with reinforcement learning from human feedback, a method that uses crowdsourced workers to rate LLM responses, which made AI more useful, responsive and aligned with human preferences.

Today, data is still the foundation. It is the raw material from which intelligence is built. But we are entering a new phase in which data alone is no longer enough. To unlock the next frontier, we must pair high-quality data with environments that allow unlimited interaction, continuous feedback and learning through action. RL environments do not replace data; they amplify what data can do, letting models apply knowledge, test hypotheses and refine behavior in realistic conditions.

In an RL environment, a model learns through a simple loop: it observes the state of the world, takes an action and receives a reward indicating whether that action helped reach the goal. Over many iterations, the model gradually discovers strategies that lead to better outcomes. The important shift is that learning becomes interactive: models don't just predict the next token, they improve through trial, error and feedback.

For example, language models can already generate code in a simple chat setup. Put them in a live coding environment where they can pull in context, run their code, debug errors and improve their solution, and something changes. They move from advising to solving problems autonomously.

In a world run by software, the ability of AI to generate and test production-grade code across vast repositories will be a major change in capabilities. That leap will not come only from larger datasets; it will come from immersive environments where agents can experiment, stumble and learn through iteration, much as human programmers do. Real-world development is messy: programmers have to deal with underspecified bugs, tangled codebases and vague requirements.

Training AI to handle that mess is the only way it will ever move from producing error-prone attempts to producing consistent, reliable solutions.

Navigating the web is messy too. Pop-ups, login walls, broken links and outdated information are woven into everyday browsing workflows. People handle these glitches almost instinctively, but AI can develop that ability only by training in environments that mimic the unpredictability of the internet. Agents must learn to recover from errors, recognize and get past user interface obstacles, and complete multi-step workflows in widely used applications.

Every major leap in AI has relied on invisible infrastructure: annotators labeling datasets, researchers training reward models, and engineers building the scaffolding that lets LLMs use tools and take actions. Finding large volumes of high-quality datasets was once the bottleneck in AI, and solving it set off the previous wave of progress. Today the bottleneck is not data; it is building RL environments that are rich, realistic and genuinely useful.

The next stage of AI progress will not be an accident of scale. It will result from combining a solid foundation of data with interactive environments that teach machines to act, adapt and reason in complex real-world scenarios. Coding sandboxes, OS and browser playgrounds, and safe simulations will turn prediction into competence.