aijooyoomJul 3, 2023
Tero Keski-Valkama

I'm very optimistic about the next steps of #LargeLanguageModels. Self-competition with #ReinforcementLearning is certainly something many people are working on.

#Google's #Gemini doesn't have many public details yet but is apparently going to use planning engines in #AlphaGo style to achieve the next level of performance in complex tasks. I believe this needs to be trained non-imitatively with reinforcement learning, so it will likely achieve the self-competition gains as well, even ignoring the gains from a possible incorporation of a search/planning engine.

We are now experiencing the quantum leap between human-level systems trained imitatively, and immeasurably above human level systems trained with self-competition in the general language domain.

We'll certainly get an unambiguous #AGI before the end of the year. The beginning is nigh!

Jun 28, 2023 at 10:39amWeb
Show threadTero Keski-ValkamaJun 28, 2023

The next interesting things after unambiguous AGI will be:
- Achieving low-energy AGI with novel learning rules, executable in very limited computing systems.
- Robotics and logistics of intelligence: How to make smart stuff happen in the correct places at the correct times. Need to make lots of factories making mobile things globally.
- Novel computing substrates designed by AGI, combining memory and execution, analog computing, optimizing the trade-off between complete exactness and speed.
- Societies of AIs and AI agents living along societies of humans.

And let's not forget the complete social transformation:
- Who are the winners, who are the losers. Hopefully we will share the fruits of progress equally.
- What happens to work and leisure, rents and wages, taxes and borders?
- Space. The never-ending frontier beginning from now. Our machines will meet all the things out there.
- Scientific integration and the speed of progress which makes all our efforts so far seem like cave painting. Medicine, engineering, art, history, ..., science and culture in all their forms will be more or less "solved" as far as humans are concerned. Humans can spend their whole endless lives appreciating the discoveries and art made by machines and not even scratch the surface.