LLMs are cousins of our brains

When we make decisions, we like to believe we’re engaging in something called “reasoning”, that we believe is a special human capability unlike logic that can be programmed. We use reasoning to decide what to have for breakfast, whether to take a job, and even to extrapolate to existential matters such as the existence or otherwise of God. C.S. Lewis famously attributed his rejection of atheism to his attempts to prove the null hypothesis of naturalism, and further considered “reasoning” not explainable without God. In “Miracles“ he claims “acts of reasoning are not interlocked with the total interlocking system of Nature as all its other items are interlocked with one another.” But instead, “something beyond nature operates whenever we reason.”

It is probably this “prior” that leads to the current somewhat contorted attempts to argue that while humans reason, LLMs are just looking like they reason, but not doing it really, because they in some ways behave differently to humans. What if we reject this prior assumption? Cognitive science research suggest that when humans think they are reasoning, they are actually behaving probabilistically, even if there is some kind of platonic ideal of logic that the best reasoners converge upon.

Even if we accept that there is something special about logic and reasoning distinct from probabilistic prediction, AI companies are rushing to integrate logic-based reasoning into LLMs, so we can expect the next generation to be more “logical”. But we should be aware that by doing this, we may be adding a constraint to LLMs (compliance with something we have invented called “logic”) as well as increasing the power of LLMs to behave like we do.

As humans and machines develop together they both face what we might call the “scope problem” - the limitation of operating only within the boundaries of their respective training data. Human knowledge emerges from a complex evolutionary process of cultural accumulation. Even pure logic is only as good as its categorical assumptions, which is only as good as the data the categories are built on. What we consider “facts” are often provisional understandings that have proven useful for navigating our world. For humans, this is the sum total of experiences and knowledge accumulated over a lifetime. For LLMs, it’s the corpus of human-generated text they’ve been trained on. In my prior posts, I observed that LLMs are on track to significantly exceed human parity (LLM limits and human parity) at least in the tests we as humans give them, and as they do so, we and they will understand more the limits of humans particularly in our capacity to effectively use massive amounts of information (Human limits and LLM parity). If both human brains and LLMs are mostly doing their thing by predicting based on historical and available data, and we incorporate “reasoning” into them, then LLMs - or their children - are clearly going to be the winners. When humans venture beyond their scope of training, the results can be just as problematic as when LLMs generate hallucinations.

If we stop treating “real” reasoning as a human birth-right and see our brains and LLMs as probabilistic cousins, a huge opportunity opens up: we can learn to talk to one another far better than we do today. The next few years should be spent moulding, framing and sometimes constraining these models so their internal predictions line up with ours—through feature discovery that surfaces what the network is really tracking, graph and structured data that anchor generations in relations we can inspect, through hybrid pipelines that fuse deterministic solvers with learned ones, through more clearly defined interfaces, and yes through the new “reasoning” layers that force the model to show its work. The prize is not a winner-takes-all contest between carbon and silicon but a joint system in which each side covers the other’s blind spots. Teaching our silicon cousins to think more like us is, in the end, a way of teaching ourselves to think a little more clearly.