In December 2017, DeepMind published a paper that quietly upended everything we thought we knew about how machine intelligence is acquired. AlphaZero, trained from nothing but the rules of chess, shogi, and Go, defeated the world’s strongest specialized programs within hours. It did not just win—it played in ways that grandmasters described as alien, pursuing strategic concepts that centuries of human play had never surfaced. The natural assumption was that this was a story about scale: more compute, more data, better hardware. But that reading misses the deeper lesson entirely.
AlphaZero’s revolution was not about how much it computed. It was about how it learned. By playing millions of games against itself—each game generating a signal about which moves led to victory—it engaged in something that cognitive scientists would recognize immediately: a form of deliberate practice. DeepMind did not use that language. But the structural parallel is hard to ignore, and we believe it holds profound implications for the next generation of AI agents.
The Deliberate Practice Framework
The science of expertise has a long and contested history, but one of its most durable contributions came from psychologist K. Anders Ericsson and his collaborators in the early 1990s. Ericsson spent decades studying elite performers—violinists, chess players, surgeons, typists—and arrived at a conclusion that flew in the face of popular intuition: raw talent and accumulated experience are poor predictors of expert performance. What distinguishes the truly exceptional is the structure of how they practice.
Ericsson called this structure deliberate practice, and he identified four essential components. The first is specific, well-defined goals: not “get better at chess” but “improve accuracy in rook endgames from equal positions.” Vague intentions produce vague results. The second is focused attention: deliberate practice demands full cognitive engagement, not the semi-automatic repetition that characterizes most “practice.” The third is immediate, informative feedback: the learner must know not merely whether they succeeded, but precisely where their performance deviated from the ideal and why. The fourth is repetition at the edge of ability: tasks must be calibrated just beyond current competence, in Vygotsky’s zone of proximal development—hard enough to demand growth, but not so hard as to produce only failure.
The famous “ten thousand hours” rule is profoundly misleading. Ten thousand hours of mindless repetition produces an experienced amateur. Ten thousand hours of deliberate practice produces a master. The difference is the quality of the practice, not its duration.
Adapted from K. Anders Ericsson, “Peak: Secrets from the New Science of Expertise”
This framework matters for AI not because we want to make machines practice scales on a digital piano. It matters because it describes, in precise mechanistic terms, the conditions under which any learning system—biological or artificial—can develop genuine expertise rather than merely accumulating experience.
Mapping Self-Play to Deliberate Practice
Return now to AlphaZero, and examine its training loop through Ericsson’s lens.
The specific goals component is handled by the game itself: win. But within that overarching objective, the Monte Carlo Tree Search that guides play constantly evaluates positions, generating micro-goals at every decision point. The system is not pursuing a vague intention; it is making precise evaluations at each state of the board.
The focused attention component maps to the depth of search. AlphaZero does not play casually—it allocates substantial computational resources to each move, exploring consequences deeply rather than relying on surface heuristics.
The immediate feedback component is the game result itself, propagated back through the value network. Every position in every game receives a retrospective evaluation: did this board state lie on the path to victory or defeat? This signal is precise, instantaneous (in training time), and directly tied to the quality of the decisions made.
Most critically, practice at the edge of ability is structurally guaranteed by self-play. AlphaZero’s opponent is always itself—and as it improves, its opponent improves at exactly the same rate. The system can never outgrow its sparring partner, because its sparring partner is its own current best self. This is the most elegant feature of the architecture, and the one that most directly mirrors what Ericsson observed in human experts.
Now consider other paradigms in modern AI training through the same lens.
Reinforcement Learning from Human Feedback (RLHF), the technique that enabled conversational AI systems like ChatGPT, can be read as introducing a human coach into the deliberate practice loop. Human raters provide the immediate, informative feedback that the model cannot generate for itself—judgments about which responses are better, and implicitly why. The limitation is that human feedback is expensive, slow, and inconsistent. A coach who can only observe your practice once a week and provide vague impressions is not the same as an endlessly patient mentor who evaluates every repetition in real time.
Constitutional AI, Anthropic’s approach to alignment through self-critique, moves the feedback mechanism partially inside the model itself. The system evaluates its own outputs against a set of principles, generating a form of self-assessment that the deliberate practice framework would recognize as self-evaluation—a capability that Ericsson identified as one of the defining traits of true experts, who internalize standards deeply enough to critique their own work without an external referee.
What is missing from all of these approaches, separately, is a unified framework that combines them: persistent specific goals, deep focused attention on weaknesses, high-quality immediate feedback, and continuous practice calibrated to the system’s current capability frontier. Each technique captures one or two components of deliberate practice. None yet captures all four in an integrated architecture.
The Etude AI Approach
This is precisely the gap our work is designed to close. Rather than one-shot training followed by static deployment, we are building systems that engage in continuous improvement through structured practice sessions—sessions explicitly designed around the four components of deliberate practice.
The musical analogy embedded in our company’s name is not incidental. An étude, in classical music, is a study piece specifically composed to develop a particular technical skill. Chopin’s études are not concert pieces that happen to be difficult; they are surgical instruments for developing specific aspects of piano technique—finger independence, octave speed, chromatic scales in thirds. A pianist preparing for performance does not simply play through their repertoire repeatedly. They identify weaknesses, seek out or compose targeted exercises, and practice those exercises with full attention until the weakness becomes a strength. Then they move to the next weakness.
Our agents follow an analogous process. Rather than simply executing tasks and discarding the experience, they engage in a four-stage cycle. First, they identify weaknesses: through structured self-evaluation after each task, they build a model of their own competency landscape—which skills are reliable, which are fragile, and which are absent. Second, they generate targeted exercises: synthetic practice scenarios calibrated to the specific weakness identified, pitched at a difficulty level just beyond current demonstrated competence. Third, they practice deliberately: engaging with these scenarios not as real tasks but as learning opportunities, with full attention on the specific skill being developed. Fourth, they measure progress: tracking improvement in the targeted skill area and updating their competency model accordingly.
The goal is an agent that knows what it does not know—and knows how to learn it. Not through passive accumulation of experience, but through active, structured engagement with its own performance frontier.
This architecture requires three technical components that do not exist in standard agent deployments: a competency model that tracks skill-level performance across a structured taxonomy, a practice memory that stores corrective strategies derived from self-evaluation, and a practice scheduler that generates targeted exercises and manages the spaced repetition of skills over time. We have described these in more detail in our earlier post on building agents that learn from practice.
Why Scale Alone Cannot Get Us There
The dominant paradigm in AI development over the past decade has been scaling: more parameters, more data, more compute. This paradigm has produced remarkable results, but there are clear signals that it is approaching diminishing returns for many capability domains. Benchmark performance on established tests continues to improve, but the rate of improvement is slowing, and the improvements are increasingly concentrated in narrow performance dimensions that do not reflect the full breadth of competence needed for real-world deployment.
More fundamentally, scaling is orthogonal to the deliberate practice framework. A larger model trained on more data is not a model that has practiced more deliberately—it is a model that has been exposed to more examples of finished human thought. This is more like reading thousands of books than like spending thousands of hours at the instrument. Reading is valuable, but it does not produce the same kind of deep, situated competence that practice produces.
Consider what AlphaZero demonstrated about the relationship between practice quality and model size. It was not the largest model ever trained on games; it was not exposed to the most game records in history. What it had was a practice loop of extraordinary quality: immediate, precise feedback, continuous calibration of difficulty, and unlimited repetition. A smaller model with better practice can outperform a larger model with worse practice. This is not a speculative claim—AlphaZero made it empirically, against the strongest engines in the world.
We believe the same principle applies to language-based agents. A model that can identify its weaknesses, generate targeted practice, and systematically improve its competency profile will, over time, develop capabilities that cannot be matched by simple scale. The key is building the machinery for practice, not just the machinery for inference.
Implications for AI Safety
Practice-based learning raises important questions for safety and alignment. An agent that modifies its own behavior through practice is an agent whose behavior can diverge from its initial specification in ways that may be difficult to predict or detect. This is a legitimate concern, and one we take seriously.
But the deliberate practice framework actually provides structural safety properties that one-shot training lacks. In our architecture, the learning that occurs during practice is stored not in opaque weight changes but in an explicit, inspectable practice memory—a structured knowledge base written in natural language. Every corrective strategy the agent develops, every new heuristic it acquires, every weakness it identifies and addresses, is recorded in a form that human operators can read, review, and override.
This is fundamentally different from fine-tuning, where behavioral changes are distributed across billions of parameters in ways that resist interpretation. It is also different from in-context learning, where adaptations exist only within a single session and cannot be audited after the fact. The practice memory provides a persistent, legible record of how the agent has changed and why—a property that becomes increasingly important as agents take on more consequential tasks.
There is a deeper alignment argument as well. Agents that can identify their own weaknesses and practice to address them are agents with a built-in mechanism for self-correction. If an agent recognizes that it has been making systematic errors in a particular domain—misinterpreting certain types of user intent, applying incorrect heuristics in edge cases—it has both the motivation and the machinery to correct those errors. This is a form of self-alignment: the agent is actively working to improve its own reliability, not merely executing a fixed policy.
We do not claim this solves alignment. An agent that practices effectively toward the wrong objectives is more dangerous, not less. But combined with appropriate oversight mechanisms and carefully constrained practice domains, self-correcting agents are, we believe, inherently safer than static agents that repeat their errors indefinitely without any mechanism for improvement.
The Road from AlphaZero
AlphaZero was trained in a domain with three properties that made deliberate practice tractable: a perfectly defined state space, a clear and unambiguous reward signal, and the ability to simulate experience at arbitrary speed. Language-based agents face none of these conveniences. The state space of human language and real-world tasks is effectively infinite; reward signals are ambiguous, contested, and expensive to obtain; and experience cannot be simulated at the speed of silicon because it involves interaction with the real world.
These are hard problems, and we do not pretend to have solved them. Our current work addresses simplified versions: structured task domains with clearer competency taxonomies, self-evaluation mechanisms that approximate the precise feedback of game outcomes, and synthetic practice scenarios that approximate the edge-of-ability calibration of self-play. Each approximation involves tradeoffs, and each represents a direction for ongoing research.
But the arc of the trajectory is clear. The path from AlphaZero to truly adaptive agents—agents that develop genuine expertise through structured engagement with their own performance—runs through deliberate practice. The cognitive science framework that Ericsson spent decades developing in the context of human expertise turns out to describe, with remarkable precision, the conditions under which any learning system can transcend its initial training.
We are at the beginning of this path. The initial results are encouraging, and the theoretical foundations are solid. But the distance between where we are and where we want to be is large, and the work required is not primarily about scaling. It is about building better practice.
If this line of research resonates with you—whether you come from cognitive science, machine learning, or practical agent engineering—we would like to hear from you. The best ideas in this space have always emerged from exactly these kinds of disciplinary intersections.