Humans Are Crooked Animals

I’ve always felt that reaching AGI and connecting models to everything in the world — but here’s the thing: humans are crooked animals.

I’ve always felt that reaching AGI means connecting models to everything in the world. But that definitely doesn’t mean simulating human behavioral patterns and adding constraints. If we treat models as living organisms, we’d find they have almost no constraints or limits. Adding those constraints is a kind of dimensionality reduction.

And the target of that reduction is inherently fuzzy. Which human’s direction do you reduce toward? There’s no ideal human. Average out all human feedback and you don’t get a superhuman — you get a mediocre human. Something that won’t make catastrophic mistakes but will never have any genuinely sharp, breakthrough ideas.

Let me be clearer: humans are crooked animals. All our great creations are byproducts of bugs. An animal’s original purpose is to survive and reproduce, but your intelligence exceeds all of that. The surplus gets used to process problems it was never meant to handle — philosophy, art, science — and those things reshape the system in turn. Do I need to name some of my favorite “superhumans” here? Musk, Feynman, Planck, von Neumann. These are the superhumans I admire. Regardless of whether their success came from stacked bugs or luck, model training can only extract their outputs. It can’t extract the distorted structure that produced those outputs. Even if survivorship bias is real, training can’t even replicate the bias itself.

Note: Ten thousand people with no sense of risk — nine thousand nine hundred and ninety-nine go broke, one becomes Musk. That analysis isn’t useful; it can’t explain why him and not someone else. It can only construct a success story after the fact for whoever already succeeded. What I’m arguing is why training — and all human-like embodied intelligence — cannot replicate these things. So whether you invoke hindsight or not doesn’t affect this argument.

Everyone right now wants to use training to make models reproduce this. The problem is that training is optimization, optimization is eliminating deviation, and eliminating deviation means eliminating bugs. So everyone is using a process specifically designed to eliminate bugs to reproduce something made entirely of bugs.

Let me go through everything: scaling is optimization at larger scale, agentic is fine-grained optimization, world models are optimized compression of the physical world, neuro-symbolic is stitching two kinds of optimization together, and embodied intelligence just moves the optimization somewhere else.

If intelligence is truly overflow, what you need isn’t better training — it’s more redundancy, more noise, more moments where the system doesn’t know what to do with itself. But no company dares do that, because an out-of-control model can’t ship.

Maybe AGI was never supposed to look human. What it should look like, no one knows — but the act of pulling it toward humanity might itself be the mistake. It might need to be something we don’t yet have words for.

But there’s a very practical problem: humans don’t trust what they can’t understand. Even if a model could operate in a genuinely non-human way, people would wrap it in a human shell anyway — otherwise they wouldn’t dare use it.

So this dimensionality reduction might not be a technical choice. It might be an unconscious political choice made by all of humanity.