Indeed, but isn't that just statistics? It detects more than us because it can go through more than us and be tireless about it. But does it understand things any more than just knowing the likelihood of C following B?
You’re closing yourself the path to answering the question by describing these things as binary options that exclude each other. They do not. There is even little reason to think that humans are anything but sophisticated statistical machinery.
The important distinction is that there are different kinds of statistical machinery. They’re not better or worse, just different. For example early LLMs already exceeded human capability in quick arithmetics (when working with numbers directly). Statistics were created to overcome limitations of human brain which can’t directly pick up relevant information from sets of data. It’s a matter of fitness for the purpose, not any generic, absolute value.
Neither LLMs nor any other statistical solutions created so far do understand in the human sense. All the non-textual input humans have, all the evolution-imposed drives, all life experiences, the serf-perception, all the social interactions — these are inherent components of human-style understanding. Whichever brain operation model we accept and how little we still know about its functions, we already know that thinking is based on a large number of interconnected and circular processes. All this is completely missing from what algorithms do now. The closest we got is agentic LLMs talking to each other and destroying things. So something like Greek philosophers, except each of them is a babbling kid with superhuman strength they can’t control.
But as with all anthropomorphisation used by AI researchers, the word has a more fuzzy meaning. LLMs can understand in more abstract sense of the word. They are stochastic parrots. But while fully accurate, this phrase is a bit snide reply to attributing AI models human-like features. It should not be taken too far beyond that, out of context. They are glorified Markov generators, yes, but no longer just Markov generators. With data so enormous, the statistical models hold not only linguistic rules. The purely textual matter at such a large scale contains patterns that represent knowledge or relationships between concepts. LLMs pick these up too and can reproduce them.
In other words: LLMs can’t understand as humans do, but since learning data captured human understanding, models can replicate that.. With no metamodel of themselves, they can’t evaluate information directly — that’s obvious. But they also process the partial output and it must map to something together with the prompt. An LLM may start generating a fairytale asked about a real event, but it means that at some point both the fairytale and the expectation of a real event will be processed together and this is where the model may fall into a path that reproduces understanding of the conflict between the two. Which of course doesn’t automatically imply it will stop or correct itself. This is not what I wanted to say. After all… it’s a stochastic parrot, not a human. It doesn’t fear that other LLMs will think badly of it because it makes mistakes.
Note that a lot of books, especially those published maybe 50 years ago, do not explicitly state that they are "fiction".
This was a minor plot point in the apocalyptic science-fiction novel A Canticle for Leibowitz (W M Miller, 1959) where a monastery has been preserving books in barrels to keep them from the book-burners.
A later expedition to read the "book-legged" books confuses R.U.R., the 1920 science-fiction play by the Czech writer Karel Čapek, with an historical text.
Which is a very human behavior. With the exception of historians, the image of history we all hold is almost entirely built on fiction. Unlike LLMs we perceive ourselves and our own thinking process, and sometimes have incentives to correct ourselves, so critical thinking may kick in and we take additional, conscious effort to verify information. But that’s the only difference.
Home
Help
Search
Login
Register
