4 ms·

On the tool-invocation point: Something that seems true to me is that LLMs are actually too smart to be good tool-invokers. It may be possible to convince them

On the tool-invocation point: Something that seems true to me is that LLMs are actually too smart to be good tool-invokers. It may be possible to convince them to invoke a purpose-specific tool rather than trying to do it themselves, but it feels harder than it should be, and weird to be limiting capability. My thought is: Could the tool-routing layer be a much simpler "old school" NLP model? Then it would never try to do math and end up doing it poorly, because it just doesn't know how to do that. But you could give it a calculator tool and teach it how to pass queries along to that tool. And you could also give it a "send this to a people LLM tool" for anything that doesn't have another more targeted tool registered. Is anyone doing it this way?

dvt 2y ago
> Is anyone doing it this way? I'm working on a way of invoking tools mid-tokenizer-stream, which is kind of cool. So for example, the LLM says something like (simplified example) "(lots of thinking)... 1+2=" and then there's a parser (maybe regex, maybe LR, maybe LL(1), etc.) that sees that this is a "math-y thing" and automagically goes to the CALC tool which calculates "3", sticks it in the stream, so the current head is "(lots of thinking)... 1+2=3 " and then the LLM can continue with its thought process.
- namaria 2y ago
  Cold winds are blowing when people look at LLMs and think "maybe an expert system on top of that?".
  - sanderjd 2y ago
    I don't think it's "on top"? I think it's an expert system where (at least) one of the experts is an LLM, but it doesn't have to be LLMs from bottom to top.
    - namaria 2y ago
      On the side, under, wherever. The point is, this is just re-inventing past failed attempts at AI.
      - sanderjd 2y ago
        Except past attempts didn't have the ability to pass on to modern foundation models. Look, I dunno if this idea makes sense, it's why I posed it as a question rather than a conviction. But I broadly have a sense that when a new technology hits, people are like "let's use it for everything!", and then as it matures, people find more success in interesting it with current approaches, or even trying older ideas but within the context of the new technology. And it just strikes me that this "routing to tools" thing looks a lot like the part of expert systems that did work pretty well. But now we have the capability to make those tools themselves significantly smarter.
        namaria 2y ago
        Expert systems are not the problem per se. The problem is that AI is very often a way of hyping software. "This is a smart product. It is intelligent". It implies lightning in a bottle, a silver bullet. A new things that solves all your problems. But that is never true. To create useful new stuff, to innovate, in a word, we need domain expertise and a lot of work. The world is full of complex systems and there are no short cuts. Well, there are, but there is always a trade off. You can pass it on (externalities) or you can hide (dishonesty) or you can use a sleight of hand and pretend the upside is so good, it's magical so just don't think about what it costs, ok? But it always costs something. The promise of "expert systems" back then was creating "AI". It didn't happen. And there was an "AI winter" because people wised up to that shtick. But then "big data" and "machine learning" collided in a big way. Transformers, "attention is all you need" and then ChatGPT. People got this warm fuzzy feeling inside. These chatbots got impressive, and improved fast! It was quite amazing. It got A LOT of attention and has been driving a lot of investment. It's everywhere now, but it's becoming clear it is falling very short of "AI" once again. The promised land turned out once again to just be someone else's land. So when people look at this attempt at AI and its limitations, and start wondering "hey what if we did X" and X sounds just like what people were trying when we last thought AI might just be around the corner... Well let's just say I am having a deja vu.
- sanderjd 2y ago
  Definitely an interesting thought to do this at the tokenizer level!

dvt 2y ago

> Is anyone doing it this way? I'm working on a way of invoking tools mid-tokenizer-stream, which is kind of cool. So for example, the LLM says something like (simplified example) "(lots of thinking)... 1+2=" and then there's a parser (maybe regex, maybe LR, maybe LL(1), etc.) that sees that this is a "math-y thing" and automagically goes to the CALC tool which calculates "3", sticks it in the stream, so the current head is "(lots of thinking)... 1+2=3 " and then the LLM can continue with its thought process.

namaria 2y ago
Cold winds are blowing when people look at LLMs and think "maybe an expert system on top of that?".
- sanderjd 2y ago
  I don't think it's "on top"? I think it's an expert system where (at least) one of the experts is an LLM, but it doesn't have to be LLMs from bottom to top.
  - namaria 2y ago
    On the side, under, wherever. The point is, this is just re-inventing past failed attempts at AI.
    - sanderjd 2y ago
      Except past attempts didn't have the ability to pass on to modern foundation models. Look, I dunno if this idea makes sense, it's why I posed it as a question rather than a conviction. But I broadly have a sense that when a new technology hits, people are like "let's use it for everything!", and then as it matures, people find more success in interesting it with current approaches, or even trying older ideas but within the context of the new technology. And it just strikes me that this "routing to tools" thing looks a lot like the part of expert systems that did work pretty well. But now we have the capability to make those tools themselves significantly smarter.
      - namaria 2y ago
        Expert systems are not the problem per se. The problem is that AI is very often a way of hyping software. "This is a smart product. It is intelligent". It implies lightning in a bottle, a silver bullet. A new things that solves all your problems. But that is never true. To create useful new stuff, to innovate, in a word, we need domain expertise and a lot of work. The world is full of complex systems and there are no short cuts. Well, there are, but there is always a trade off. You can pass it on (externalities) or you can hide (dishonesty) or you can use a sleight of hand and pretend the upside is so good, it's magical so just don't think about what it costs, ok? But it always costs something. The promise of "expert systems" back then was creating "AI". It didn't happen. And there was an "AI winter" because people wised up to that shtick. But then "big data" and "machine learning" collided in a big way. Transformers, "attention is all you need" and then ChatGPT. People got this warm fuzzy feeling inside. These chatbots got impressive, and improved fast! It was quite amazing. It got A LOT of attention and has been driving a lot of investment. It's everywhere now, but it's becoming clear it is falling very short of "AI" once again. The promised land turned out once again to just be someone else's land. So when people look at this attempt at AI and its limitations, and start wondering "hey what if we did X" and X sounds just like what people were trying when we last thought AI might just be around the corner... Well let's just say I am having a deja vu.
sanderjd 2y ago
Definitely an interesting thought to do this at the tokenizer level!