What? This is a massive misunderstanding. It’s easy to get truly novel ideas from LLMs, unless your definition of “new things” is so strict that no human can do so either.
The training set is about patterns, not facts or specific configurations. Yes, it’s possible to extract (some) of the training set verbatim, but that doesn’t mean it’s all you can do.
>unless your definition of “new things” is so strict that no human can do so either.
Humans rarely think of new things. We're a weak hivemind species. One or two individuals figure something out, and the rest of the troop of monkeys imitates. Brains are too fuel hungry for every brain to be innovating, "innovate and copy to the other brains" is the norm.
They can think and reason better than most humans. Most problems they're pointed at are not in their training set, but in certain ways they resemble things that are—maybe there are a few different resemblances to different problems in its training set—so it's able to pull these disparate similarities together and apply the patterns it finds to come up with a solution. Much like human brains do.
It does not think at all. It vibes based on its training and any additional bolted on constraints. It is a quite simple automation that only works by huge amount of existing data.
Modern man has grown quite dumb. He only seems to be able to "invent" by massive scaling things that are decades or centuries old..
Electricity runs from simple batteries (600 BCE) to today’s power grids.
RF was predicted but not demonstrated by Maxwell in the 1860’s. His work built on Faraday’s (1840’s) and Coulomb’s (1780’s). Coulomb built on Franklin and Newton, among others. Or do you mean Marconi and Tesla, who merely implemented what Maxwell predicted?
The same is true for lasers and transistors but it’s tedious. There was no single “back in the day people invented things from whole cloth” moment.
In what way is electricity an invention? Electricity is a physical phenomenon. Various machines for doing work with electrical energy, storing electrical energy, converting other types of energy to electrical energy, etc. are certainly inventions... Heck, rubbing an amber rod with a fur is an invention. The static charge transferred is not.
I would put it differently. Those inventions came from humans interacting with the physical world.
When LLMs were first introduced, they didn't have much of a feedback loop. They wrote code, but they couldn't compile it. Not surprisingly, the code had bugs.
Now, they run with harnesses that allow them to compile the code, and react to the issues they observe. They can fix their own bugs and solve problems that they create, just like humans.
Give an agent access to the physical world, and it seems highly likely that they will be able to "invent" things based on feedback they receive while working towards goals.
Of course, there are some well-known limitations of LLMs, one of the biggest being that they're pretrained. So there may be some things where they're not as good. Just like how some humans aren't as good at certain tasks, depending on their genetics and/or how they've been trained.
Those are not merely scaling. I can get “build upon other works”, but there’s a lot of scientific insights needed for observing and modeling a phenomena. It may even requires a boost of creativity to theorize an effect based on that model and then make it possible in an experiment.
I was pointing out two things: first, your understanding of LLM capabilities is very outdated; and second, that in this respect, you're behaving much like an LLM with a training cutoff.
That further touches on the idea that the differences between you and an LLM may not be as large as you imagine. In particular, "cobbling something together if it's in their training set" is pretty much what all humans do.