I would take most of these bets (that is, bet that these will happen by 2027).
(1,10) Are possible today, but a matter of scaling up to longer contexts and more complex problems.
(2,3,5) Are probably already possible
(4) Interesting challenge, but probably doable.
Only for (6,7,8,9) I'm more skeptical. What is needed there is what Cantwell-Smith calls "judgment": the AI needs to be an entity in the world that in some sense "takes responsibility for its actions".
@pbloem Disagreed on 4 especially, having recently played some cognitively more involved video games, I realised there's no way current day AI can reasonably beat those anytime soon. (Think: Outer Wilds, Tunic, Animal Well).
10 is definitely already possible, if with frequent or many errors
@EmilevanKrieken We definitely need some sort of ImageNet moment on 4 for world models and interaction. But in three years, we may well get one. There's a lot of push behind those two topics.
"Doable" was probably too optimistically phrased, but I think I might put the probability above 0.5.
@EmilevanKrieken I feel like there's a Moravec paradox at work here.
ChatGPT is already quite good at zero shotting logic puzzles in natural language (not flawless, but impressive enough to expect a lot of progress to come). On the other hand models like SORA seem to have absolutely no coherent world model despite having seen god-knows-how-many hours of video.
So, maybe the difficult part is the world modeling and interaction, not the generalized puzzle solving (maybe).
Peter Bloem
Emile van Krieken