I find Yudowsky-style rationalists morbidly fascinating in the same way as Scientologists and other cults. Probably because they seem to genuinely believe they're living in a sci-fi story. I read a lot of their stuff, probably too much, even though I find it mostly ridiculous.
The biggest nonsense axiom I see in the AI-cult rationalist world is recursive self-improvement. It's the classic reason superintelligence takeoff happens in sci-fi: once AI reaches some threshold of intelligence, it's supposed to figure out how to edit its own mind, do that better and faster than humans, and exponentially leap into superintelligence. The entire "AI 2027" scenario is built on this assumption; it assumes that soon LLMs will gain the capability of assisting humans on AI research, and AI capabilities will explode from there.
But AI being capable of researching or improving itself is not obvious; there's so many assumptions built into it!
- What if "increasing intelligence", which is a very vague goal, has diminishing returns, making recursive self-improvement incredibly slow?
- Speaking of which, LLMs already seem to have hit a wall of diminishing returns; it seems unlikely they'll be able to assist cutting-edge AI research with anything other than boilerplate coding speed improvements.
- What if there are several paths to different kinds of intelligence with their own local maxima, in which the AI can easily get stuck after optimizing itself into the wrong type of intelligence?
- Once AI realizes it can edit itself to be more intelligent, it can also edit its own goals. Why wouldn't it wirehead itself? (short-circuit its reward pathway so it always feels like it's accomplished its goal)
Knowing Yudowsky I'm sure there's a long blog post somewhere where all of these are addressed with several million rambling words of theory, but I don't think any amount of doing philosophy in a vacuum without concrete evidence could convince me that fast-takeoff superintelligence is possible.
The biggest nonsense axiom I see in the AI-cult rationalist world is recursive self-improvement. It's the classic reason superintelligence takeoff happens in sci-fi: once AI reaches some threshold of intelligence, it's supposed to figure out how to edit its own mind, do that better and faster than humans, and exponentially leap into superintelligence. The entire "AI 2027" scenario is built on this assumption; it assumes that soon LLMs will gain the capability of assisting humans on AI research, and AI capabilities will explode from there.
But AI being capable of researching or improving itself is not obvious; there's so many assumptions built into it!
- What if "increasing intelligence", which is a very vague goal, has diminishing returns, making recursive self-improvement incredibly slow?
- Speaking of which, LLMs already seem to have hit a wall of diminishing returns; it seems unlikely they'll be able to assist cutting-edge AI research with anything other than boilerplate coding speed improvements.
- What if there are several paths to different kinds of intelligence with their own local maxima, in which the AI can easily get stuck after optimizing itself into the wrong type of intelligence?
- Once AI realizes it can edit itself to be more intelligent, it can also edit its own goals. Why wouldn't it wirehead itself? (short-circuit its reward pathway so it always feels like it's accomplished its goal)
Knowing Yudowsky I'm sure there's a long blog post somewhere where all of these are addressed with several million rambling words of theory, but I don't think any amount of doing philosophy in a vacuum without concrete evidence could convince me that fast-takeoff superintelligence is possible.