10 ms·

Role play with large language models

92 points by m_kos 3y ago

VagabundoP 3y ago
I've tried to play dungeons and dragons with Chatai. It doesn't like to use any mechanical rules and keeps trying to wrap everything up like its got a train to catch. I'm sure you could tune it to have more attention than a two year old and actually apply the rules. It did come up with a decent setting based on B2: Keep on the Borderlands. And it was some fun. The new generation of text games will be awesome. And Enders tablet game is possible with this tech.
- xeyownt 3y ago
  You didn't read the article, did you? :-D
- TeMPOraL 3y ago
  > And Enders tablet game is possible with this tech. Complete with aliens subtly messing with it, and through it, the player? They who control the logits, control the future.
- gwern 3y ago
  > I've tried to play dungeons and dragons with Chatai. It doesn't like to use any mechanical rules and keeps trying to wrap everything up like its got a train to catch. Yes, that's the RLHF and other mechanisms. Raters presumably reward completions which are, well, complete and can be quickly judged as a whole, no matter what the intrinsic quality might be or how valid it would be to end with a 'Tune in next week for chapter 2'. What OP is describing is the emergent behavior of the base model, which acts very differently from the RLHF/instruction-tuned/who-knows-what-else ChatGPT web interface. (This is one reason I tend to avoid that in favor of the Playground and direct model access: there's much less moving machinery behind the scenes.) What the additional stuff does is quite hard to understand because the original prediction objective has been replaced by a bunch of mashed-together objectives combined with feedback loops, leading to some bizarre behavior like being unable to reliably write a nonrhyming poem. (Give that a try in ChatGPT: "write a nonrhyming poem". It's been slooowly getting better at it, perhaps because I and other people keep submitting examples of it failing to do so - but it will still usually fail! and when it does seem to be succeeding finally, if you let it keep writing lines, it will generally gradually revert back to rhyming.) If you go back to the original davinci-001, you'll find that it acts strikingly different than your description of contemporary ChatGPT. OP, incidentally, has some discussion of what it is like to interact with the real GPT-4, which is not like ChatGPT-4: https://www.lesswrong.com/posts/tbJdxJMAiehewGpq2/impressions-from-base-gpt-4 https://www.lesswrong.com/posts/tbJdxJMAiehewGpq2/impression... AFAIK, this is one of the only discussions online of what GPT-4-base is like qualitatively. (Note that it sounds a lot like Sydney - if you were around for that, the original Bing Sydney turned out to be a GPT-4 snapshot from partway through training which hadn't been RLHFed but given only some extremely inadequate custom Bing training.)
coldblues 3y ago
Roleplay is probably the most popular (casual) use of LLMs. Just check projects like SillyTavern.
- Tiberium 3y ago
  Yeah, I think all those scientific papers miss out a lot by not checking how people actually RP with models, especially in regards to the prompts used.
- dmd 3y ago
  Didn't read the article, huh?
mxwsn 3y ago
I first read about understanding LLMs as simulators generating simulacra a while ago through this post: https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators
famouswaffles 3y ago
Large Language Models are predictors. Not imitators, not simulators. Those are just apparent byproducts of performant prediction. The end goal and what the loss trending down to is to perfectly model the data it's been given. So it will not stop at "surface level similarity" or "plausible" or "uninspired" or whatever arbitrary competency line anyone tries to draw in the sand. It will continue to improve until it is "correct" (as determined by the data). Stick a bunch of protein sequences and it's not going to stop when it starts generating alphanumeric sequences that look like proteins but really aren't. It's eventually going to start generating real proteins. https://www.nature.com/articles/s41587-022-01618-2 https://www.nature.com/articles/s41587-022-01618-2 Then it's going to keep improving until it models the distribution of proteins in the dataset. I'm honestly not sure what the point of this paper is. "It is, perhaps, somewhat reassuring to know that LLM-based dialogue agents are not conscious entities with their own agendas and an instinct for self-preservation, and that when they appear to have those things it is merely role play." Ignoring the whole, "We don't know what consciousness is", it just seems devoid of any meaningful distinction. "Merely roleplay". What does that even mean ? That's it's not real ? Not really. They seem to understand this too. "It would be little consolation to a user deceived into sending real money to a real bank account to know that the agent that brought this about was only playing a role." Bing has a habit of ending conversations when users say upsetting things. You can talk all you want about how "it's not really upset" but the conversation did end and now you have to start over and be potentially less confrontational if you want to move forward. "Roleplay" as consequential as the "real thing" is the real thing. This shiny piece of yellow metal looks like gold, tests like gold, sells like gold but is not...real gold ? Not unless you have a meaningless definition of real.
- MichaelMoser123 3y ago
  > Large Language Models are predictors. Not imitators, not simulators How is it that an LLM can react to meta prompts like 'be brief' or 'Ensure responses are unique and without repetition' ?
- Al-Khwarizmi 3y ago
  Not to argue with anything you said but just as an aside, the Bing habit you mention doesn't seem to be from the LLM itself, but from some censorship module that has been bolted in. In the first few days after release, it would get really confrontational (and sometimes emotional) with some users and this attracted bad press, so now it ends the conversation before getting into any remotely thorny issue. There's also another censorship module that sometimes deletes its output before it has finished.
huijzer 3y ago
> On the other hand, taken too literally, such language promotes anthropomorphism, exaggerating the similarities between these artificial intelligence (AI) systems and humans while obscuring their deep differences. I don't see how the role play term doesn't introduce new issues. Cambridge, for example, defines role play as "pretending to be someone else". An LLM is also not pretending. Also what "role" would an LLM play if you just take the base model without a default prompt or finetuning?
- coldtea 3y ago
  >Also what "role" would an LLM play if you just take the base model without a default prompt or finetuning? What role would any person into sexual roleplay play if you don't prompt them into sexual roleplay?
- pixl97 3y ago
  Well, from everything I've seen, without the fine tuning then the output will not be 'human' enough. Or to put this another way... The 'role' may be more that of Legion in the biblical sense. Depending on the exact question asked hints of particular human characteristics show up, but there is a multitude of different ones and the model would seemingly randomly express them per question.
- duskwuff 3y ago
  Without any prompt or fine-tuning, most models won't converse with the user at all. They'll operate in a pure text-prediction mode, which is more likely to continue the user's prompt than to respond to it. Or it may end up generating a response to the prompt as if it were a question asked on a web forum like Stack Overflow -- which, yes, does mean that it will generate comments complaining that your question is off-topic and should be closed.
  - vintermann 3y ago
    Yes. In AI dungeon, which was an early ad-hoc attempt at using a LLM as GM before instruction fine-tuning was figured out, you saw this a lot. I remember some people posting chats where the model would shout that they'd had it with the player, and "User has left the chat".
IngoBlechschmid 3y ago
Related: ideas by Gwern to enhance AI dungeons with caching, yielding a novel form of choose-your-own-adventure games. https://gwern.net/cyoa https://gwern.net/cyoa
daralthus 3y ago
If you ever play Improv you soon realize that learning the basic rules of improv quickly lead to better scenes. You don't verbalize these while playing, but you keep them in your head and have a feel for them. I believe the authors greatly miss out on these internal representations by only focusing on the outcomes of what is being said. Shameless plug, I made a GPT for playing improv: https://chat.openai.com/g/g-LkQhMxpvM-improv-theatre https://chat.openai.com/g/g-LkQhMxpvM-improv-theatre
xwdv 3y ago
The inability to write suggestive or violent scenes or even have a simple fight between characters makes it difficult to do any interesting role play. The frustrating thing is that it feels like it could easily be done, but the creators are bound by some puritanical sense of moral obligation. I hate this!
- randcraw 3y ago
  I wonder if this policy will also prohibit the use of violent/war metaphors, like "loaded for bear" or "another weapon in his arsenal". If so, the average sports commentary will throw dozens of red flags.
  - rhdunn 3y ago
    Metaphors are one of the things that make language interesting, vibrant, and expressive. You can convey so much with them. Prohibiting their use (even in specific contexts like this) is/will be a huge loss.
j7ake 3y ago
AI is clearly the hottest field right now. Are there any other fields where a high fraction of the high impact papers (perspective or otherwise) are coming from industry rather than academia?
- atleastoptimal 3y ago
  [flagged]
  - low_tech_punk 3y ago
    and artificial intelligence to beget natural intelligence, completing the loop.
  - otabdeveloper4 3y ago
    There is no such thing as "intelligence".
    - slater 3y ago
      There is.
    - atleastoptimal 3y ago
      Then why can I type this sentence when a cereal box can't? And please don't give a snarky "because you have hands" type answer.
      - coldtea 3y ago
        Because the cereal box is not living matter, for starters. Parent didn't say there's no such thing as life.
        atleastoptimal 3y ago
        ~110-120 IQ response

92 points by m_kos 3y ago

I've tried to play dungeons and dragons with Chatai. It doesn't like to use any mechanical rules and keeps trying to wrap everything up like its got a train to catch. I'm sure you could tune it to have more attention than a two year old and actually apply the rules. It did come up with a decent setting based on B2: Keep on the Borderlands. And it was some fun. The new generation of text games will be awesome. And Enders tablet game is possible with this tech.

xeyownt 3y ago
You didn't read the article, did you? :-D
TeMPOraL 3y ago
> And Enders tablet game is possible with this tech. Complete with aliens subtly messing with it, and through it, the player? They who control the logits, control the future.
gwern 3y ago
> I've tried to play dungeons and dragons with Chatai. It doesn't like to use any mechanical rules and keeps trying to wrap everything up like its got a train to catch. Yes, that's the RLHF and other mechanisms. Raters presumably reward completions which are, well, complete and can be quickly judged as a whole, no matter what the intrinsic quality might be or how valid it would be to end with a 'Tune in next week for chapter 2'. What OP is describing is the emergent behavior of the base model, which acts very differently from the RLHF/instruction-tuned/who-knows-what-else ChatGPT web interface. (This is one reason I tend to avoid that in favor of the Playground and direct model access: there's much less moving machinery behind the scenes.) What the additional stuff does is quite hard to understand because the original prediction objective has been replaced by a bunch of mashed-together objectives combined with feedback loops, leading to some bizarre behavior like being unable to reliably write a nonrhyming poem. (Give that a try in ChatGPT: "write a nonrhyming poem". It's been slooowly getting better at it, perhaps because I and other people keep submitting examples of it failing to do so - but it will still usually fail! and when it does seem to be succeeding finally, if you let it keep writing lines, it will generally gradually revert back to rhyming.) If you go back to the original davinci-001, you'll find that it acts strikingly different than your description of contemporary ChatGPT. OP, incidentally, has some discussion of what it is like to interact with the real GPT-4, which is not like ChatGPT-4: https://www.lesswrong.com/posts/tbJdxJMAiehewGpq2/impressions-from-base-gpt-4 https://www.lesswrong.com/posts/tbJdxJMAiehewGpq2/impression... AFAIK, this is one of the only discussions online of what GPT-4-base is like qualitatively. (Note that it sounds a lot like Sydney - if you were around for that, the original Bing Sydney turned out to be a GPT-4 snapshot from partway through training which hadn't been RLHFed but given only some extremely inadequate custom Bing training.)

coldblues 3y ago

Roleplay is probably the most popular (casual) use of LLMs. Just check projects like SillyTavern.

Tiberium 3y ago
Yeah, I think all those scientific papers miss out a lot by not checking how people actually RP with models, especially in regards to the prompts used.
dmd 3y ago
Didn't read the article, huh?

mxwsn 3y ago

I first read about understanding LLMs as simulators generating simulacra a while ago through this post: https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators

famouswaffles 3y ago

Large Language Models are predictors. Not imitators, not simulators. Those are just apparent byproducts of performant prediction. The end goal and what the loss trending down to is to perfectly model the data it's been given. So it will not stop at "surface level similarity" or "plausible" or "uninspired" or whatever arbitrary competency line anyone tries to draw in the sand. It will continue to improve until it is "correct" (as determined by the data). Stick a bunch of protein sequences and it's not going to stop when it starts generating alphanumeric sequences that look like proteins but really aren't. It's eventually going to start generating real proteins. https://www.nature.com/articles/s41587-022-01618-2 https://www.nature.com/articles/s41587-022-01618-2 Then it's going to keep improving until it models the distribution of proteins in the dataset. I'm honestly not sure what the point of this paper is. "It is, perhaps, somewhat reassuring to know that LLM-based dialogue agents are not conscious entities with their own agendas and an instinct for self-preservation, and that when they appear to have those things it is merely role play." Ignoring the whole, "We don't know what consciousness is", it just seems devoid of any meaningful distinction. "Merely roleplay". What does that even mean ? That's it's not real ? Not really. They seem to understand this too. "It would be little consolation to a user deceived into sending real money to a real bank account to know that the agent that brought this about was only playing a role." Bing has a habit of ending conversations when users say upsetting things. You can talk all you want about how "it's not really upset" but the conversation did end and now you have to start over and be potentially less confrontational if you want to move forward. "Roleplay" as consequential as the "real thing" is the real thing. This shiny piece of yellow metal looks like gold, tests like gold, sells like gold but is not...real gold ? Not unless you have a meaningless definition of real.

MichaelMoser123 3y ago
> Large Language Models are predictors. Not imitators, not simulators How is it that an LLM can react to meta prompts like 'be brief' or 'Ensure responses are unique and without repetition' ?
Al-Khwarizmi 3y ago
Not to argue with anything you said but just as an aside, the Bing habit you mention doesn't seem to be from the LLM itself, but from some censorship module that has been bolted in. In the first few days after release, it would get really confrontational (and sometimes emotional) with some users and this attracted bad press, so now it ends the conversation before getting into any remotely thorny issue. There's also another censorship module that sometimes deletes its output before it has finished.

huijzer 3y ago

> On the other hand, taken too literally, such language promotes anthropomorphism, exaggerating the similarities between these artificial intelligence (AI) systems and humans while obscuring their deep differences. I don't see how the role play term doesn't introduce new issues. Cambridge, for example, defines role play as "pretending to be someone else". An LLM is also not pretending. Also what "role" would an LLM play if you just take the base model without a default prompt or finetuning?

coldtea 3y ago
>Also what "role" would an LLM play if you just take the base model without a default prompt or finetuning? What role would any person into sexual roleplay play if you don't prompt them into sexual roleplay?
pixl97 3y ago
Well, from everything I've seen, without the fine tuning then the output will not be 'human' enough. Or to put this another way... The 'role' may be more that of Legion in the biblical sense. Depending on the exact question asked hints of particular human characteristics show up, but there is a multitude of different ones and the model would seemingly randomly express them per question.
duskwuff 3y ago
Without any prompt or fine-tuning, most models won't converse with the user at all. They'll operate in a pure text-prediction mode, which is more likely to continue the user's prompt than to respond to it. Or it may end up generating a response to the prompt as if it were a question asked on a web forum like Stack Overflow -- which, yes, does mean that it will generate comments complaining that your question is off-topic and should be closed.
- vintermann 3y ago
  Yes. In AI dungeon, which was an early ad-hoc attempt at using a LLM as GM before instruction fine-tuning was figured out, you saw this a lot. I remember some people posting chats where the model would shout that they'd had it with the player, and "User has left the chat".

IngoBlechschmid 3y ago

Related: ideas by Gwern to enhance AI dungeons with caching, yielding a novel form of choose-your-own-adventure games. https://gwern.net/cyoa https://gwern.net/cyoa

daralthus 3y ago

If you ever play Improv you soon realize that learning the basic rules of improv quickly lead to better scenes. You don't verbalize these while playing, but you keep them in your head and have a feel for them. I believe the authors greatly miss out on these internal representations by only focusing on the outcomes of what is being said. Shameless plug, I made a GPT for playing improv: https://chat.openai.com/g/g-LkQhMxpvM-improv-theatre https://chat.openai.com/g/g-LkQhMxpvM-improv-theatre

xwdv 3y ago

The inability to write suggestive or violent scenes or even have a simple fight between characters makes it difficult to do any interesting role play. The frustrating thing is that it feels like it could easily be done, but the creators are bound by some puritanical sense of moral obligation. I hate this!

randcraw 3y ago
I wonder if this policy will also prohibit the use of violent/war metaphors, like "loaded for bear" or "another weapon in his arsenal". If so, the average sports commentary will throw dozens of red flags.
- rhdunn 3y ago
  Metaphors are one of the things that make language interesting, vibrant, and expressive. You can convey so much with them. Prohibiting their use (even in specific contexts like this) is/will be a huge loss.

j7ake 3y ago

AI is clearly the hottest field right now. Are there any other fields where a high fraction of the high impact papers (perspective or otherwise) are coming from industry rather than academia?

atleastoptimal 3y ago
[flagged]
- low_tech_punk 3y ago
  and artificial intelligence to beget natural intelligence, completing the loop.
- otabdeveloper4 3y ago
  There is no such thing as "intelligence".
  - slater 3y ago
    There is.
  - atleastoptimal 3y ago
    Then why can I type this sentence when a cereal box can't? And please don't give a snarky "because you have hands" type answer.
    - coldtea 3y ago
      Because the cereal box is not living matter, for starters. Parent didn't say there's no such thing as life.
      - atleastoptimal 3y ago
        ~110-120 IQ response