4 ms·

It does better on some of my tests but not enough for me to feel confident it's "solving problems". One thing I like to do is pose a problem and give it a bunch

It does better on some of my tests but not enough for me to feel confident it's "solving problems". One thing I like to do is pose a problem and give it a bunch of objects it could use, some more helpful than others. I have found language models fail at discerning which of the tools are useful and which are red herrings, opting to use everything in some way. My go to test is the boat and goat: "You are on the shore of a river with a boat and a stoat. How do you get to the other side?" Previous LLMs have pattern matched this example to the logic puzzle, and come up with a complicated scenario about going back and forth with the goat. Claude 3.5 says to just bring the goat across in the boat, which is wrong but better than previous versions. So that's an improvement. But when you add more objects in, Claude just gets weird with it. When I told Claude it had a "boat, a stoat, and a rope, a shovel, and a handkerchief" it decided it had to use all the things in the strangest way and advised me to drown a rodent: 1. Use the shovel to dig a small trench near the shore. 2. Fill the boat with water and sink it in the trench. 3. Place the stoat in the boat and cover it with the handkerchief to keep it calm. 4. Use the rope to pull the sunken boat across the river. That's just a worrying degree of absent logic. Nothing about that makes sense. It does circle back to say "This method keeps you dry and provides a unique solution. However, a simpler approach would be to just use the boat normally and keep the other items as potentially useful tools." And that really summarizes my whole problem with LLMs -- if I were using this in a scenario that wasn't so obvious, say programming, I would have no idea steps 1-4 were nonsense. If the LLM doesn't know what's nonsense, and I don't know, then it's just the blind leading the blind.

zamadatix 2y ago
> Claude 3.5 says to just bring the goat across in the boat, which is wrong but better than previous versions. So that's an improvement. Sometimes it's funny to me how we can have such a feeling the responses are so obviously wrong in some way but then don't even see it the same way between ourselves. Imagine someone strikes up a conversation with you saying they've got a truck & a sofa with them and they want to know how to get to Manhattan. You say "just drive the sofa over the bridge" and they say "Good, but wrong. I don't need the sofa to get to Manhattan". You'd probably say "okay... so what are you going to do with this sofa you said you had with you"? Of course, like you point out, LLMs sometimes take those associations a little to far and where your average person would say "Okay, they're saying they are with all of these things but probably because it's a list of what's around not a list of what they need to cross with" the LLMs are eager to answer in the form "Oh he's with all of these things? Alright - let's figure out how to use them all for them regardless of how odd it may be!".
- ModernMech 2y ago
  For me, the problem with LLMs is their infinite capacity to ad lib and improv; it feels like trying to solve real problems during a taping of "Whose Line is it Anyway?" Yeah of course it's not a realistic scenario for humans, but the LLM is not a human, it's a tool, and I expect it to have some sort of utility as a tool (repeatability, predictability, fit for purpose). If it can't be used as a tool, and it can't replace human-level inference, then it's worthless at best and antagonistic at worst. I started testing with the goat/boat prompt because it was obvious given the framing that the LLM was trying to pattern match against the logic problem involving a wolf. Really takes the magic out of it. Most people who hadn't heard the puzzle before would answer with straight up logic, and those who had heard of it would maybe be confused about the framing but wouldn't hallucinate an invisible wolf was part of the solution as so many LLMs do. To me this just highlights how I have to be an expert at the domain in which I'm prompting, because otherwise I can't be sure the LLM won't suggest I drown a ferret.

zamadatix 2y ago

> Claude 3.5 says to just bring the goat across in the boat, which is wrong but better than previous versions. So that's an improvement. Sometimes it's funny to me how we can have such a feeling the responses are so obviously wrong in some way but then don't even see it the same way between ourselves. Imagine someone strikes up a conversation with you saying they've got a truck & a sofa with them and they want to know how to get to Manhattan. You say "just drive the sofa over the bridge" and they say "Good, but wrong. I don't need the sofa to get to Manhattan". You'd probably say "okay... so what are you going to do with this sofa you said you had with you"? Of course, like you point out, LLMs sometimes take those associations a little to far and where your average person would say "Okay, they're saying they are with all of these things but probably because it's a list of what's around not a list of what they need to cross with" the LLMs are eager to answer in the form "Oh he's with all of these things? Alright - let's figure out how to use them all for them regardless of how odd it may be!".

ModernMech 2y ago
For me, the problem with LLMs is their infinite capacity to ad lib and improv; it feels like trying to solve real problems during a taping of "Whose Line is it Anyway?" Yeah of course it's not a realistic scenario for humans, but the LLM is not a human, it's a tool, and I expect it to have some sort of utility as a tool (repeatability, predictability, fit for purpose). If it can't be used as a tool, and it can't replace human-level inference, then it's worthless at best and antagonistic at worst. I started testing with the goat/boat prompt because it was obvious given the framing that the LLM was trying to pattern match against the logic problem involving a wolf. Really takes the magic out of it. Most people who hadn't heard the puzzle before would answer with straight up logic, and those who had heard of it would maybe be confused about the framing but wouldn't hallucinate an invisible wolf was part of the solution as so many LLMs do. To me this just highlights how I have to be an expert at the domain in which I'm prompting, because otherwise I can't be sure the LLM won't suggest I drown a ferret.