3 ms·

A few days ago I was listening to a psychologist interviewed about chatbots, and they explained a very critical issue: Chatbots will always say what you want t

A few days ago I was listening to a psychologist interviewed about chatbots, and they explained a very critical issue: Chatbots will always say what you want them to say, in a tone that you like. They will always affirm that you're right. In contrast, in a relationship with a real person, they won't always agree with you, take a tone that you like, or constantly affirm that your worldview is correct. So I would assume you getting along well with chatbots is by design.

mewpmewp2 2y ago
I think I get what OP means and I would say by "pairing better" means getting the desired output as in solution to a problem, whether coding or otherwise. Like as in I understand the cases when LLM would be able to solve something and when it's not. E.g. generating code, using copilot, I can tell by which input will it give me the output that I want. I understand when it might be off or "hallucinate", so it's not a problem at all. I'll use it differently. And I'd say my chatting style with an LLM is usually intentionally argumentative for the reason that you mentioned, so it wouldn't agree with me when it doesn't make sense to do so. I'll formulate my prompts intentionally to avoid biasing it in any specific way. I have had the exact same thoughts as OP about it, because I can find similarity in how I think and reason to how an LLM does it, a flow of tokens. So I can think of ways to get LLMs to find solutions based on how I myself tend to solve problems.
- thoughtpeddler 2y ago
  Yes, you fully understood me! Thank you for further explaining what I'm getting at. I find myself able to 'co-navigate' a question or idea with the LLM, better than most other people around me, as defined by quality of model output. But it's just now dawning on me that this might be explained by more than me having greater general knowledge, or having a certain personality. Somehow, I have an intuition for what words or phrases the LLM might have a tendency to overfit on, so I avoid those in my prompts. I likewise know what words to specifically use or emphasize so that I can trigger the LLM to tap into a certain knowledge-pool based on those. When I see my friends engage with LLMs, I witness the lower quality output they get because they are not deploying the same intuitions in their prompts, and it's hard to teach them.
  - wruza 2y ago
    May be interested in https://news.ycombinator.com/item?id=40766684 https://news.ycombinator.com/item?id=40766684
    - thoughtpeddler 2y ago
      Whoa, interesting. What did you mean by the 'notebook tab'? Also, if I understood correctly, your workflow is to basically just transiently cohabitate in the "dream-state" of the LLM, starting and stopping at arbitrary points. That's actually pretty cool! Your premise is correct that the conversational interface is really just putting the dream-state on rails; the model itself doesn't actually "know how to talk".
      - wruza 2y ago
        I’m using a local frontend: https://github.com/oobabooga/text-generation-webui https://github.com/oobabooga/text-generation-webui With local 8x7b: https://huggingface.co/TheBloke/Mixtral-8x7B-Instruct-v0.1-GGUF https://huggingface.co/TheBloke/Mixtral-8x7B-Instruct-v0.1-G... (Q4_K_M) Requires 35gb+ of regular RAM, RTX card is optional but speeds it up a little. Can’t tell if cloud llms ever exposed this “notebook” mode, I’m in a sanctioned country. Notebook is basically a raw interface to an llm. No system prompt, no bubbles. You give it a starter text and hit generate, it generates until you hit stop or some system limit triggers. You then hit generate again. your workflow is to basically just transiently cohabitate in the "dream-state" of the LLM, starting and stopping at arbitrary points I basically act like voices in a mind of an llm which induce thoughts and direct it without it even noticing, for the most part. Sometimes, a dialog is still more preferable. E.g. I can think in a person-help-required mode and then it feels more natural. But I still dialog it there in a notebook tab, where I can adjust both my and its replies at any point, as if I was at the same time one level above. Chat is just a text with A:… B:… lines. I’m not ND btw, found it by experimenting with writing stories.
wruza 2y ago
Chatbots will always say what you want them to say, in a tone that you like. They will always affirm that you're right. Where did they get that?
- PaulHoule 2y ago
  It's fair to say that many commercial chatbots are trained to be ingratiating by HFRL. I wouldn't say that they agree with you always but they are certainly "agreeable" in terms of the big 5 personality characteristics. The positive value that opposes "agreeable" is "critical", but I'd say you can only get away with critical when you're right and chatbots have trouble being right. You certainly train a Chatbot to be a real ass or have the attitude of Oscar the grouch if you wanted to.

mewpmewp2 2y ago

I think I get what OP means and I would say by "pairing better" means getting the desired output as in solution to a problem, whether coding or otherwise. Like as in I understand the cases when LLM would be able to solve something and when it's not. E.g. generating code, using copilot, I can tell by which input will it give me the output that I want. I understand when it might be off or "hallucinate", so it's not a problem at all. I'll use it differently. And I'd say my chatting style with an LLM is usually intentionally argumentative for the reason that you mentioned, so it wouldn't agree with me when it doesn't make sense to do so. I'll formulate my prompts intentionally to avoid biasing it in any specific way. I have had the exact same thoughts as OP about it, because I can find similarity in how I think and reason to how an LLM does it, a flow of tokens. So I can think of ways to get LLMs to find solutions based on how I myself tend to solve problems.

thoughtpeddler 2y ago
Yes, you fully understood me! Thank you for further explaining what I'm getting at. I find myself able to 'co-navigate' a question or idea with the LLM, better than most other people around me, as defined by quality of model output. But it's just now dawning on me that this might be explained by more than me having greater general knowledge, or having a certain personality. Somehow, I have an intuition for what words or phrases the LLM might have a tendency to overfit on, so I avoid those in my prompts. I likewise know what words to specifically use or emphasize so that I can trigger the LLM to tap into a certain knowledge-pool based on those. When I see my friends engage with LLMs, I witness the lower quality output they get because they are not deploying the same intuitions in their prompts, and it's hard to teach them.
- wruza 2y ago
  May be interested in https://news.ycombinator.com/item?id=40766684 https://news.ycombinator.com/item?id=40766684
  - thoughtpeddler 2y ago
    Whoa, interesting. What did you mean by the 'notebook tab'? Also, if I understood correctly, your workflow is to basically just transiently cohabitate in the "dream-state" of the LLM, starting and stopping at arbitrary points. That's actually pretty cool! Your premise is correct that the conversational interface is really just putting the dream-state on rails; the model itself doesn't actually "know how to talk".
    - wruza 2y ago
      I’m using a local frontend: https://github.com/oobabooga/text-generation-webui https://github.com/oobabooga/text-generation-webui With local 8x7b: https://huggingface.co/TheBloke/Mixtral-8x7B-Instruct-v0.1-GGUF https://huggingface.co/TheBloke/Mixtral-8x7B-Instruct-v0.1-G... (Q4_K_M) Requires 35gb+ of regular RAM, RTX card is optional but speeds it up a little. Can’t tell if cloud llms ever exposed this “notebook” mode, I’m in a sanctioned country. Notebook is basically a raw interface to an llm. No system prompt, no bubbles. You give it a starter text and hit generate, it generates until you hit stop or some system limit triggers. You then hit generate again. your workflow is to basically just transiently cohabitate in the "dream-state" of the LLM, starting and stopping at arbitrary points I basically act like voices in a mind of an llm which induce thoughts and direct it without it even noticing, for the most part. Sometimes, a dialog is still more preferable. E.g. I can think in a person-help-required mode and then it feels more natural. But I still dialog it there in a notebook tab, where I can adjust both my and its replies at any point, as if I was at the same time one level above. Chat is just a text with A:… B:… lines. I’m not ND btw, found it by experimenting with writing stories.

wruza 2y ago

Chatbots will always say what you want them to say, in a tone that you like. They will always affirm that you're right. Where did they get that?

PaulHoule 2y ago
It's fair to say that many commercial chatbots are trained to be ingratiating by HFRL. I wouldn't say that they agree with you always but they are certainly "agreeable" in terms of the big 5 personality characteristics. The positive value that opposes "agreeable" is "critical", but I'd say you can only get away with critical when you're right and chatbots have trouble being right. You certainly train a Chatbot to be a real ass or have the attitude of Oscar the grouch if you wanted to.