mxx1's avatar

AI hypothesis

If you look closely at how large language models work you see an uncomfortable reflection of how most humans work.

LLMs are trained on huge quantities of written words, compressed into patterns. They use this data to guess the next word. I suspect that is not far from what most human brains do.

Billions of neurons form a web of associations, strengthened by repetition and context, until the act of thinking is mostly prediction. The familiarity of an LLM’s mistakes, the hallucinations, the confident nonsense, all supports this assertion.

Way back when, language turned memory into an external system, a way to stabilise and share mental models. Culture was built out of words and stories, but it did not make humans more rational; it simply socialised the biases.

Rational thinking and the scientific method was a later addition; a set of procedures designed to distrust intuition; state a hypothesis, test it, measure outcomes, critique and replicate results.

It broke the closed loop of pattern recognition; it is a tool we invented to work around a brain optimised for survival, not truth. In fact it’s invention marked the transition from survival to exponential growth on the exploitation of the environment.

LLMs are a mirror of the lower cognitive stack: a predictive system with vast memory but no metacognitive layer.

They mimic intelligence convincingly because much of what we call intelligence is just pattern prediction at scale. But they do not critique themselves or perform experiments, they do not doubt themselves, nor do they measure or care about success.

Without a scientific layer, they are locked in the same intuitive world humans occupied before the Enlightenment. But they are much faster.

So it’s the next step in AI development that will be the most revolutionary. When we encode rational thinking and the scientific method into the machines.

This will not come from bigger models but from programming machines to work like humans do when they’re doing rational thinking or science, layering a process of hypothesis, testing, critique and correction on top of raw prediction and heuristics.

The scientific method was humanity’s hack for escaping intuition, a framework that let us test truth rather than assume it. Embedding that same discipline into AI would create systems that do not just sound right but can be innovative.

The path to real machine reasoning could start small, with narrow domains and simple protocols rather than grand ambitions. A minimal system could focus on a tightly defined field like logic puzzles, geometry puzzles or formal proofs, using just a handful of agents: one to propose arguments, one to critique them, and one to verify each step against explicit rules.

Once it works, then we scale it.

In the meantime, the more useful LLMs are to you,  the more you’re living in the pre-Enlightenment mental space, like most people.