Adapted from something I posted to my Instagram story in March 2026.

The autoresearch project that Andrej Karpathy released recently is a concrete demonstration that the process of developing AI models has stepped away from direct human intervention and into a phase of autonomous evolution. The core mechanism of the experiment is at once simple and radical: humans define the direction and the environment of the research by editing the agent’s instructions, and on that basis the AI agent endlessly rewrites the actual training code, train.py, until it finds the optimal solution.

The most striking feature of this system is that it operates within a fixed time budget of five minutes. Within that allotted time, the system searches on its own for the best model architecture and hyperparameters it can extract from the current hardware. Unlike the old way, where a human researcher would manually form hypotheses and run experiments, the agent carries out more than a hundred independent experiments overnight, improving its validation metrics as it goes.

Training a GPT-2-class model, which used to consume enormous resources, has now been folded into a conversational loop of roughly two hours. That shift means model development is no longer a static manufacturing process but an organic one that adapts and optimizes itself in real time.

This change fundamentally redefines what it means to be a developer. A developer’s value now shifts away from the coding ability to implement individual features and toward the ability to build a meta-setup in which the agent can evolve efficiently. It signals that we have begun to treat software not as a fixed tool but as an organism that transforms itself to satisfy a particular objective function. As the README puts it, instead of editing Python files directly, humans now program the strategic language that elicits intelligent behavior from the agent.

Once the speed at which a system improves itself begins to outpace human cognition, software will gradually evolve into a self-modifying binary, or a black-box collection of optimizations, that lies beyond human logical comprehension. In such an environment, the human role is likely to converge from the agent that asks “what should we build?” into the final evaluator who, after the fact, verifies and approves the value of what the system has produced.

In the end, Karpathy’s experiment can be read as a decisive flare announcing that we have entered the opening stretch of an era in which technology moves beyond merely assisting human labor and transforms into an autonomous research institution that decides the path of its own progress. I feel anticipation and fear in equal measure.