The Jung-Freud problem (which you mentioned before in another podcast) is very interesting! The trouble is, not only can you not trust the LLM, but you can't even trust the pre-training data, at least not with the way that pre-training is conventionally done today. A similar issue could arise in fictional, hypothetical, and historical scenarios. For example, it's good for ChatGPT to think the world is round, but in a fictional setting with a flat world, that'd be undesirable. So this is a very generally applicable problem.
extremely interesting. one note on the ending: it is possible that these LLM derived minds will not be conscious. eg David Chalmers guesses that there is a 10% probability that current models are conscious and a 20% probability that future models will be conscious. Ofc no one really knows, but the prospect of getting replaced by completely unconscious beings seems unappealing to say the least...
I wouldnt call the idea that there are no human brains in 1 million years provocative. Youd have to have been stupid to believe otherwise given the rate of advancement.
You said it has implications at the level of the universe that so soon after our ape brain evolving we can build AGI. What did you mean? I presume you mean that biological advanced intelligence is probably rare and transient and that the universe should be filled instead with AI. It so where are they?
Taken further, it suggests we might live in a simulation. If you can create AGIs you can create virtual worlds filled with agents who may not know they are AIs... The timescale during which most sentient beings are evolved biological creatures (versus mostly AIs, perhaps living in virtual worlds) seems very small (transient) compared to the age of the universe.
![ปฏิเสธไม่ไหว (Crush on you) - LIPTA feat. No One Else [Official MV]](http://i.ytimg.com/vi/NBw1jF342vw/mqdefault.jpg)
The Jung-Freud problem (which you mentioned before in another podcast) is very interesting! The trouble is, not only can you not trust the LLM, but you can't even trust the pre-training data, at least not with the way that pre-training is conventionally done today.
A similar issue could arise in fictional, hypothetical, and historical scenarios. For example, it's good for ChatGPT to think the world is round, but in a fictional setting with a flat world, that'd be undesirable. So this is a very generally applicable problem.
Can you pls point me towards the podcast you mentioned in your comment?
@@sachetanbhat I think it was Danny In The Valley's 2023-04-20 episode, interviewing Hsu.
extremely interesting.
one note on the ending: it is possible that these LLM derived minds will not be conscious. eg David Chalmers guesses that there is a 10% probability that current models are conscious and a 20% probability that future models will be conscious. Ofc no one really knows, but the prospect of getting replaced by completely unconscious beings seems unappealing to say the least...
Pretty interesting and provocative transhumanist perspective towards the end
I wouldnt call the idea that there are no human brains in 1 million years provocative. Youd have to have been stupid to believe otherwise given the rate of advancement.
That ending for me is very pessimistic, but I’m guessing Steve would describe it as optimistic. Regardless, it seems more or less inevitable
You said it has implications at the level of the universe that so soon after our ape brain evolving we can build AGI. What did you mean? I presume you mean that biological advanced intelligence is probably rare and transient and that the universe should be filled instead with AI. It so where are they?
Taken further, it suggests we might live in a simulation. If you can create AGIs you can create virtual worlds filled with agents who may not know they are AIs... The timescale during which most sentient beings are evolved biological creatures (versus mostly AIs, perhaps living in virtual worlds) seems very small (transient) compared to the age of the universe.
👍👍👍
Well that was pessimistic