Obviously one could e.g. fill a small wiki with ChatGPT’s thoughts on the alignment problem, by continuing to ask for questions, and then answers to the questions…
The main benefit of the discussion was that in order to correct ChatGPT’s vague (but characteristically confident) descriptions of the proposal, I had to refine my own understanding of it!
Feburary 2023 is ending. I just got access to Bing AI (described as an “answer engine” rather than a search engine).
I missed out on the “Sydney” era, but I found that, once I was past the part where it’s just summarizing material from the web, it can still be a stimulating and challenging conversation partner, even when discussing specific ideas for AI alignment:
I just had a go at using GPT-J to solve the alignment problem. This was its output:
https://pastebin.com/iTHAZsdV
Not actually a solution, but interesting to see its attempts.
Should I write a post about this? Have there been any more successful attempts to use language models to solve the problem?
It’s now January 2023, seven months later. Here is what one can now do in one minute with ChatGPT:
https://pastebin.com/0EUBbvGd
Obviously one could e.g. fill a small wiki with ChatGPT’s thoughts on the alignment problem, by continuing to ask for questions, and then answers to the questions…
It’s now February 2023. I discussed some aspects of a specific alignment proposal with ChatGPT:
https://pastebin.com/B1ntAxm7
The main benefit of the discussion was that in order to correct ChatGPT’s vague (but characteristically confident) descriptions of the proposal, I had to refine my own understanding of it!
Feburary 2023 is ending. I just got access to Bing AI (described as an “answer engine” rather than a search engine).
I missed out on the “Sydney” era, but I found that, once I was past the part where it’s just summarizing material from the web, it can still be a stimulating and challenging conversation partner, even when discussing specific ideas for AI alignment:
https://pastebin.com/DPMenYHU