this post was submitted on 23 Nov 2023
184 points (91.8% liked)
Technology
59440 readers
3572 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Being a layperson in this, I’d imagine part of the promise is that once you’ve got reliable arithmetic, you can get logic and maths in there too and so get the LLM to actually do more computer-y stuff but with the whole LLM/ChatGPT wrapped around it as the interface.
That would mean more functionality, perhaps a lot more of it works and scales, but also, perhaps more control and predictability and logical constraints. I can see how the development would get some excited. It seems like a categorical improvement.
If that's the case then bad news for OpenAI's "moat" (and for people arguing for restraint in general): there's been some recent breakthroughs in getting open-source LLMs trained to understand math as well.
It'd be hilarious if OpenAI's board went through huge turmoil, tanked tens of billions of dollars worth of investments, disrupted their partnership with Microsoft to protect this huge revolution they've got brewing in their most secret and secure of laboratories... and then someone posts "hey, I got my AI Waifu to count good, check out this github to see how I did it" on Reddit.
It also brings into question (well, it adds to the questions, they were already brought up) the whole premise of IP law that "if we don't protect it properly, no one will want to invent things". It seems to me like people like creating things and humanity has a strange habit of converging on new inventions from multiple directions. Kinda like how calculus was invented by two different people at the same time.
Always wondered why the text model didn't just put its output through something like MATLAB or Mathematica once it got as far as having something which requires domain-specific tools.
Like when Prof. Moriarty tried it on a quantum physics question and it got as far as writing out the correct formula before failing to actually calculate the result
There is definitely a lot of effort in this direction, seems very likely that a hybrid system could be very powerful.
More or less, yeah. It's been possible to do math with llm's by hooking in an interpreter, but it's a pain in the ass and in my limited experience required tons of custom prompting. Now, hooking in an interpreter and having it do math directly on the processor instead of using the neural net architecture is going to be orders of magnitude faster and more power-efficient, but also, with rudimentary reasoning comes the possibility of improving an llm's ability to hook itself into external tools without lots of help from the user.