They couldn't find a more apt demnonstration of what an LLM is and does if they tried.
An LLM doesn't know more than what's in the training data.
In Michael Crichton's The Great Train Robbery (published in 1975, about events that happened in 1855) the perpetrator, having been caught, explains to a baffled court that he was able to walk on top of a running train "because of the Bernoulli effect", that he misspells and completely misunderstands. I don't remember if this argument helps him get away with the crime? Maybe it does, I'm not sure.
> At this point, the prosecutor asked for further elucidation, which Pierce gave in garbled form. The summary of this portion of the trial, as reported in the Times, was garbled still further. The general idea was that Pierce--- by now almost revered in the press as a master criminal--- possessed some knowledge of a scientific principle that had aided him.
> An LLM doesn't know more than what's in the training data.
Post-training for an LLM isn't "data" anymore, it's also verifier programs, so it can in fact be more correct than the data. As long as search finds LLM weights that produce more verifiably correct answers.
I know that some specific parts of what's in my training data is false, even though it was in there often. I am not just the average-by-volume of everything I've read.
It's a good question, but there are things I figured out by myself, that weren't in my training data, some, even, where my training data said the exact opposite.
An LLM doesn't know more than what's in the training data.
In Michael Crichton's The Great Train Robbery (published in 1975, about events that happened in 1855) the perpetrator, having been caught, explains to a baffled court that he was able to walk on top of a running train "because of the Bernoulli effect", that he misspells and completely misunderstands. I don't remember if this argument helps him get away with the crime? Maybe it does, I'm not sure.
This is another attempt at a Great Robbery.