It doesn’t actually solve those math problems though, does it? It replies with a solution if it has seen one often enough in training data or something that looks like a solution but isn’t. At the end, the human still needs to proof it.
Same for short stories, it doesn’t actually write new stories, it rehashes stories it (probably illegally) ingested in training data.
LLMs are good at mimicking the content they were trained on, they don’t actually adopt or extend the intelligence required to create that content in the first place.
Oh, I remember those talks. People actually checking whether an LLM's response is something that was in the training data, something that was online that it replicated, or something new.
They weren't finding a lot of matches. That was odd.
That was in the days of GPT-2. That was when the first weak signs of "LLMs aren't just naively rephrasing the training data" emerged. That finding was controversial, at the time. GPT-2 couldn't even solve "17 + 29". ChatGPT didn't exist yet. Most didn't believe that it was possible to build something like it with LLM tech.
I wish I could say I was among the people who had the foresight, but I wasn't. Got a harsh wake-up call on that.
And yet, here we are, in year 20-fucking-25, where off-the-shelf commercially available AIs burn through math competitions and one shot coding tasks. And people still say "they just rehash the training data".
Because the alternative is: admitting that we found an algorithm that crams abstract thinking into arrays of matrix math. That it's no longer human exclusive. And that seems to be completely unpalatable to many.
Same for short stories, it doesn’t actually write new stories, it rehashes stories it (probably illegally) ingested in training data.
LLMs are good at mimicking the content they were trained on, they don’t actually adopt or extend the intelligence required to create that content in the first place.