True Reasoning vs Statistical Pattern Matching
For me, one of the clear distinctions between true reasoning and pattern matching, is what happens when you remove access to new information. Many argue that LLMs are not plagiarism machines, they learn like humans do. They consume knowledge from teachers and books, developing an understanding along the way.
As a professional researcher, I too learned most of what I know by reading the works of far greater researchers. But there became a point where I knew enough to perform my own original research in uncharted waters. I now can and regularly do research topics where there is no other source to check my work against. This is not something LLMs are good at, or debatably, can even do at all.
Giving an LLM access to the entire internet and all of recorded human knowledge, then testing them with a quiz designed for humans is obviously just a cheap parlor trick. I, too, could score 100% on a multiple-choice exam if you let me Google all the answers. The true measure of LLM intelligence and reasoning should not be refactoring existing information, but the ability to produce truly novel works.
Sure, an LLM could probably create a brand-new pop song, because it has plenty of existing songs to analyze, allowing it to produce something seemingly new, but really just based on existing patterns. Yet, every time I tried to get LLMs to perform novel research, they fail because they don’t have access to existing literature on the topic. Whereas, humans, on the other hand, discovered everything humanity knows.
Where people get tied up is with the argument, “well most humans don’t produce novel work either”. But this is not because they’re fundamentally incapable of it. The average person is simply just sandbagged with an unfulfilling job. Ideally, without the perverse incentives of shareholder value, LLMs would automate all the busy work, allowing humans to focus on more meaningful pursuits.
I’ve made this argument many times before. But if humans could come up with all the groundbreaking discoveries they have, reading only as many books or research papers as they realistically could. Where are all the major LLM discoveries? An individual human may be limited in their ability to make novel discoveries as a result of competing with 8 billion other humans. But LLMs have access to the knowledge of all.
You’d think that a machine with access to every book ever written, every paper ever published, every speech ever recorded, and every study ever produced could do a lot better than some person who can read maybe 1 book per day. Yet, nothing. There’s the odd “maybe the LLM did something novel, but we don’t know yet” posts here and there, but if they could actually think, with as much information as they have, there’d be groundbreaking discoveries literally falling from the sky.
In reality, all we’ve created is a bot which is almost perfect at mimicking human-like natural language use, and the rest is people just projecting other human qualities on to it. Quite simply, “LLMs are doing reasoning” is the “look, my dog is smiling” of technology. In exactly the same way that dogs don’t convey their emotions via human-like facial expressions, there’s no reason to believe that even if computer could think, it’d perfectly mirror what looks like human reasoning.