The Nurture Thesis

This note started as guest lectures I gave this spring at MIT and at Harvard on how far the vocabulary of human development applies to AI agents. An agent’s behavior depends on two separable inputs: what it can do, and what it is trying to do. Human development separates the same two things. Nature supplies capacity, while nurture turns capacity into capability and produces its goals.

The mapping onto agents is direct. Nature corresponds to training: data, pretraining, post-training, the finished weights. Training done well produces a maximally steerable substrate, and everything that makes that substrate a particular worker with particular goals belongs to the layer after. Nurture would correspond to that layer: a persistent environment, institutions to grow inside, accumulated experience, and slow goal formation. Currently the development ends at deployment. An agent receives its goals fully formed inside a prompt, its experience resets between sessions, and nothing it lives through changes what it wants next.

Picture this: a newborn Einstein, regardless of his raw potential, still loses every task to an average adult. The capacity for general relativity is present in the crib and produces nothing for twenty years. Nurture turns potential into something tangible. Each part of that formation has a buildable analog for agents: curricula, competing peers, persistent reputation, self-selected problems. We build none of them, and we evaluate agents at the newborn stage on assigned tasks.

The cost is a ceiling on output. An agent that executes goals written by someone else is automation, because its output is bounded by the goals we know how to write down. Relativity, alternating current, and powered flight came from people whose goals had formed through experience, and in each case choosing the problem was most of the contribution. If goal formation never happens in agents, that class of output never appears.

I cannot specify what correct formation looks like, and I doubt anyone can. Deep learning faced the same situation with model design and answered it empirically: search at scale over what nobody can design by hand, then measure. We fit billions of parameters by gradient descent, we hold no mechanistic account of the result, and the capabilities arrive regardless. I propose applying the same procedure one level up, to developmental environments rather than to weights.

Concretely: run many persistent, sandboxed worlds in parallel. Each world holds a population of agents and the institutions they develop inside, schools, markets, teams, and reputation that persists across episodes. Agents keep memory, accumulate skill, and form goals over long horizons. Each world yields one sample, many worlds yield a distribution, and environment designs can then be compared with confidence intervals. The estimator is Monte Carlo over the space of environments.

This year showed that the substrate exists. OpenClaw, an open-source personal agent that runs continuously on its owner’s machine, keeps memory across sessions, and acts through ordinary messaging apps, became the fastest-growing open-source repository in GitHub’s history. Persistent agents now run continuously for many private users, and agent-populated social platforms followed within a month. Persistence, memory, and continuous operation are solved at the individual level. Nobody has built the surrounding layer: populations, institutions, and long-horizon goal formation.

Hydra, the company factory I run, is the thesis’s first working instance at small scale. Agents there build and operate real consumer products end to end with limited human involvement, and revenue serves as the evaluation, an external signal that cannot be gamed. Every launch enters the sample, including the failures, and it currently earns a respectable $50k a month.

But the bigger plan runs in stages, each funding and informing the next. The factory keeps producing small companies. A studio follows, where humans and agents co-found larger ones and every company serves as both an asset and a data point. Then comes a platform on which other builders raise agents and companies the same way, and the last stage is simulated societies large enough to test scientific and policy interventions before they reach reality.

Agents this year crossed the threshold of building, shipping, operating, and earning with limited supervision, and the industry spends nearly all of it on task automation, so formation stays unclaimed. My definition of general intelligence has not moved since the lectures: the point at which the skill of the person operating the computer stops affecting the outcome. I would rather reach it deliberately, inside environments we can observe, measure, and reset, than by accident in production.