31–35 of 35 posts
Natural selection is a massive amount of computation. It's not like we got to skip that. Any learning that we do without computation is based on structures that were computed by evolution. One way or another the "learning" process is happening somewhere.
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…
At the same time, will we ever get General AI just by throwing more data into the models?
Our biological brains definitely have way more data than the models we have right now but they also have different mechanisms which we barely understand.
I'm not confident we'll get human-like behaviour just by throwing infinite amount of data into a model, I think we will need significant changes to how we do NN as well.
Or maybe we just grow a brain in the lab, wire some inputs and outputs, train it for 20 years and then treat is a CPU.
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…
Apparently GPT-3 trained on code gained the ability for reasoning, but plain language, or even multi-task fine tuning, were not enough. Maybe the quality of the data makes a difference in the robustness of the model. 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 archi…
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…
Apparently GPT-3 is not capable of multi-step reasoning unless trained on code. So it seems having code in the training set generalises an ability to reason in natural language.