[perrygeo]

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 other words, computer programs should stop trying be something they are not. They are not AI. Computers are expensive machines that can (very efficiently, and with economies of scale) calculate and present anything that the author of the software desires. It takes actual human intelligence, economics, and ethics to translate that into action.

[nonrandomstring]

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 calculated matrices?

[visarga]

> Shall we abandon expert knowledge? Procedural knowledge?

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.

[lucidrains]

well, you are assuming these efficient matrices cannot become superhuman teachers that show us the inner workings of the universe in a more optimal way.

[visarga]

I think it says something else - when you can marry search with learning, you can surpass anything else, of course paying the price of compute cost.

[manmal]

Can’t a program that searches and learns not become an AI? I think there’s a leap in your argument that you haven’t documented.