It’s currently on 12%. 12% seems really xrisk-relevant to me. I think xrisk could increase quite a bit if such a model would be much more lightweight (>=2 OOMs) than LLMs.
Also I think this would still be very relevant if the timeline was longer. Maybe a new post for eg five years?
That’s the level where manifold can’t really tell you about exact probability because betting no ties up a lot of capital for minimal upside.
Also by 2026 I’d expect to have GPT 4 level LLMs with 1/10th the parameter count just due to algorithmic improvements (maybe I’m wildly wrong here) so doing the same but with different architecture isn’t necessarily as indicative as it seems.
Doing the same with a different architecture could open up the possibility of doing better down the road. I’d be equally interested in how fast it gets better as in how good it is. It would also beg the question: if two architectures can do this, how many more? Do they all max out at the same point or not at all? I think it could be quite important. Would be curious how big experts think the probability is that a different architecture could do LLM level thinking on a reasonably wide range of tasks in say five or ten years.
It’s one of the perks of reading some articles quite a long time after they were published: the hindsight lol!
I am obviously still curious about that on longer time scales though.
The above second prediction bet “trying to predict any significant architectural shifts over the next year and a half (until end of 2026)” is at time of this comment (10th June 2026) on 27% which is interesting. I would also be interested in the same prediction bets for end of 2027, 2028, 2029, 2030 and 2035 (each “end of <year>”).
Personally I am currently (10th June 2026) 50% about any significant architectural shifts before end of 2026, but 90% before end of 2027.
During 2027 it is likely the technological singularity, i.e. basically RSI (at autonomous levels between 90% and 100%), will be in full swing, so I would be surprised if no better architecture than the current SOTA LLMs has been found/used by end of 2027.
I made a manifold post for this for those who wish to bet on it: https://manifold.markets/JasonBrown/will-a-gpt4-level-efficient-hrm-bas?r=SmFzb25Ccm93bg
It’s currently on 12%. 12% seems really xrisk-relevant to me. I think xrisk could increase quite a bit if such a model would be much more lightweight (>=2 OOMs) than LLMs.
Also I think this would still be very relevant if the timeline was longer. Maybe a new post for eg five years?
Currently on 9%
That’s the level where manifold can’t really tell you about exact probability because betting no ties up a lot of capital for minimal upside.
Also by 2026 I’d expect to have GPT 4 level LLMs with 1/10th the parameter count just due to algorithmic improvements (maybe I’m wildly wrong here) so doing the same but with different architecture isn’t necessarily as indicative as it seems.
Doing the same with a different architecture could open up the possibility of doing better down the road. I’d be equally interested in how fast it gets better as in how good it is. It would also beg the question: if two architectures can do this, how many more? Do they all max out at the same point or not at all? I think it could be quite important. Would be curious how big experts think the probability is that a different architecture could do LLM level thinking on a reasonably wide range of tasks in say five or ten years.
I’ve ended up making another post somewhat to this effect, trying to predict any significant architectural shifts over the next year and a half: https://manifold.markets/Jasonb/significant-advancement-in-frontier
Well 0% now and resolved to “no”!
It’s one of the perks of reading some articles quite a long time after they were published: the hindsight lol!
I am obviously still curious about that on longer time scales though.
The above second prediction bet “trying to predict any significant architectural shifts over the next year and a half (until end of 2026)” is at time of this comment (10th June 2026) on 27% which is interesting. I would also be interested in the same prediction bets for end of 2027, 2028, 2029, 2030 and 2035 (each “end of <year>”).
Personally I am currently (10th June 2026) 50% about any significant architectural shifts before end of 2026, but 90% before end of 2027.
During 2027 it is likely the technological singularity, i.e. basically RSI (at autonomous levels between 90% and 100%), will be in full swing, so I would be surprised if no better architecture than the current SOTA LLMs has been found/used by end of 2027.