This path feels correct to me. It feel like what we do as humans and seems like a reasonable way to start to construct "mode 2" thinking.
IDK if our current models have enough of "mode 1" to power this system. It's also plausible that our current "mode 1" systems are more than powerful enough and that inference speed (and thus the size/depth of the tree that can be explored) will be the most important factor.
I hope that the major players are looking at this and trying it out at scale (I know Deepmind wrote the orginal paper, but their benchmarks were quite unimpressive). It's plausible that we will have an AlphaGo moment with this scheme.
I believe you are correct here, yet at the same time I think we're about 2 orders of magnitude off on the amount of compute power needed to do it effectively.
I think the first order of mag will be in tree of thought processing. The amount of additional queries we need to run to get this to work is at least 10x, but I don't believe 100x.
I think the second order of mag will be multimodal inference so the models can ground themselves in 'reality' data. Saying, "the brick layed on the ground and did not move" and "the brick floated away" are only deciable based on the truthfulness of all the other text corpus it's looked at. At least to me it gets even more interesting when you tie it into environmental data that is more likely to be factual, such as massive amounts of video.
Yeah looks very promising. Naively, it multiplies computation time by a factor of 20x though? If they are taking 5x samples per step, and multiple steps per problem.
As this gets explored further, I believe we will start finding out why human minds are constructed the way they are, from the practical/necessity direction. Seems like the next step is farming out subtasks to smaller models, and adding an orthogonal dimension of emotionality to help keep track of state.
I’m sympathetic to the idea of new types of specialized models to assist in this effort. We’re using our one hammer for all problems.
In particular, it jumps out that a “ranking model” (different, I think from current ranking models) to judge which paths to take and which nodes to trim would make some level of sense.
Not sure if it's relevant, but the OpenAI APIs generally support taking multiple responses in a single API call. I'm unsure what the generalized effect on processing time of that is however. From what I've read, this is sub-linear, so could reasonably be more effective than 20x, and I'd bet there are probably speedups to be had on the model side of this that make the extra time cost negligible.
IDK if our current models have enough of "mode 1" to power this system. It's also plausible that our current "mode 1" systems are more than powerful enough and that inference speed (and thus the size/depth of the tree that can be explored) will be the most important factor.
I hope that the major players are looking at this and trying it out at scale (I know Deepmind wrote the orginal paper, but their benchmarks were quite unimpressive). It's plausible that we will have an AlphaGo moment with this scheme.