> "You can’t just train an LLM with more and more parameters and more and more tokens and expect it to be smarter than the data it was trained on." This is wrong. I saw so many people who think this though, even smart people. But it's just clearly not true if you think about it. It's like saying that you can't train a model to predict a trend from a scatter plot because the model can't be smarter than the average poi…
Even if the statement were true, the data that these models are trained on is, collectively, smarter than the smartest person who ever lived. An individual person simply cannot absorb, in an 80-year-or-so lifetime, even a tiny fraction of the amount of information contained in the training data. It is a truly mind-boggling amount of data. The most frightening thing is that, even with pretty naive and primitive algori…
"Smarter" in a horizontal sense. But not vertical.
The latter being what people really mean by "smarter". You know, that thinking-outside-the-box stuff.
That's why LLMs aren't able to "solve" problems that don't have existing writeups somewhere (beyond certain problems more or less equivalent to combinatorial search). Or produce text useful beyond the mild amusement level (or for spam and/or research fraud -- at which it of course excels quite brilliantly).
And still trips over itself narratively, left, right and center. And still makes all kind of cute-sounding, and superficially correct (but ultimately misleading) pronouncements about the world that a true expert would never make.
Taking an example from a recent NYT article: "Give me some useful mnemonics for learning Spanish" (paraphrased slightly):
Hola: "Hola, cola" - sounds like "cola" (Coke)
Adiós: "Adiós, my toes" - rhymes with "toes"
Gracias: "Gracias, grassy ass" - sounds like "grassy ass"
Mañana: "Mañana, banana" - rhymes with "banana"
Uno: "Uno, you know" - sounds like "you know"
Tres: "Tres, dress" - rhymes with "dress"