The current Silicon Valley promise is if we feed a machine enough of your data, it'll become you.

Science fiction has warned us this is a lie. In the Black Mirror episode "Be Right Back," a clone is built from a dead man's social media. It sounds like him, jokes like him, but it feels hollow.

But how can you tell the nuances of human behavior? If human agent is simply a complex system that takes actions based on sensory input, can't AI agent approximate the same computational function?

The current architecture of LLMs across major labs follows the same paradigm: Pre-training ingests the vast corpus of the internet, effectively learning the "average" of human behavior. Post-training and RL steers them towards desired & expert behavior. However, we have not yet steered into individualism, diverse sometimes outlier behaviors that are unique to humans. When we converge on the most helpful answer, we are stripping away idiosyncrasies, biases, and "inefficient" choices that define a specific person.

Through the lens of Richard Sutton, a person in the language of AI can be described through policies and values. Your policy is your behavior and the specific, often irrational actions you take in the moment. But underneath that policy lies your value function, which is human taste. Take Steve Jobs. In an era where the policy was optimized for "computing power per dollar," Jobs was operating on a value function that heavily weighted aesthetics.

We did not get Steve Jobs by averaging the men in suits. In fact, the hippie movement and counterculture had profound impact. If you train AI on every painting before 1907, you may get a better landscape or portrait, but it'll not create Cubism. If you fed it all the physics before 1905, it would give you a better Newton but not Einstein and relativity.

These were novel discoveries, shaped by uniqie human experiences.

In 1907, Picasso visited the Trocadéro Ethnographic Museum in Paris. He encountered African masks and sculptures that were not attempting to mimic reality (realism) but were attempting to capture a spirit. This experience directly led to Les Demoiselles d'Avignon. The faces of the women on the right side of the painting are not human faces; they are jagged, mask-like reconstructions.

After dropping out of Reed College, Steve Jobs didn't leave campus. He audited classes that interested him, specifically a course on calligraphy taught by a Trappist monk named Robert Palladino. Ten years later, when designing the first Macintosh, this experience resurfaced. He insisted that the Mac shouldn't just display green command lines; it should have beautiful typography.

Why can't AI create open-ended discoveries like these using prior art? Each of these human geniuses in our history seem to have accumulated decades of experience to shape their view of the world that are different from others at the time. Innovation requires divergence. Divergence requires bias. Only when an AI is allowed to drift away from the average can it create true novelty.