1 Answers sorted by
Mar 18, 2024
72Your belief that "a good explainer should be able to convince a no-experience person to abandon their initial assumptions" is quite false in practice and you should abandon this initial assumption. Good communication is hard, communication against an already formed opinion is much harder still, and doing so in 3 short lines of conversation is verging on impossible.
It may or may not be true that Alice is providing a good argument for T1, and Bob may or may not have a valid point about A but there's no way to know since all the relevant information about that has been stripped from the scenario. It's certainly not true that in general Alice has made a poor argument.
If this sort of exchange happened in a real conversation (and I have seen some similar cases in practice), I would interpret the last line to mean that Alice is writing off helping Bob as a waste of time since any further communication is likely to result in ongoing argument and no useful progress (see first paragraph).
It is very unlikely that Bob has tried T1, since he gave a very weak theoretical argument in favour of using T2 instead, rather than a much stronger and practical argument "I tried T1, and here's why it didn't help". Likewise he claims to know about T2 and believes it to be better than T1 but fairly clearly hasn't tried that one either. Why not?
2Thanks for your reply. It helps me refine my thinking. I have come up these questions, hope you can help me answering them:
- Is it correct that when Alice says "You should try it first. You are reasoning too much", she is no longer giving reasons on why T₁ is better?
- Is it correct that the experience state of Bob isn't relevant to the reasons why T₁ is better?
- Is it correct that it's possible for one to get closer to the truth via reasoning and conversing with people with direct experience, while not having direct experience themself? After all there are theor
This seems WAY over-abstracted. There are important differences in what kinds of evidence is obtainable by what techniques for problems in very different domains.
Also, this seems unnecessarily adversarial between Bob and Alice. Have they not agreed on the problem or on what would constitute a solution? If they can reframe to a shared seeking of knowledge, it may be easier to actually talk about what each believes and why.
Do you mean that how the evidence is obtained in different problems and domains determines whether saying "You should try it first. You are reasoning too much" is still giving reasons on why T₁ is better? Can you elaborate or give examples?
Have they not agreed on the problem or on what would constitute a solution?
I suppose it's because Bob isn't aware of all the things he needs to say before posting the question, and Alice assumes on what he needs while he thinks he doesn't need it.
Yes, different "it" will have VASTLY different costs and potential evidence from trying, so the discussion doesn't generalize very well. "you are reasoning too much" implies "you are empirically testing too little", which could easily be true or false, or neither (it could be "you are reasoning badly from evidence we agree on" or "you need to BOTH measure and reason a lot more clearly").
For some (but not all) topics, "just try it" is INCREDIBLY unhelpful - most people are pretty bad observers, and a lot of things don't separate experiential elements in ways that are easy to analyze which parts are evidence, which parts are idiosyncratic triggers of internal states.
I suppose it's because Bob isn't aware of all the things he needs to say before posting the question, and Alice assumes on what he needs while he thinks he doesn't need it.
Without actual specifics, it's hard to know WHY the disconnect is happening. It does seem that Alice and Bob aren't in agreement over what the question is, but it's unclear which (if either) is closer to something useful.
I can of course provide full detailed, but in the interest of keeping it as general as possible (or, to understand where further generalization is not possible), I want to get deeper on this before provide more specifics.
I agree that "you are reasoning badly from evidence we agree on" or "you need to both measure and reason a lot more clearly" are also reasonable interpretations of "you are reasoning too much". But isn't that the other sentence "You should try it first" makes the "you are empirically testing too little" interpretation more likely?
Can you simplify "idiosyncratic triggers of internal states"? Also, if most people are bad observers, then wouldn't that it's more helpful for them to have direct experience with it?
Can you simplify "idiosyncratic triggers of internal states"? Also, if most people are bad observers, then wouldn't that it's more helpful for them to have direct experience with it?
It matters a lot what "it" is. Common targets of "just try it" are mystic or semi-mystic experiences around drugs, meditation, religion, etc., These tend to be hard to communicate because they're not actually evidence of outside/objective phenomena, they're evidence of an individual's reaction to something. I have no clue whether that applies here or not - that's my primary point: one size does not fit all.
Note that Bob is making an error if he flatly denies Alice's experiences, rather than acknowledging that the experiences can be real without the underlying model being correct.
I see your point. I think it's time to tell the specifics.
- [Bob posts a problem about data classification]
- Alice: You should use LLM. It especially suites your problem
- Bob: In my understanding LLM is only strong where the context is large. If the context is small then using regex gives better result? Also, regex has advantages of high accuracy, fast, understandable and debuggable?
- Alice: Not really. Also, regex cannot work with synonyms and it must be in form. LLM is trained on multiple data, so if you make good prompt then it's much better
- Bob: But the nature of catching synonyms is still depending on context. As the context is small then there is not much synonyms at the beginning. If even human cannot get them then how can machine recognize them?
- Alice: You should try it first. You are reasoning too much
There are some notes:
- By "regex" Bob actually means rule-based approach. He thought in the context of NLP people generally understand regex and rule-based approach as one
- He mistakes synonym with homonym. Had he been aware of that he might have not said "the nature of catching synonyms is still depending on context"
These info are only revealed later on. I understand that at the time of saying these are enough for Alice to conclude that Bob needs to learn more. And again, I understand that she doesn't want to waste time. But I think shutting down a conversation because of the experience state of the conversations partner is a sign of ad hominent. If she wants to save her time then she can just abandon the conversation. If she still want to give feedback while not having the burden to elaborate then she can tell Bob to read more about synonyms/context/LLM. Is that a correct thinking?
Oh, both of them are stuck in a false dichotomy. The right answer is obviously a mix of both. They could have a valid debate about budgets (both dev and runtime costs) and precision-recall tradeoffs, and long-tail utterances, and whether the problem is simple enough that it's anywhere close to true that the rules will fit in someone's head in order to be testable and debuggable. They can talk about what parts of the problem are best addressed by what tools. But a general discussion of "if we have only one tool, should it be a hammer or a screwdriver" is probably unhelpful.
"Just try it" is probably fairly costly to do, so I understand Bob's reluctance. He's wrong, because "try it" includes "learn enough about it to actually understand it's strengths and weaknesses". Alice is wrong because "you're reasoning too much" is not what's happening, and trying it is not an alternative to reasoning, it's an aid to reasoning better.
I don't find much use in trying to analyze whether Alice is trying to shut down the conversation, or whether Bob is confusing things that are basic enough that Alice is understandably frustrated and unable to give good advice on a practical level. Again, specifics of the discussion matter - is this a friendly, low-stakes brainstorm, or are they teammates trying to solve a specific problem, or is it some other relationship?
That said, I'm now shutting down this conversation :) I'll read any further responses, but probably won't further engage.
That said, I'm now shutting down this conversation :) I'll read any further responses, but probably won't further engage.
Thanks for your patience. Your response is by far the most useful one I have. I really appreciate that.
If you find replying a short sentence isn't much burden, can you tell why you don't want to continue this conversation? Is it because continuing it doesn't stimulate your interest as you have on it in the beginning? Can you share why?
I don't know that my reasons are all that legible, but some combination of:
- I've now heard enough details that my curiosity about the framing of the discussion is satisfied.
- I actually know quite a bit about the object-level discussion (having worked on large NLP projects, though before the dominance of LLMs), and I don't think there's much value to you or to me in trying to go deeper on a problem I'm not invested in.
- My estimate of future enjoyment or education from further exploration of the topic is pretty low. I got good value from a small time/thought investment in early conversation, and it feels like we've passed the peak of marginal enjoyment for additional effort.
- I'm still a little annoyed at the incorrect and wasteful generalization at the start, when the specifics dominate the interpretation that should be used.
Ok, that sounds a lot more harsh than I actually feel. I probably should have said "I think we've explored the overall question about Alice and Bob, and that's all I was looking for in the first place."
I see. Thanks for replying. I wonder if we move the topic to how your interest works, would that be your interest?
They could have a valid debate about budgets (both dev and runtime costs) and precision-recall tradeoffs, and whether the problem is simple enough that it's anywhere close to true that the rules will fit in someone's head in order to be testable and debuggable.
I'm not sure what "it's anywhere close to true that the rules will fit in someone's head in order to be testable and debuggable" means. I guess you just mean if the problem is simple enough then the rule-based approach is better. Is that correct?
I'll assuming yes here. While Bob indeed doesn't mention explicitly about budgets, I guess we can all assume that everyone wants budgets as low as possible. And since Bob thinks that since the context is small here, I think he thinks that the problem is simple enough, and if using the rule-based approach then the precision-recall tradeoffs won't bee applied here. Overall, I think that in his mind the rule-based approach is the perfect tool for the problem.
They can talk about what parts of the problem are best addressed by what tools. But a general discussion of "if we have only one tool, should it be a hammer or a screwdriver" is probably unhelpful.
So while I really agree with this, I suppose the reason both sides stuck is because both assume that their tool has superiority over the other on the problem. Alice sees the problem and concludes that this must be a nail and can never be a screw, so obviously the screwdriver is bad. Bob sees the problem and concludes that this must be a screw and can never be a nail, so obviously the hammer is bad.
So I guess instead of arguing whether the solution should be a screwdriver or a hammer, they should argue on whether the problem is a screw or a nail. Perhaps it's a combination of screw and nail. Bob does give more detail about the problem when he says "the context is small", and I guess in his mind it's a distinctive feature of a screw, but perhaps the reality is that it's a shared feature of both. He sees the shank and mistakes it with the helical thread.
I hope this make sense.
Context
Imagine a conversation:
Reasoning
On one hand, I understand the rationale of Alice:
But still, isn't that Alice doesn't focus on the important topic: giving reasons to use T₁ in this problem? She derails on the experience of Bob, which is not relevant to the argument. It's possible that Bob has already tried T₁ but fails to see its strengths on the problem, or consulted other experts and documentations to see that his assumptions haven't been challenged successfully. Regardless of the experience state, it doesn't contribute to the reasons to use T₁. I think a good explainer should be able to convince a no-experience person to abandon their initial assumptions. Of course Alice is not obligated to continue the conversation, but if she decides to continue, then she should stick to the real topic.
Questions
Therefore, I think these questions would help me validate the assumptions in the reasoning part above: