Lex Clips
1 HN recommender · recommended in 9 months, 2022–2025
18:27 · 2.5M views · Nov 26, 2019
An excerpt of Richard Feynman's 1985 lecture Q&A about whether machines can think. He discusses the difficulty of defining what computers can never do and compares machine intelligence to machines that don't copy animals.
HN uses it as an early authoritative take in debates about AI, reasoning, and AGI.
10 comments from Hacker News
Relevant — decades earlier, Righard Feynman discussing AI like a boss. He discusses exactly the problems being solved in the article with the Haar approach, but from his time. He mentioned Lenat, who was the OG and who the “LeNet” by Yann LeCun was named after probably (it is mentioned in the article). this video
Well, you have to define what you mean by "intelligence". I think it's just not been enough time, we can take the current LLM technology and just put it in a pipeline that includes 24/7 checking work and building up knowledge bases. A lot of the stuff that you think is "new and original ideas" are just like prompting an LLM to "come up with 20 original variations" or "20 original ways to combine" some building blocks it already has been trained on or have been added into its context. If you do this frequently enough, and make sure to run acceptance tests (e.g. unit testing or whatever is in your domain) then you can really get quite far. In fact, you can generate the tests themselves as well. What's missing, essentially, is autonomous incremental improvements, involving acceptance testing and curation, not just generation. Just like a GAN does when it generates novel images. "Exceptional pattern matching does not imply understanding." - You'll have to define what you mean by "understanding". I think we have to revisit the Chinese Room argument by John Searle. After all, if the book used by the person in the room is the result of training on Chinese, then "the whole Chinese room" with the book and operator may be said to "understand" Chinese. It's not just pattern matching but emergent structures in the model, that is a non-von-neumann architecture, when it's being trained. Those structures are able to manipulate symbols in ways that are extremely useful and practical for an e
Sure, but intent is a very fickle thing. Consider zero and single click deployments in IT operations. With single click deployments, you need to have everything automated, but the go sign is still given by a human. With zero click, you'll have a deployment policy instead - the human decision is now out of the critical path completely, and only plays part during the authoring and later editing of said policy. And you can also then generate those policies, and so on. Same can be applied to AI. You can have canned prompts that you keep refining to encode your intent and preferences, then you just use them with a single click. But you can also build a harness that generates prompts and continuously monitors trends and the world as a whole for any kind of arbitrary criteria (potentially of its own random or even shifting choice), and then follows that: a reward policy. And then like with regular IT, you can keep layering onto that. Because of this, I don't think that intent is the point of differentiation necessarily, but the experience and shared understanding of human intent . That people have varying individual, arbitrary preferences, and are going through life of arbitrary and endless differences, and then source from those to then create. Indeed, this is never going to be replicated, exactly because of what I said: this is humans being human, and that giving them a unique, inalienable position by definition. It's like if instead of planes we called aircraft "mechanical birds"
> It's very difficult to find some way of defining rather precisely something we can do that we can say a computer will never be able to do. There are some things that people make up that say that, "While it's doing it, will it feel good?" or, "While it's doing it, will it understand what it's doing?" or some other abstraction. I rather feel that these are things like, "While it's doing it, will it be able to scratch the lice out of it's hair?" No, it hasn't got any hair nor lice to scratch from it, okay? > You've got to be careful when you say what the human does, if you add to the actual result of his effort some other things that you like, the appreciation of the aesthetic... then it gets harder and harder for the computer to do it because the human beings have a tendency to try to make sure that they can do something that no machine can do. Somehow it doesn't bother them anymore, it must have bothered them in earlier times, that machines are stronger physically than they are... - Feynman this video
Nah. I don’t “feel the AGI”. I think the AGI is a silly quest, just like having a plane flap its wings. Feynman had it right in the 80s: this video I think the future is lots of incremental improvements that get replicated everywhere and humans outclassed in nearly every field, where they stop relying on each other. As far as LLMs yes I think they are the best placed to know if some code or invention is novel, because of their vast training. Can be far better than a patent examiner, if trained on prior art, for instance. What you’re not used to is an LLM being fed stuff that you statistically / heuristically would expect to be average but is in fact the polished result of years of work. The LLM freaks out, you get surprised. You think it was the prompts. The prompts are changed, the END result is the same (scroll to the bottom). I want to see whether foundational LLMs can be used as a good first filter for dealflow and evaluating actual projects.