So certain aspects of the lesswrong ‘consensus view’ I have a lot of doubts on. But they are all on the end of what’s happening right now. (nanotechnology, can an ASI just ‘hack anything’ and ‘persuade anybody’ and ‘coordinate with itself’, or “optimize itself to fit into cheap computers that are available”, or “it’s pointless to try to control it”?)
I would have to ask, if AGI is defined as “median human ability”, and the argument around the GPT-3.5 release was “it can only read and emit text”, how do you explain:
Above median human ability on many tests with GPT-4
The ‘context length’ barrier was lifted with Gemini 1.5 and Claude.
The multimodality limit (I had doubts on the integration of extra modalities) was lifted with Gemini and Claude.
4. If it’s just “hype”, why does GPT-4 with plugins solve various basic engineering and physics problems about as well as an undergrad? How do you explain Claude writing fairly long chunks of code that works? (I have personally tried both)
5. How do you explain the rate of progress.
6. Have you considered that all those “sciencedirect” news you saw over the years were from tiny startups and single professor university labs? That more is being pumped into AI every year than fusion power since the beginning? Scale matters.
7. Why are so many investors voting with their money? Are they just stupid and tricked by hype or do they know something as a group that you don’t?
It seems to me that ‘all’ we have to do to reach AGI is integrate:
online learning, robotics i/o (to a system 1 model that directly controls the robot), video perception, audio perception, internal buffer for internal monologue
Note that all already exist in someone’s lab.
Such a machine would have about the breadth and skill as the median human. That’s AGI. People keep pushing that definition further (“expert human level at everything”, “can research AI autonomously and faster”) but a median human worker you can print is a gamechanger.
So certain aspects of the lesswrong ‘consensus view’ I have a lot of doubts on. But they are all on the end of what’s happening right now. (nanotechnology, can an ASI just ‘hack anything’ and ‘persuade anybody’ and ‘coordinate with itself’, or “optimize itself to fit into cheap computers that are available”, or “it’s pointless to try to control it”?)
I would have to ask, if AGI is defined as “median human ability”, and the argument around the GPT-3.5 release was “it can only read and emit text”, how do you explain:
Above median human ability on many tests with GPT-4
The ‘context length’ barrier was lifted with Gemini 1.5 and Claude.
The multimodality limit (I had doubts on the integration of extra modalities) was lifted with Gemini and Claude.
4. If it’s just “hype”, why does GPT-4 with plugins solve various basic engineering and physics problems about as well as an undergrad? How do you explain Claude writing fairly long chunks of code that works? (I have personally tried both)
5. How do you explain the rate of progress.
6. Have you considered that all those “sciencedirect” news you saw over the years were from tiny startups and single professor university labs? That more is being pumped into AI every year than fusion power since the beginning? Scale matters.
7. Why are so many investors voting with their money? Are they just stupid and tricked by hype or do they know something as a group that you don’t?
It seems to me that ‘all’ we have to do to reach AGI is integrate:
online learning, robotics i/o (to a system 1 model that directly controls the robot), video perception, audio perception, internal buffer for internal monologue
Note that all already exist in someone’s lab.
Such a machine would have about the breadth and skill as the median human. That’s AGI. People keep pushing that definition further (“expert human level at everything”, “can research AI autonomously and faster”) but a median human worker you can print is a gamechanger.
I wrote a long reply to your points, but ultimately decided it was a derail to original topic. I’ll PM you just for fun though.