I'm not sure about current LLM techniques leading us there.
Current LLM-style systems seem like extremely powerful interpolation/search over human knowledge, but not engines of fundamentally new ideas, and it’s unclear how that turns into superintelligence.
As we get closer to a perfect reproduction of everything we know, the graph so far continues to curve upward. Image models are able to produce incredible images, but if you ask one to produce something in an entirely new art style (think e.g. cubism), none of them can. You just get a random existing style. There have been a few original ideas - the QR code art comes to mind[1] - but the idea in those cases comes from the human side.
LLMs are getting extremely good at writing code, but the situation is similar. AI gives us a very good search over humanity's prior work on programming, tailored to any project. We benefit from this a lot considering that we were previously constantly reinventing the wheel. But the LLM of today will never spontaneously realise there there is an undiscovered, even better way to solve a problem. It always falls back on prior best practice.
Unsolved math problems have started to be solved, but as far as I'm aware, always using existing techniques. And so on.
Even as a non-genius human I could come up with a new art style, or have a few novel ideas in solving programming problems. LLMs don't seem capable of that (yet?), but we're expecting them to eventually have their own ideas beyond our capability.
Can a current-style LLM ever be superintelligent? I suppose obviously yes - you'd simply need to train it on a large corpus of data from another superintelligent species (or another superintelligent AI) and then it would act like them. But how do we synthesise superintelligent training data? And even then, would they be limited to what that superintelligence already knew at the time of training?
Maybe a new paradigm will emerge. Or maybe things will actually slow down in a way - will we start to rely on AI so much that most people don't learn enough for themselves that they can make new novel discoveries?
[1] https://www.reddit.com/r/StableDiffusion/comments/141hg9x/co...
> Can a current-style LLM ever be superintelligent? I suppose obviously yes - you'd simply need to train it on a large corpus of data from another superintelligent species
This is right, but we can already do that a little bit for domains with verification. AlphaZero is an example of alien-level performance due to non-human training data.
Code and math is kind of in the middle. You can verify it compiles and solves the task against some criteria. So creative, alien strategies to do the thing can and will emerge from these synthetic data pipelines.
But it's not fully like Go either, because some of it is harder to verify (the world model, unstated requirements, meta-level questions). That's the frontier challenge. How to create proxies where we don't have easy verification, from which alien performance can emerge? If this GPTZero moment arrives, all bets are off.
The main issue with novel things is that they look like random noise / trashy ideas / incomprehensible to most people.
Even if LLMs or some more advanced mechanical processes were able to generate novel ideas that are "good", people won't recognize those ideas for what they are.
You actually need a chain of progressively more "average" minds to popularize good ideas to the mainstream psyche, i.e. prototypically, the mad scientist comes up with this crazy idea, the well-respected thought leader who recognizes the potential and popularizes it to people within the niche field, the practitioners who apply and refine the idea, and lastly the popular-science efforts let the general public understand a simplified version of what it's all about.
Usually it takes decades.
You're not going to appreciate it if your LLM starts spewing mathematics not seen before on Earth. You'd think it's a glitch. The LLM is not trained to give responses that humans don't like. It's all by design.
When you folks say AI can't bring new ideas, you're right in practice, but you actually don't know what you're asking for. Not even entities with True Intelligence can give you what you think you want.
Certain classes of problems can be solved by searching over the space of possible solutions, either via brute force or some more clever technique like MCTS. For those types of problems, searching faster or more cleverly can solve them.
Other types of problems require measurement in the real world in order to solve them. Better telescopes, better microscopes, more accurate sensing mechanisms to gather more precise data. No AI can accomplish this. An AI can help you to design better measurement techniques, but actually taking the measurements will require real time in the real world. And some of these measurement instruments have enormous construction costs, for example CERN or LIGO.
All of this is to say that there will color point at our current resolution of information that no more intelligence can actually be extracted. We’ve already turned through the entire Internet. Maybe there are other data sets we can use, but everything will have diminishing returns.
So when people talk about trillion dollar superclusters, that only makes sense in a world where compute is the bottleneck and not better quality information. Much better to spend a few billion dollars gathering higher quality data.