4 ms·

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 ri

by blagie 3y ago

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.

steve_adams_86 3y ago
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.
Buttons840 3y ago
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.
- blagie 3y ago
  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.
raverbashing 3y ago
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.
janekm 3y ago
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 https://www.youtube.com/watch?v=Ipe9xJCfuTM There's enough in there to approximate the sensorial inputs that humans get through their life.

by blagie 3y ago

steve_adams_86 3y ago

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.

Buttons840 3y ago

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.

blagie 3y ago
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.

raverbashing 3y ago

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.

janekm 3y ago

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 https://www.youtube.com/watch?v=Ipe9xJCfuTM There's enough in there to approximate the sensorial inputs that humans get through their life.