I don't think that's a limit. If you give me $10M and 5 years, I can tell you a half-dozen ways to train models better than GPT4, primarily by having them do richer tasks than text completion, by having them evaluate themselves in richer ways, by having workflows around them, and by having shared models perform multiple types of tasks (e.g. text, image, controls, etc.).

It's not that I'm especially smart; many others could tell you ways to do the same thing now that there's an initial proof-of-concept. It's that these things are new enough no one has had those five years yet. GPT4 came out two months ago, ChatGPT maybe seven months ago, and GPT3 three years ago.

I can't predict if these things will level off, grow linearly, grow exponentially, Moore's Law style, or explode off into the singularity.

I can say GPT4 is nowhere close to the limit.

There are plenty of efforts at the moment which indicate that you’re right.

I think we’ll learn to compose models for more diverse and capable outputs, validate those outputs better, and do all of it more efficiently. The end result seems like it could be far greater than the sum of its parts.

This reminds me of reinforcement learning and the Bellman equations.

Let a neural network based agent act in an environment. Record its rewards. Update the neural network so that the best actions become a little more likely, and the worse actions become a little less likely. Thus, we have demonstrated a training step that improves the agent.

Now just run that training step in a loop forever and you'll improve each step and reach Godlike intelligence, right? Right? Well, no, there is instability in "pulling yourself up by the bootstraps", and while I expect to see more GTP improvements, I believe there may be limitations we don't know about yet.

That sounds like technobabble, which is to say big words like "instability" and "Bellman equations," but with little connecting logic.

If there is meaning behind those words, please write something with multiple paragraphs.

I don't see any reason why there would be a limit, instability, or otherwise, any more or less than there is on biological evolution. I also don't see any reason why there wouldn't be a limit. We just have no idea at this point.

I do think a diversity of tasks is critical. An AGI should be able to not just complete human language but:

- Play a diversity of strategic games like chess or Starcraft

- Perform machine vision like Stable Diffusion

- Predict the output of Python programs (and vice-versa)

- Write (correct) mathematical proofs and arguments

- Control a robotic arm or airplane

... and so on.

We can train all of those independently -- and we could do so in a single network now -- but we're at the very early stages of architecture for how those things would integrate.

At some point, the best solution is a very sophisticated intelligence. How close deep learning in current -- and future -- architectures gets us to super-intelligence is an open question.

Especially as we're not training the LLMs on really specialized stuff

Internet content does not go as deep as it would ideally be in some areas.

For example, do training sets contain ArXiv data? Or other stuff behind a paywall?

LLMs know very little about XX century literature.

Absolutely. There's no doubt teams already training multi-modal models on subsets of the videos in YouTube (there's already published examples training on Minecraft videos).

There's even tons of POV videos now literally showing human life experience in many scenarios: https://www.youtube.com/watch?v=Ipe9xJCfuTM

There's enough in there to approximate the sensorial inputs that humans get through their life.

I don't get all the negativity unlike most humans LLM can be made to discard bad or incorrect data. Quickly transition to new ways of doing something. A human spending all their life in a room in front of the computer would not can not go through all the millions of scientific papers and journals. They would not look at unlimited amount of code or go through all the past and present law of every country . No human is capable of that.

We are entering into the real culmination of the information age where previously information was available but not exactly accessible or usable for a common person. Which like all tools will result in good and bad outcomes.

Ask any LLM about the cheese inventors in far East china. It'll make something up very plausible. Then all you have to do is to dig through libraries and historic records to verify that this all really is true and is really rooted in reality.

That's why you are better served asking LLMs for factual information when they are connected to tools like search. ChatGPT with search enabled, Bing.chat or https://www.phind.com/ will give you sourced and much more reliable output.

I'm not saying it's not useful. I'm saying that once it learns the content of the data sources that are fed to it, that's it. It will spit out some overlooked connections then stop. Maybe the cure for cancer can be found just by reading in between the lines of millions of old scientific papers but I doubt it.