logoalt Hacker News

handfuloflightlast Monday at 9:09 PM1 replyview on HN

Then you're missing the point of my rebuttal. You say submarines don't swim [like fish] despite both moving through water, the only distinction is mechanism. Can AI recursively create new capabilities like thinking does, or just execute tasks like submarines do? That's the question.


Replies

gwdyesterday at 11:42 AM

> Can AI recursively create new capabilities like thinking does, or just execute tasks like submarines do? That's the question.

Given my experience with LLMs, I think that they could, but that they're handicapped by certain things at the moment. Haven't you ever met someone who was extremely knowledgable and perceptive at certain tasks, but just couldn't keep on target for 5 minutes? If you can act as a buffer around them, to mitigate their weak points, they can be a really valuable collaborator. And sometimes people like that, if given the right external structure (and sometimes medication), turn out to be really capable in their own right.

Unfortunately it's really difficult to give you a sense of this, without either going into way too much detail, or speaking in generalities. The simpler the example, the less impressive it is.

But here's a simple example anyway. I'm developing a language-learning webapp. There's a menu that allows you to switch between one of the several languages you're working on, which originally just had the language name; "Mandarin", "Japanese", "Ancient Greek". I thought an easy thing to make it nicer would be to have the flag associated with the language -- PRC flag for Mandarin, Japanese flag for Japanese, etc. What do do for Ancient Greek? Well, let me see it looks and then maybe I can figure something out.

So I asked Claude what I wanted. As expected, it put the PRC and Japanese flags for the first two languages. I expected it to just put a modern Greek flag, or a question mark, or some other gibberish. But it put an emoji of a building with classical Greek columns (), which is absolutely perfect.

My language learning system is unusual; so without context, Claud assumes I'm making something like what already exists -- Duolingo or Anki or something. So I invested some time creating a document that lays out in detail. Now when I include that file as a context, Claude seems to genuinely understand what I'm trying to accomplish in a way it didn't before; and often comes up with creative new use cases. For example, at some point I was having it try to summarize some marketing copy for the website; in a section on educational institutions, it added a bullet point for how it could be used that I'd never thought of.

The fact that they can't learn things on-line, that they have context rot, that there's still a high amount of variance in their output -- all of these, it seems to me, undermine their ability to do things, similar to the way some people's ADHD undermines their ability to excel. But it seems to me the spark of thinking and of creativity is there.

EDIT: Apparently HN doesn't like the emojis. Here's a link to the classical building emoji: https://www.compart.com/en/unicode/U+1F3DB