I appreciate the rigor you have brought to this ambiguous topic. It seems to me that given the binding constraint of compute capacity in all of these future scenarios that somehow obtaining and utilizing said scarce resource might be incorporated into further iterations on this concept. Of course, that brings up the whole question of the interface between AI systems and the physical world which is a whole other question…
I think the labs believe RSI is similar to the "continuous learning" outlined in https://ai-2027.com. Doesn't seem to far fetched that a suite of algorithms could improve upon their own data and reasoning traces in an unguided way, much like how "self-play" led to emergent properties in OpenAI's https://openai.com/index/emergent-tool-use/ Might look like a combination of self-play from AlphaGo that updates a static model's weights in near realtime. Could be cool, I guess.
I appreciate the rigor you have brought to this ambiguous topic. It seems to me that given the binding constraint of compute capacity in all of these future scenarios that somehow obtaining and utilizing said scarce resource might be incorporated into further iterations on this concept. Of course, that brings up the whole question of the interface between AI systems and the physical world which is a whole other question…
Yes, this raises a whole different question regarding what happens after RSI.
One recent tweet about this:
https://x.com/deredleritt3r/status/2073126436064387266
I think the labs believe RSI is similar to the "continuous learning" outlined in https://ai-2027.com. Doesn't seem to far fetched that a suite of algorithms could improve upon their own data and reasoning traces in an unguided way, much like how "self-play" led to emergent properties in OpenAI's https://openai.com/index/emergent-tool-use/ Might look like a combination of self-play from AlphaGo that updates a static model's weights in near realtime. Could be cool, I guess.