Tabula Rasa Learning: From Imitating Human Thought to Searching for What Actually Works

When I draw the distinction in this figure, I am contrasting two fundamentally different approaches to building intelligent systems.

The first approach is imitation, or behavior cloning. It trains a model on examples, programs, solutions, explanations, and reasoning traces produced by humans. The second approach is first-principles search. It begins with an objective, an environment, and relatively weak assumptions, then searches for strategies according to whether they succeed — not according to whether they resemble human behavior.

My central claim is not that human-generated data is useless. Human knowledge is extraordinarily valuable. It can compress centuries of experimentation into a tractable starting point. My claim is narrower and more consequential:

A system trained primarily to reproduce human-generated solutions will tend to concentrate its probability mass around the kinds of solutions humans already produce.

That tendency may remain even when we add more training data, more test-time computation, longer reasoning traces, or recursive self-improvement. These mechanisms can deepen and refine exploration within the learned distribution, but they do not automatically create an efficient path toward radically different regions of the solution space.

The Full Solution Space and the Human Subspace

In the figure, I represent the full function or program space as FFF. This is the space of all possible procedures, algorithms, policies, proofs, designs, and strategies that could solve a given class of problems.

Within FFF, I identify a smaller region, HHH, which I call the human-generated subspace. It contains the types of solutions that humans have historically discovered, recorded, explained, and judged intelligible enough to preserve.

The distinction between FFF and HHH is not merely about scale. It concerns structure.

Human solutions are shaped by numerous constraints:

• limited working memory;
• limited lifespans and experimentation budgets;
• biological perception and motor systems;
• cultural conventions;
• educational traditions;
• institutional incentives;
• the need to communicate ideas to other humans;
• preference for explanations that appear intuitive;
• historical path dependence.

Consequently, the solutions in HHH are not an unbiased sample of all effective solutions. They are the output of a particular species, operating under particular cognitive and social constraints.

This means that many useful solutions may exist outside HHH. I denote this region as F∖HF \setminus HF∖H: the portion of the complete solution space that is not represented by conventional human-generated strategies.

Some of those solutions may be incomprehensible to humans. Others may be comprehensible only after they have been discovered. Some may look inelegant, alien, or even irrational from the perspective of human intuition while nevertheless performing better according to the actual objective.

What Imitation Learning Really Optimizes

On the left side of the figure, examples, programs, and reasoning traces are fed into a model. The model learns statistical regularities across these human-generated artifacts.

This can produce extremely capable systems. Such a model may synthesize ideas, correct errors, combine distant concepts, and generate outputs that no individual human has previously written. But its training signal still originates primarily from HHH.

The model is therefore rewarded for answering questions such as:

• What would a competent human write?
• What sequence of reasoning steps resembles a successful demonstration?
• What program is probable given the programs in the training distribution?
• What explanation would human evaluators recognize as plausible?

These are not identical to the question:

What policy or program maximizes success in the underlying environment?

The difference can be subtle. In familiar domains, human-likeness and effectiveness are strongly correlated. Human-generated mathematics, engineering, and science contain abundant structure. But when we approach the frontier — where humans do not already know the answer — the correlation can weaken.

A model trained to imitate successful reasoning may learn the surface architecture of discovery without acquiring an efficient mechanism for searching through genuinely novel hypotheses.

Why More Test-Time Compute May Not Be Enough

A common argument is that the limitations of imitation learning can be overcome through test-time computation. Even if the model begins inside HHH, perhaps it can reason for longer, generate many candidates, critique itself, use tools, conduct experiments, and progressively move beyond the human distribution.

This is possible. I do not claim that a model trained on human data can never leave HHH. Neural networks can interpolate, recombine, and sometimes extrapolate. Search procedures layered on top of a learned model can also generate surprising results.

My concern is probability and efficiency.

Test-time search is guided by a proposal distribution. The model does not explore all possible programs uniformly. It proposes continuations, hypotheses, and actions according to what its training has made probable. If the model assigns extremely low probability to a useful but non-human strategy, then enormous computational resources may be required to discover it.

Suppose the target solution lies in F∖HF \setminus HF∖H, but the model’s learned distribution places nearly all its probability mass near HHH. Increasing test-time compute may generate more samples, yet those samples can remain highly correlated. The system may repeatedly explore variations of familiar approaches rather than crossing into a qualitatively different region.

More computation is not equivalent to unbiased exploration.

If I sample one million times from a distribution that assigns a probability of 10−2010^{-20}10−20 to the relevant class of solutions, I should not expect the correct strategy to appear. A vast search budget can still be negligible relative to the effective size of the space.

The practical question is therefore not whether departure from HHH is logically possible. It is whether it is computationally probable within realistic budgets.

Recursive Self-Improvement Does Not Eliminate the Initial Bias Automatically

A stronger version of the argument invokes recursive self-improvement. A model could improve its own reasoning procedures, produce better training data, design stronger successors, and repeat the process. Perhaps this recursive loop would eventually escape the human-generated subspace.

Again, I regard this as possible but not guaranteed.

Recursive self-improvement is not magic. Each improvement step requires:

1. generating candidate modifications;
2. evaluating those modifications;
3. identifying genuine progress;
4. preserving useful changes;
5. avoiding catastrophic regressions;
6. maintaining alignment between proxy metrics and the actual objective.

If candidate improvements are proposed by a system whose search distribution is heavily shaped by HHH, then the recursion may continue to optimize within the same broad basin. The system might become an increasingly powerful executor of human-like cognitive strategies without discovering fundamentally different forms of computation.

Recursion amplifies the dynamics already present. It does not necessarily erase their biases.

A locally optimizing process can move very far while remaining inside the same attractor. Greater depth does not guarantee a change in direction.

The AlphaGo–AlphaZero Analogy

The left and right sides of the figure can be understood through the contrast between AlphaGo and AlphaZero, although the analogy should not be treated as exact.

AlphaGo incorporated human expert games. It learned from demonstrations before extending its abilities through self-play and search. Human knowledge gave it an advantageous initial distribution.

AlphaZero, by contrast, began without human game examples. Given the rules and a performance objective, it developed its strategy through self-play. It was not required to imitate the historical conventions of human chess or Go.

The important distinction is not that one system used search and the other did not. Both relied on powerful optimization and search mechanisms. The distinction is that AlphaZero’s learning signal was grounded more directly in the consequences of actions within the game.

It was rewarded for winning, not for playing in a recognizably human manner.

This allowed it to discover strategies that initially appeared unconventional. Once those strategies were demonstrated, humans could often interpret them retrospectively. But human interpretability was not a prerequisite for their discovery.

That is what I mean by the phrase at the bottom of the right-hand panel:

It learns what solves.

The left-hand system primarily learns how humans solve. The right-hand system searches for whatever actually satisfies the objective.

What “Tabula Rasa” Should Mean

The phrase tabula rasa can be misleading if interpreted literally. No learning system is completely blank.

Every system contains inductive biases. Its architecture, optimization algorithm, representation, action space, objective function, memory structure, and environment all constrain what it can learn.

Therefore, by tabula rasa, I do not mean a system without assumptions. I mean a system that minimizes dependence on human demonstrations of the solution itself.

A first-principles learner may still receive:

• the rules of the environment;
• a measurable objective;
• computational primitives;
• sensory inputs;
• an action interface;
• mechanisms for memory, prediction, and search.

What it does not receive is a detailed prescription of how humans believe the problem should be solved.

This is an important distinction. We cannot remove inductive bias, but we can choose whether the bias comes primarily from human behavioral imitation or from more general structural assumptions about computation, causality, exploration, and optimization.

Search, Evaluation, and Update

The right panel contains a recurrent loop: search, evaluate, update.

The system begins with an objective and a set of possible actions or programs. It generates candidates, tests them against the environment, measures their consequences, and updates its policy or search procedure.

This creates an epistemic advantage. The environment can reject strategies that appear plausible but do not work. It can also reward strategies that appear strange but succeed.

Human-generated supervision often evaluates outputs through human expectations. First-principles search evaluates them through external consequences.

That distinction becomes crucial in domains where human intuition is unreliable or where the relevant phenomena are too complex for direct human judgment. Examples may include protein design, materials discovery, algorithm synthesis, chip layout, theorem discovery, control systems, and scientific experimentation.

In these domains, the most effective solution might not resemble anything in the historical record. A system should therefore be allowed to treat human knowledge as a prior rather than as the boundary of legitimate thought.

The Strongest Objection to My Argument

The strongest objection is that HHH is not a closed subspace.

Human-generated data includes mathematics, programming languages, experimental methods, search algorithms, and descriptions of creativity itself. A sufficiently capable model may learn general procedures that permit it to construct solutions far beyond its demonstrations.

This objection is valid.

A model trained on human artifacts is not limited to memorizing those artifacts. It can learn compositional rules. It can infer latent structures. It can use human-produced algorithms as launchpads for non-human discoveries. In that sense, training on HHH can provide tools that reach beyond HHH.

Therefore, the crude statement “training on human data permanently traps a model inside HHH” is too strong.

The more defensible claim is distributional:

Training predominantly on human-generated solutions creates a strong prior toward regions of solution space that are reachable through human-like representations, decompositions, and search trajectories.

That prior may be beneficial initially. It can greatly reduce sample complexity. But it may become restrictive near the frontier unless the system possesses sufficiently strong mechanisms for environmental interaction, experimentation, objective-grounded evaluation, and diversity-preserving search.

The real issue is not whether escape is mathematically possible. It is the expected computational cost of escape.

We Do Not Have Infinite Compute

In theory, an infinitely powerful search process could enumerate every possible program, test every candidate, and eventually discover every solvable strategy.

But intelligence operates under constraints.

We have finite energy, finite hardware, finite data, finite experiment budgets, and finite time. More importantly, many environments are partially observable and nonstationary. Actions can have irreversible consequences. Evaluation may be noisy or delayed.

Life is closer to a partially observable Markov decision process than to a clean optimization benchmark. An agent does not know the complete state of the world. It receives incomplete observations, maintains uncertain beliefs, and chooses actions whose effects may only become visible much later.

Although the world may not possess a literal finite endpoint, any real agent operates with a finite effective horizon. It has only a limited number of decisions before resources are exhausted, opportunities disappear, or the environment changes.

Under these conditions, the quality of the search distribution matters enormously.

We cannot rely on infinite test-time computation to compensate for a poor prior. We need systems whose learning dynamics direct computation toward productive novelty.

Human Knowledge as Scaffolding, Not a Cage

My argument should not be interpreted as a demand to discard human data. That would be wasteful.

Human knowledge should serve as scaffolding. It can provide language, abstractions, tools, heuristics, causal hypotheses, and accumulated empirical findings. The danger arises when the system is trained to treat human-generated trajectories as the ultimate target rather than as provisional evidence.

The more promising architecture may be hybrid:

1. use human knowledge to establish a strong initial model;
2. ground learning in external objectives and empirical feedback;
3. permit autonomous experimentation;
4. reward valid discoveries even when their reasoning is unfamiliar;
5. maintain multiple competing hypotheses rather than collapsing prematurely;
6. use search procedures capable of crossing low-probability regions;
7. continuously revise the model when reality contradicts inherited assumptions.

In this architecture, imitation accelerates early learning, while first-principles search prevents inherited human conventions from becoming permanent boundaries.

Toward More Intelligent Systems

The deepest issue is what we want an intelligent system to optimize.

Do we want it to reproduce the cognitive paths humans have already taken? Or do we want it to discover effective solutions under the actual structure of reality?

These objectives overlap, but they are not identical.

Human reasoning traces reveal valuable strategies, yet they also contain accidents of history, pedagogical simplifications, cognitive limitations, and social conventions. A system that learns them perfectly may become an exceptional simulator of human intelligence without becoming a fundamentally different form of intelligence.

To reach substantially more capable systems — and perhaps eventually artificial general intelligence — we may need to shift the center of gravity from imitation toward autonomous, objective-grounded discovery.

The decisive transition is from asking:

“How would a human solve this?”

to asking:

“What solves this, under the real constraints of the environment?”

That is the meaning of the figure. The left side inherits a map drawn by humans and explores it with increasing sophistication. The right side begins with an objective, confronts the terrain directly, and constructs its own map.

The first approach learns our paths.

The second may discover paths that no human would have thought to search.