John Carmack’s ‘Different Path’ to Artificial General Intelligence
dallasinnovates.com
John Carmack’s ‘Different Path’ to Artificial General Intelligence
1–10 of 506 posts
Re: John Carmack’s ‘Different Path’ to Artificial General Intelligence
#2Fun interview but barely any meat for those in the field. Just very general questions and answers, saying that the road is murky, but nothing about e.g. if Transformers / attention are the way to go forward, multi-modal models, reinforcement learning + self-supervised learning.
Re: John Carmack’s ‘Different Path’ to Artificial General Intelligence
#3What would be the difference between GPT3 and actual AGI? (Want to make sure we don’t keep moving goalposts)
Re: John Carmack’s ‘Different Path’ to Artificial General Intelligence
#4What would be the difference between GPT3 and actual AGI? (Want to make sure we don’t keep moving goalposts)
AGI is usually defined as anything a normal human could do. Examples of some things that GPT is currently lacking:
1) Understanding images and video
2) Learning and remembering things outside the 2048 token context window
3) Interacting with the environment (either through a computer, or in the real world)
4) Doing basic math (1662 + 239 = ?) and logic
Re: John Carmack’s ‘Different Path’ to Artificial General Intelligence
#5Carmack seems to be simulating a vtuber AI watching TV and playing video games.
If anyone here is doing that too, I would recommend taking a quick look at Neuro-sama on Twitch. They're using RL to play OSU, Minecraft, and Pokemon, and voice input + video image analysis to react to Twitch streams and documentaries. While being watched by 6.5K people.
The url is twitch.com/vedal
Re: John Carmack’s ‘Different Path’ to Artificial General Intelligence
#6What would be the difference between GPT3 and actual AGI? (Want to make sure we don’t keep moving goalposts)
The same as the difference between GPT3 and humans’ cognitive abilities.
Re: John Carmack’s ‘Different Path’ to Artificial General Intelligence
#7Fun interview but barely any meat for those in the field. Just very general questions and answers, saying that the road is murky, but nothing about e.g. if Transformers / attention are the way to go forward, multi-modal models, reinforcement learning + self-supervised learning.
There is a comment in the article about having models watch TV and play video games, and he's talked about that before too in his Lex Friedman interview. Seems like his approach is to take existing model architectures, apply some tweaks and experimental ideas and then use datasets consisting of TV (self supervised learning maybe?) and classic video games (RL I guess?).
The video game part at least sounds like what Deepmind is already doing. I guess we'll just have to wait and see what he plans to do differently.
It seems to me like his expertise would be most valuable in optimizing model architectures for hardware capabilities to improve utilization and training efficiency. That will be important for AGI especially as the cost of training models skyrockets (both time and money). If I was a startup doing AI hardware like Cerebras or Graphcore I would definitely try to hire Carmack to help with my software stack. Though he doesn't seem interested in custom AI hardware.
Re: John Carmack’s ‘Different Path’ to Artificial General Intelligence
#8What would be the difference between GPT3 and actual AGI? (Want to make sure we don’t keep moving goalposts)
A first step would be to make inductive inferences from first principles. Current GPT models struggle with basic math. If you tell it "sorry you are wrong" it will regurgitate back the incorrect answer you gave. The language model is dumb enough that no amount of string escaping will prevent it from dumping its prompt.
The biggest problem right now is online information. We still don't have a good way to teach it new information aside from single epoch training or prompt engineering. If we want a model to constantly learn and update itself, then we need a robust way of information retrieval and storage, possibly either through knowledge graphs or child network modules. (Are neural Turing machines still a thing? What happened to capsules?)
Re: John Carmack’s ‘Different Path’ to Artificial General Intelligence
#9Fun interview but barely any meat for those in the field. Just very general questions and answers, saying that the road is murky, but nothing about e.g. if Transformers / attention are the way to go forward, multi-modal models, reinforcement learning + self-supervised learning.
I do not question for a second that Carmack is a computing genius, but “murky” is an understatement. There is nothing of technical, scientific or business value in this interview. It’s just generic words strung together.
Re: John Carmack’s ‘Different Path’ to Artificial General Intelligence
#10>And one of the things that I trained myself to do a few decades ago is pulling ideas out and pursuing them in a way where I’m excited about them, knowing that most of them don’t pan out in the end. Much earlier in my career, when I’d have a really bright idea that didn’t work out, I was crushed afterwards. But eventually I got to the point where I’m really good at just shoveling ideas through my processing and shooting them down, almost making it a game to say, ‘How quickly can I bust my own idea, rather than protecting it as a pet idea?’
Cool life lesson there