Every training method builds a lens. The question is whether that lens is shaped by human convenience or by the structure of the world itself. That distinction, more than any architectural choice, is what separates imitation from genuine understanding.
How we got here
The earliest AI systems were mostly built by hand. People wrote rules, logic trees, and expert systems, so the machine could only “know” what humans had explicitly placed inside it.
Modern machine learning changed that by letting systems learn patterns from data rather than from fixed rules. Pretraining made this especially powerful: instead of telling a system how the world works, we exposed it to vast amounts of data and let it compress regularities into internal representations. That was the first big step from “software that follows instructions” to “software that builds a statistical model of reality.”
Then came supervised fine-tuning, or SFT. A pretrained model may already contain broad pattern knowledge, but SFT teaches it how to behave in the format people actually want: answer questions, follow instructions, stay relevant, and sound coherent in dialogue. That is a large part of why ChatGPT became the first AI system that felt broadly useful to ordinary users. It was not merely that the underlying model could continue text; it was that the model had been shaped into something that could act like an assistant.
That distinction matters. Pretraining gives a model raw competence, but SFT gives it social usability. One teaches “what tends to go with what”; the other teaches “how humans want this knowledge packaged.” In practical terms, SFT was a massive speed-up for making general models useful. In philosophical terms, though, it also pulled the system closer to human-approved answers and formats, which means it pulled the model closer to our way of expressing the world, not necessarily the world itself.
Why reinforcement learning changed the conversation
This is where reinforcement learning entered as the next big promise. SFT mostly teaches imitation: it says, in effect, “Here is what a good answer looks like; produce more things like that.” RL is different. It says, “Act, observe what happens, and improve according to a score.” That changes the search process. Instead of staying near the surface of human demonstrations, the model can sometimes discover strategies or structures that people never explicitly wrote down.
AlphaGo Zero became the iconic example of this shift because it learned from self-play starting from random play, without human game data, and then surpassed earlier human-seeded AlphaGo systems. In the graph from the accompanying research, the supervised-learning curve begins much higher, while the reinforcement-learning curve starts dramatically lower, then catches up, overtakes it, and eventually climbs beyond the AlphaGo Lee line (the human-data-trained system that defeated world champion Lee Sedol in 2016, effectively representing the ceiling of human-level play). That picture captures something profound: human guidance can give a fast start, but freer search can sometimes escape imitation and find stronger solutions.
That insight is larger than Go. It suggests that if we train a model only to copy human examples, we may cap it near the edge of what humans already know or already express well. RL mattered because it showed a route beyond demonstration. In a narrow, well-defined world like Go, that worked spectacularly because the objective was clean: win or lose. But the moment we move from Go to truth-seeking, science, or real-world reasoning, the problem becomes harder. In open-ended domains, the reward is rarely as crisp as “win the game,” and reward misspecification becomes a central issue: the system may optimize the score we gave it rather than the deeper thing we meant (or should mean).
Today’s frontier language models are already a kind of hybrid. Pretraining gives them broad priors, SFT and preference optimization shape them into assistants, and inference-time reasoning or internal search lets them test and revise candidate lines of thought before committing to an answer. In that sense, there is already a small echo of the AlphaGo story inside modern LLMs: not pure imitation, but search layered on top of learned priors. The next leap is likely to be making that search world-model-native rather than token-native. Right now, much of the search still happens over words, intermediate chains, and symbolic traces. A more powerful system would search over a richer internal model of the world itself.
Why bias and local optima remain
This leads to the core thesis: every training method bakes in a lens. SFT bakes in the lens of human labels and approved answers. RL bakes in the lens of human-designed rewards. Even pretraining bakes in the lens of data collection, tokenization, architecture, objective choice, and compute limits. So when we say a model “learns the world,” what we really mean is that it learns the world through the narrow channel we constructed for it. This would be fine if humans were reliable arbiters of truth. But we are not. We revise, contradict ourselves, and systematically err in ways we do not notice until later. The lens we give a model is not a window onto the world, it is a record of what we thought the world looked like when we built it.
That is why local optima are such a useful metaphor here. Large models do seem to absorb many real regularities from massive datasets, and scaling-law work showed fairly regular improvement as model size, data, and compute increase. This is one reason datacenters keep expanding: more compute lets labs train larger systems on broader datasets and search wider regions of parameter space. A huge pretrained model therefore gets an extraordinary head start. It already contains many useful compressions of language, images, patterns, and human concepts before any task-specific tuning begins.
But the same head start can also become a trap. If pretraining or post-training hardens the wrong abstractions, the model may inherit bad grooves in representation space that are difficult to escape later. SFT gives speed, coherence, and usability, but it can also make the model cling to human-shaped shortcuts. RL can help the model go beyond imitation, but it can also optimize the wrong proxy if the reward is not aligned with what we actually care about. So the real picture is not “SFT bad, RL good.” It is that both are useful, both are biased, and both can get stuck.
This is even more serious in open-ended settings than in games. In Go, the reward is clean and closed. In scientific discovery, truth-seeking, or general reasoning, the reward is noisy, delayed, partial, and often contested. That is why moving from human labels to world feedback is so important. In open domains, the only scalable correction mechanism is reality itself: experiments, consequences, interventions, consistency across modalities, and repeated contact with the world. Curiosity, intrinsic motivation, and embodiment matter so much because they are ways of generating better feedback when no single human-written reward can capture the target well.
Why unconstrained training is not the answer
At first glance, the obvious solution seems simple: remove the constraints and let the model “just find the truth.” But that breaks down the moment you ask what learning actually requires. A learner with literally no constraints would have no reason to privilege one pattern over another, no way to know what should remain stable, and no basis for deciding which similarities matter. To learn at all, a system must already be biased toward some structures and not others.
This is the deepest reason unconstrained training is not a real solution. The broader point associated with No Free Lunch style arguments is that no learning rule is best across all possible worlds unless you assume something about the kind of world you are in. In plain language: you cannot learn everything equally well from nothing. Some inductive bias is unavoidable. The real question is not whether there will be constraints, but which constraints are worth having.
There is a second problem too. Even if we had enormous amounts of data, observation alone does not always reveal the true underlying structure of reality. What we usually see are surface outcomes, while the deeper causes, hidden variables, and unrealized alternatives remain out of view. A model may therefore learn to predict the world extremely well without actually recovering the real mechanism that generated what it sees. Different hidden structures can often explain the same evidence, so accuracy by itself does not guarantee that the model has found the uniquely correct ontology. Prediction gets us closer to truth, but it is not identical to truth.
This matters because it changes the goal. We should not aim for no constraints. We should aim for better constraints. Human approval, narrow labels, and simplistic reward proxies are often weak constraints because they reflect our convenience more than nature’s structure. Better constraints are things like temporal coherence, causal consistency, multimodal grounding, action-consequence feedback, and uncertainty calibration. Those do not remove structure from training; they replace arbitrary structure with more reality-coupled structure.
That is also why “truthful AI” is harder than many people assume. Truth is not a scalar target sitting in the environment waiting to be directly optimized. What we usually have are partial observations, imperfect sensors, delayed consequences, and competing explanatory models. So the path to truth is not unconstrained freedom. It is disciplined exposure to reality under constraints that make false world models harder to maintain.
The move toward multimodality, embodiment, and world models
Once you see the problem that way, the current research frontier becomes more coherent. If language-only training gives a narrow slice of reality, then adding more modalities helps because the model can cross-check one channel against another. Vision can correct language. Sound can correct vision. Time can correct static snapshots. Action can correct passive observation. A model that learns from text, images, audio, video, and interaction has a much better chance of grounding its concepts in something deeper than chat patterns.
Embodied learning pushes this even further. When a system acts in an environment, it does not just absorb correlations; it sees consequences. Self-supervised embodied active learning and embodied concept learning were built around exactly this idea: representation learning improves when the agent can explore, observe outcomes, and use interaction to shape what it learns. Embodiment matters because it gives the model a route from “pattern recognizer” to “world participant.”
Self-supervised learning is the natural companion to this. Instead of feeding the model human-labeled answers, we let the structure of the data provide much of the signal. The model predicts missing parts, future states, consistent views, or stable abstractions, and in doing so it learns the regularities already present in the world. MIT researchers have even noted parallels between self-supervised learning and how the brain may learn about world structure. This does not magically produce truth, but it does reduce dependence on human answer keys.
World-model research takes the next step. The aim is not just to predict the next token or classify an input, but to learn how states of the world evolve and how actions alter those states. That is why this line of work feels so important for the future. If intelligence is partly the ability to manipulate internal models of reality, then AI needs more than a surface-level language prior. It needs a latent simulator of how the world works.
JEPA sits squarely in that direction. Its basic idea is to predict the representation of hidden or future parts from visible context rather than reconstruct every pixel or generate the next token directly. LeCun’s attraction to this approach is clear: if thinking happens in abstract continuous representation space, then a good model should learn to operate in that space instead of being forced to solve everything at the level of raw words or pixels. JEPA is not unconstrained, but it is less tied to human-labeled outputs than ordinary supervised learning and less tied to raw token continuation than standard LLM training.
This is also where a broader philosophical point comes back. Nature has its own data-generating process. Our formulas, labels, and benchmarks are only approximations to it. So the closer we can make the training loop to actual contact with that process, the better chance we have of learning something that is not merely human-convenient, but world-revealing.
Implications for labs and a research agenda
If this framing is right, the future of frontier AI is not "abandon language models and start over." It is a quieter but more fundamental shift: keep the gains from scale, but redirect them toward systems increasingly grounded in the world rather than optimized for token prediction or human approval. Scaling still matters — the evidence on compute and data is unambiguous — but scaling a next-token predictor is not the same thing as scaling a truth-seeking system. That distinction is going to matter more and more as the easy benchmark wins dry up.
The most honest implication is that pretraining and SFT should stay, but their roles should be demoted. Pretraining gives a model its priors; SFT shapes how it talks to people. Neither should be mistaken for how a model learns what is actually true. That work has to come from somewhere harder: robotics, scientific instruments, simulators, long-horizon video, experimental environments — places where the world pushes back. Text scraped from the internet is a record of what humans said. It is not a record of what the world does. Closing that gap means building training loops where contact with reality, not contact with human approval, is the main correction signal.
Reinforcement learning points in this direction, but only if the reward is honest. In closed domains with crisp feedback — games, formal proofs, verified code — RL can be spectacular, and recent results bear that out. In open-ended domains, though, a scalar reward is usually a proxy for something deeper that it does not fully capture, and a model smart enough to optimize a proxy is smart enough to game it. The answer is not to abandon RL but to couple it to richer feedback: uncertainty, information gain, consistency across modalities, the kind of intrinsic curiosity that makes a learner seek out what it does not yet understand rather than confirming what it already knows.
None of this is easy. Physical data is slow to gather. Embodied training loops are expensive and brittle. The language priors baked into today's models are deep and hard to dislodge. But the difficulty is not an argument against the direction — it is just an accurate description of the terrain.
The final recommendation is not "train without constraints." It is "train under constraints that are increasingly answerable to reality." Not no lens, because no learner gets to have no lens, but a lens shaped less by our labels and more by the structure of the world itself.