4 ms·
He actually says the areas in which AI has had the novel successes are those which can be evaluated (like coding or Go). Not that it can’t happen at all.
by zeroonetwothree 4mo ago
He actually says the areas in which AI has had the novel successes are those which can be evaluated (like coding or Go). Not that it can’t happen at all.
- simianwords 4mo agoThat’s my point, he says ai does well where evaluation is neurosymbolically closed. But so do humans? How do humans make discoveries without having formal ways to evaluate? In my pharma drug example, humans could evaluate only because they had access to the physical realm. I can’t think of an example of humans evaluating a discovery in a way that LLMs can’t. can you?
- arowthway 4mo agoI don't think there is any "humans are metaphysically superior to LLMs" subtext to this talk, it's just a technical/educational observation. Access to some forms of evaluation and selective retention is inherent to humans and it's not inherent to LLMS. But it can be somehow bolted on and that's when they work best. It makes sense that more focus on those principles can yield better AI. I think the retention part is the real limitation of LLMs, because it's limited to stuffing things in context window.
- simianwords 4mo ago> Access to some forms of evaluation and selective retention is inherent to humans and it's not inherent to LLMS I'm not sure I understood - what forms of evaluation is inherent to humans? If you don't give humans tools or access to the physical world, how can they evaluate?
- arowthway 4mo agoThere's no such thing as a human without access to the physical world.
- simianwords 4mo agoSo technically the only reason AI can't do discovery is access to physical word. When you give AI and humans access, they both do discoveries - that is the clean summary of the author's position. Its not too interesting.. we already know that giving AI access to compilers and tools make them better.
- MortalMan 4mo ago[dead]