4 ms·

As much of a curmudgeon as I am on AI I do sincerely believe that one area that it's effective at is going from absolute beginner to decent understanding on som

As much of a curmudgeon as I am on AI I do sincerely believe that one area that it's effective at is going from absolute beginner to decent understanding on something completely new and foreign. Recently I've been somewhat curious about Skia, the graphics library that Google uses. A while back I fielded a few questions about Skia to Claude. Nothing crazy, just questions on how to draw a few primitive shapes, but I felt it would have taken some effort to find the answers on my own. And I must say I was pretty satisfied with what it gave me back.

lxgr 2y ago
It is indeed incredibly useful for getting from zero to an undergrad level of understanding of most subjects. In my personal experience, it's also been largely useless beyond that. That's not to diminish the potential impact of having an almost-free grad-student-level tutor available 24/7! Even if AI stops improving here, this alone will have a huge impact on future science (there being more trained people around capable of helping solve hard problems and all). But we're also definitely some way away from AIs doing their own research, and I'm not sure if scaling the current architecture alone will get us there.
- throwup238 2y ago
  > But we're also definitely some way away from AIs doing their own research, and I'm not sure if scaling the current architecture alone will get us there. Hot take: I think it's fairly obvious the current architecture is not enough BUT it's already or almost at the point where it can be the backbone for a proper AGI/ASI system. Linking human communication and some light language-based reasoning to other AI models that handle the rest of the world model and reasoning, using special tokens to coordinate them. I think it's just really early and most of this stuff hasn't even been attempted yet. There's tools/function calling which is kind of a proof of concept, and a lot of academic labs researching transformers applied to other fields like robotic control and learning, but no one's put it all together yet.
  - lxgr 2y ago
    What's critically missing in the current architectures, in my view, is models that are able to "keep learning" (in a way beyond just providing a summary of the previous conversation, which grants surface-level insights at most). If we keep the current iteration cycle (retrain an LLM every couple of months, incorporating the current "status quo", including all of its previous conversations), we might get somewhat interesting results, but even the least motivated grad student has an iteration cycle orders of magnitude faster than that.
devjab 2y ago
As an expert who uses LLMs daily on my field of expertise I’m not looking forward to dealing with works created by people who were taught by them. They are wrong far too often to be used for any sort of learning in my opinion. You can feed them a book and they’ll give you answers which don’t exist in any of the written material, and that’s the good part of them. Simply asking a LLM about things will give you answers that are hopelessly wrong and unless you’re an expert you won’t know it. Which isn’t necessarily worse than learning things from search engines or YouTube. I recently had ChatGPT give me a recipe for bread which was certainly better than the top 10 results on Google except for one things. The recipe listed 3x the amount of salt which you should ever reasonably use. I asked it a few other times, I even asked it for my friend Tommy’s recipe and all those answers were perfect. It obviously doesn’t know Tommy but it pretended to know and just gave me a pretty basic recipe. Salt is a harmless error. Most people would know not to use that much salt, and even if they did, it wouldn’t harm them (much). But imagine if you had used it for something electric or chemistry.

lxgr 2y ago

It is indeed incredibly useful for getting from zero to an undergrad level of understanding of most subjects. In my personal experience, it's also been largely useless beyond that. That's not to diminish the potential impact of having an almost-free grad-student-level tutor available 24/7! Even if AI stops improving here, this alone will have a huge impact on future science (there being more trained people around capable of helping solve hard problems and all). But we're also definitely some way away from AIs doing their own research, and I'm not sure if scaling the current architecture alone will get us there.

throwup238 2y ago
> But we're also definitely some way away from AIs doing their own research, and I'm not sure if scaling the current architecture alone will get us there. Hot take: I think it's fairly obvious the current architecture is not enough BUT it's already or almost at the point where it can be the backbone for a proper AGI/ASI system. Linking human communication and some light language-based reasoning to other AI models that handle the rest of the world model and reasoning, using special tokens to coordinate them. I think it's just really early and most of this stuff hasn't even been attempted yet. There's tools/function calling which is kind of a proof of concept, and a lot of academic labs researching transformers applied to other fields like robotic control and learning, but no one's put it all together yet.
- lxgr 2y ago
  What's critically missing in the current architectures, in my view, is models that are able to "keep learning" (in a way beyond just providing a summary of the previous conversation, which grants surface-level insights at most). If we keep the current iteration cycle (retrain an LLM every couple of months, incorporating the current "status quo", including all of its previous conversations), we might get somewhat interesting results, but even the least motivated grad student has an iteration cycle orders of magnitude faster than that.

devjab 2y ago

As an expert who uses LLMs daily on my field of expertise I’m not looking forward to dealing with works created by people who were taught by them. They are wrong far too often to be used for any sort of learning in my opinion. You can feed them a book and they’ll give you answers which don’t exist in any of the written material, and that’s the good part of them. Simply asking a LLM about things will give you answers that are hopelessly wrong and unless you’re an expert you won’t know it. Which isn’t necessarily worse than learning things from search engines or YouTube. I recently had ChatGPT give me a recipe for bread which was certainly better than the top 10 results on Google except for one things. The recipe listed 3x the amount of salt which you should ever reasonably use. I asked it a few other times, I even asked it for my friend Tommy’s recipe and all those answers were perfect. It obviously doesn’t know Tommy but it pretended to know and just gave me a pretty basic recipe. Salt is a harmless error. Most people would know not to use that much salt, and even if they did, it wouldn’t harm them (much). But imagine if you had used it for something electric or chemistry.