8 ms·

I have studied Rich's RL book (both editions), and enjoy his work and occasional talks on YouTube. Carmack is obviously talented. In the referenced YouTube anno

I have studied Rich's RL book (both editions), and enjoy his work and occasional talks on YouTube. Carmack is obviously talented. In the referenced YouTube announcement Carmack basically said that the average person should not really care about what they are doing right now, and the way they are funded they can take their time, not rush out any public systems or projects, etc. - at least that is the way I interpreted what he said. Off topic, but: I used to think AGI was likely, but probably not until maybe 2040. After doing a deep dive into LLMs in the last 2 years, and generally deep learning for the eight years before that, I now think there is some real chance of having something that I consider AGI by 2030. Lately LLMs have become a little useful for graph datastore, KGs, etc. I think the missing piece is the ability of LLMs to handle reasoning, and seeing GPT-4's ability to write Prolog code given a problem description, it doesn't strain my imagination too much to think a breakthrough for reasoning could be here in a year or less.

fnordpiglet 3y ago
There are a fair amount of formal reasoning systems out there for various structured reasoning tasks. LLMs may not be able to reason, but they can specify things to be reasoned about in Gallina or OWL DT or whatever and defer. I think folks seem to think AGI will be a single model, while if you look at our primary example of human intelligence it’s many models interconnected. There’s no reason for, say, LLMs to have the ability to play chess or solve differential equations or recall facts or reason formally. These already have excellent solutions that you might as well use. It’s interesting to see how far they can go in these areas, but I’m not holding my breath for an LLM to be able to out math mathematica.
computerex 3y ago
If a breakthrough in reasoning could occur in "a year or less" then why do you still have such a conservative upper bound of 2030?
- mark_l_watson 3y ago
  I think integrating better reasoning is just the next step. Thank you for considering my 2030 prediction to be “conservative.” That is refreshing for me. In 1982 I came home from AAAI with a bumper sticker “AI, It’s for Real” that I embarrassingly had on my car for years.

fnordpiglet 3y ago

There are a fair amount of formal reasoning systems out there for various structured reasoning tasks. LLMs may not be able to reason, but they can specify things to be reasoned about in Gallina or OWL DT or whatever and defer. I think folks seem to think AGI will be a single model, while if you look at our primary example of human intelligence it’s many models interconnected. There’s no reason for, say, LLMs to have the ability to play chess or solve differential equations or recall facts or reason formally. These already have excellent solutions that you might as well use. It’s interesting to see how far they can go in these areas, but I’m not holding my breath for an LLM to be able to out math mathematica.

computerex 3y ago

If a breakthrough in reasoning could occur in "a year or less" then why do you still have such a conservative upper bound of 2030?

mark_l_watson 3y ago
I think integrating better reasoning is just the next step. Thank you for considering my 2030 prediction to be “conservative.” That is refreshing for me. In 1982 I came home from AAAI with a bumper sticker “AI, It’s for Real” that I embarrassingly had on my car for years.