11–20 of 35 posts
I wonder if we can even go beyond that and find ways to abstract "discoverability" so that AI agents can evolve into diverse species, perhaps entirely different from our specific ways of discovering. As an analogue, I think of the nervous system of octopuses, which is very different from ours, but capable of amazing feats.
Four very interesting concepts in a short space: - "general methods that leverage computation are ultimately the most effective, and by a large margin" - "[search and learning are] methods that continue to scale with increased computation" - "We should stop trying to find simple ways to think about the contents of minds" - "We want AI agents that can discover like we can, not which contain what we have discovered" In…
Four very interesting concepts in a short space: - "general methods that leverage computation are ultimately the most effective, and by a large margin" - "[search and learning are] methods that continue to scale with increased computation" - "We should stop trying to find simple ways to think about the contents of minds" - "We want AI agents that can discover like we can, not which contain what we have discovered" In…
Intuition is a curse though. The data to hand clearly indicates that the big (compute), dumb (search and learning) approach is the only way that’s worked. Emergent properties aren’t very satisfying, particularly when we can’t model or understand, but This is The Way.
This is a philosophical and epistemological matter that is often undiscussed. Cognitive philosophers are still hung up on questions from the 70s and their offshoots ("hard problem") etc. On the other hand i am not sure if the "computational" people often know what they are doing. Looking at something like the deep Transformer models, one has to ask if there is any rhyme or reason there, or the thing just works becaus…
This interesting discovery, when applied to brains, means the same brain trained with different data would be able to display emerging abilities. Maybe these abilities are more in the data than the architecture.
If we want better AI we need to come up with better data.
> [the critics of brute force] said that ``brute force" search may have won this time, but it was not a general strategy, and anyway it was not how people played chess. No expert in AI or chess, but I assume that this observation is wrong. Chess players do indeed construct moves based on a search of an interior mental library... a sort of pattern recognition. Any chess players here agree/disagree? As an artist who is…
I don't think so. It will explode into a kaleidoscope of diverse styles and ideas, some amazing, some meh, and some horrible. We'll be able to rank by preference and aesthetic scores to get to the ones we like.
"We want AI agents that can discover like we can, not which contain what we have discovered." I wonder if we can even go beyond that and find ways to abstract "discoverability" so that AI agents can evolve into diverse species, perhaps entirely different from our specific ways of discovering. As an analogue, I think of the nervous system of octopuses, which is very different from ours, but capable of amazing feats.
Combine "search" with "learning". Or content generation with content validation, and retrain on the clean outputs. Or, run many simulations such as AlphaGo, and learn from the outcomes.
In general the idea is to use lots of compute to generate interesting and hard to come by training data for the next iteration. This approach is necessary because we have exhausted most of the good training data and need a path forward. You can't copy money, but you can copy the model and data.
Four very interesting concepts in a short space: - "general methods that leverage computation are ultimately the most effective, and by a large margin" - "[search and learning are] methods that continue to scale with increased computation" - "We should stop trying to find simple ways to think about the contents of minds" - "We want AI agents that can discover like we can, not which contain what we have discovered" In…
Sutton also raises one uncomfortable question. Where are the limits of "our field". If we follow Sutton's (interesting and thought-provoking) advice, where do we stop throwing away human knowledge in favour of general, bare methods? Shall we abandon expert knowledge? Procedural knowledge? Structural knowledge? Algorithms and data structures? Should the rest of computer science surrender in the face of efficiently cal…
I think this bitter lesson needs to be taken for a several grains of salt. Number one, the progress in a particular AI field tends to go, at first, from custom to more general algorithms, exactly as Professor Richard Sutton described. However, there is a second part to this progress, where, once we "understood" (which we never really do) the new level of general algorithms (say Transformers in NLP), we begin to put b…
A counter example would be showing a number of successful examples in, say, computer vision, where handcrafted features do better than learned features. This is largely not the case. In, say, both NLP and Computer Vision, learned features dominate, even at companies with less compute (they use pretrained models).
(Disclaimer: I work with Rich.)
The Bitter Lesson (2019) - https://news.ycombinator.com/item?id=30889873 - April 2022 (37 comments)
The Bitter Lesson - https://news.ycombinator.com/item?id=28409314 - Sept 2021 (1 comment)
The Bitter Lesson (From AI Research) - https://news.ycombinator.com/item?id=27924335 - July 2021 (1 comment)
The Bitter Lesson (2019) - https://news.ycombinator.com/item?id=23781400 - July 2020 (85 comments)
The Bitter Lesson - https://news.ycombinator.com/item?id=19393432 - March 2019 (53 comments)