AI and Creativity

Evolution is the creation of knowledge through variation and selection. And yet, it is amazing that the DNA code has enough reach to describe everything from simple cells to dinosaurs and humans. Somenthing we still cannot fully explain.

Nature holds many unresolved mysteries, among them the phenomenon of creativity, the nature of qualia (feelings), and the “hard problem” of consciousness. I am not sure that trying to achieve these artificially, without ever discovering those unknowns, can work. As long as we cannot explain the common language of art and music, why certain flowers look beautiful, or why we feel happy or sad, how can we expect to simulate any of it in a computer program?

As David Deutsch said, we, humans, are universal explainers, capable of creating explanations. New explanations create new problems, inevitable but soluble problems.

AI is a computational process; it uses pre-existing knowledge to predict the next token. The question here is whether AI will be able to help us in this infinite expansion of knowledge. I believe the answer to that question is the same as the answer to whether AI can be creative and conscious.

Deutsch argues that digital computers are universal too, similar to human brains. But the problem lies in the program that runs on them. Deutsch rejects induction (learning by extrapolating patterns from data) as a valid theory of knowledge at all. He holds that real creativity means generating new explanations through bold conjecture and error-correction, ideas that go beyond, and sometimes against, the data, rather than ones derived from it. Hardware is not the source of creativity. Even if we were able to recreate the brain, it wouldn’t be creative on its own. What matters is the program running on it, the conjecture and criticism that generate new knowledge.

Current AI is built on inductive principles: optimizing against training data or objective functions. So, it can recombine and extrapolate existing patterns, but it cannot originate genuinely new explanatory knowledge or question its own goals (so far).

On the other hand, Penrose says no; a computational process, by definition, cannot be creative. He draws on Gödel’s theorem: in any consistent formal system, there are specific true statements that cannot be proven within that system. In other words, any sufficiently powerful, consistent formal system will contain true statements that cannot be proven using its own rules. Because computation is bound by these formal systems, Penrose argues it cannot replicate the human mind’s ability to recognize these unprovable truths. This suggests consciousness accesses something beyond computation.

Penrose proposes that consciousness arises from quantum collapse in microtubules. When a quantum superposition reaches a threshold of mass-energy, the wave function collapses spontaneously. Penrose ties consciousness to these collapses (which are not yet proven).

He imagines three worlds: math, physics, and mind, linked in a loop. Math comes first: it exists on its own, timeless and independent of us. Only a tiny piece of math describes the physical world. From the physical world, only a small part gives rise to conscious minds. And from the mind, a small part of our thinking, our ability to understand math, somehow reaches back and grasps the entire mathematical world we started with.

In March 2016, AlphaGo played Move 37 against Lee Sedol, an unconventional move that had only a one-in-ten-thousand probability of being played by a human. Still, I believe that when AlphaGo searches a game tree, it is a physical system instantiating abstract rules. But instantiating a rule is not the same as understanding why the rule works. A computer can execute an algorithm perfectly without grasping the explanation behind it. Humans, by contrast, build good explanations, theories that let us predict, explain, and create in domains we have never seen before.

If Deutsch is right, AGI is a software problem; that is, we need an algorithm for conjecture or a different architecture (see Yann LeCun).

If Penrose is right, AGI is a hardware/physics problem (we need quantum biological computing).

And if Altman is right, true understanding is an emergent property of scaling laws, where a future iteration like GPT-8 might be able to solve quantum gravity through data alone.

The day AI becomes conscious and creative, able to explain new knowledge on its own, will be the day we reach AGI.

But even if that day never comes, the potential to apply AI to improve our daily lives remains huge, a once-in-a-lifetime opportunity. And as Popper reminds us, it is our duty to remain optimists.