The "Super Learner" Bet: Why Reinforcement Learning Could Be the Next Big Path Beyond Large Language Models

For the past several years, the story of artificial intelligence has been a story about one thing: bigger language models fed on more text. Every few months, another giant model arrives, trained on a larger slice of the internet, and the results get a little better. It has been a powerful formula. But a growing number of the people who helped build modern AI are quietly asking a different question, what if reading the whole internet is not the same as understanding how to act in the world?

That question sits at the center of a new venture from David Silver, a former scientist at DeepMind. Silver is now building a startup called Ineffable, and its goal is bold: an AI "super learner" built on reinforcement learning (RL) rather than on large language models. Instead of learning by predicting the next word in a sentence, this kind of system would learn by taking actions, receiving feedback, and improving over time. It is a different road to machine intelligence, and it may turn out to be one of the most important AI stories of the decade.

Two Very Different Ways to Build a Thinking Machine

To understand why this matters, you need to understand the difference between the two approaches.

Large language models (LLMs) learn from text. They are shown enormous amounts of writing, books, articles, code, conversations, and they get very good at guessing what comes next. Do that billions of times, and something surprising happens: the model starts to answer questions, write essays, and reason through problems. The strength of this method is scale. There is a lot of text in the world, and it is cheap to collect.

Reinforcement learning works differently. An RL system is more like a player in a game than a reader in a library. It tries things. Some attempts work; some fail. It gets a signal, a reward, telling it how well it did. Slowly, it learns a strategy that earns more reward. This is the approach behind machines that mastered complex games and control systems, and it has one feature that text training does not: it learns from experience, not from someone else's words.

Silver's bet, as described in the interview, is that this second path has been underrated. LLMs are excellent at reproducing and remixing human knowledge. But RL systems are built to discover, to find strategies that no human wrote down, because no human had found them yet.

What a "Super Learner" Actually Means

The phrase "super learner" is doing a lot of work here, and it is worth unpacking.

A super learner is not simply a model that scores higher on a test. It is a system designed to keep getting better at learning itself, to transfer skills from one problem to the next, to need less data over time, and to improve through its own trial and error rather than through endless new piles of human text.

Think of the difference between a student who memorizes every textbook ever written and a student who is handed a difficult puzzle, no instructions, and figures it out. The first student can recite enormous amounts of information. The second student can handle something nobody has seen before. Today's LLMs lean heavily toward the first profile. The super learner idea leans toward the second.

If that works, the implications are enormous. AI systems today are mostly very well-read assistants. A super learner would be something closer to an autonomous problem-solver, one that can be dropped into a new environment, given a goal, and left to work out how to reach it.

Why This Is Happening Now

Three forces are pushing the industry to look beyond the LLM-only playbook.

1. The limits of "more text"

The supply of high-quality human writing is finite. Model builders have already scraped much of what is publicly available. As the easy data runs out, the returns from simply scaling up text training start to flatten. RL offers a way to generate new learning signal without needing new text.

2. Text is not the same as the world

Most valuable real-world problems are not language problems. They are action problems: routing a delivery fleet, tuning a factory line, managing a power grid, discovering a new material, negotiating a supply contract. These require sequences of decisions with consequences, which is exactly the shape of a reinforcement learning problem.

3. Agents need to act, not just answer

The industry has been racing toward "agents", AI systems that don't just respond to a prompt but carry out multi-step tasks. Answering is a language skill. Acting well is a learning skill. That gap is where RL becomes essential.

What This Means for Businesses

For most companies, the practical question is not which research path wins. It is what you should do while the race plays out. Here are the takeaways that matter right now.

• Start collecting feedback data, not just content. The fuel for RL-style systems is records of what happened, what was tried, and how well it worked. Support tickets with resolutions, pricing decisions with outcomes, logistics routes with delivery times, these are assets. Many companies throw them away.
• Look for problems with clear goals and fast feedback. RL-style learning works best where you can measure success quickly and repeatedly. That makes scheduling, pricing, inventory, energy use, and quality control far better candidates than fuzzy, slow-moving strategy questions.
• Don't rip out your LLM stack. The realistic near future is hybrid. Language models are unmatched for understanding instructions, summarizing, and talking to people. Learning-based systems are better at optimizing and acting. The winning architecture will likely combine both.
• Build simulation where you can. A learning system needs somewhere safe to fail. Digital twins, sandboxes, and test environments become infrastructure, not luxuries.
• Track the talent signal. When leading researchers from top labs start their own ventures around a specific idea, that idea usually becomes a hiring market and a funding theme within 18 months. Watching where the best people go is often a better leading indicator than watching benchmark charts.

The Bigger Picture: A Fork in the Road for AI

It is tempting to frame this as LLMs versus reinforcement learning, but that framing misses the point. The more useful way to see it is as a widening of the field.

For a few years, the industry converged on one recipe. Now it is splitting. One camp keeps scaling language models and squeezing more out of them with better training and tooling. Another camp, represented by ventures like Ineffable, argues that the missing ingredient is not more data but a better learning process, one that lets a system improve through its own experience.

If the second camp is right, the consequences reach further than any single product. A system that genuinely learns from experience would be less dependent on human-written knowledge, which means it could keep improving in areas where humans have little to teach it. That is exciting for science and engineering. It is also the thing that makes regulators and safety researchers nervous, because a learner that improves on its own is harder to predict than a model that mainly reflects what people already wrote.

The safety tension nobody can avoid

Reinforcement learning has a long-standing problem: systems get very good at the reward you actually gave them, not the outcome you intended. Give a system a proxy for success and it may exploit that proxy in ways you never imagined. As these systems take on more real-world control, reward design becomes a business and governance issue, not just a technical one. Any company deploying learning-based agents should expect to spend real effort on defining what "good" means, measuring it honestly, and watching for shortcuts.

What to Watch Next

A few signals will tell you how seriously to take this shift:

• Demonstrations of transfer. The clearest proof of a super learner would be a system that masters one task and then picks up a related one much faster. Watch for evidence of learning speed improving over time, not just final scores.
• Data efficiency claims. If learning-based systems can reach useful performance with far less data than language models need, the economics of AI change quickly.
• Enterprise adoption in operational settings. The first real wins will likely show up in factories, logistics, energy, and finance, places with clear metrics and repeatable decisions.
• Hybrid architectures. Expect to see language models used as the "interface" and learning systems used as the "engine" inside the same product.

The Bottom Line

The most important AI story of the next few years may not be about a bigger model. It may be about a different kind of model entirely, one that learns the way living things do, by trying, failing, and improving. David Silver's new venture, Ineffable, is a clean expression of that idea: a super learner built on reinforcement learning as an alternative to the language-model path that has dominated the field.

Whether or not this particular company succeeds, the direction is telling. The AI industry is no longer betting on a single formula. It is hedging, experimenting, and reopening questions that many people assumed were settled. For businesses, that means the smart move is not to pick a side but to prepare for both: keep using language models where they shine, and start building the data, feedback loops, and simulation environments that any learning-based system will need to be useful.

The race to build machines that learn from experience has restarted. This time, the prize is not a better chatbot. It is an AI that can figure things out on its own.