4 ms·

I skimmed the article and really enjoyed it. I admire the authors effort and will come back to read it in full soon. I have spent a lot less effort on a simila

I skimmed the article and really enjoyed it. I admire the authors effort and will come back to read it in full soon. I have spent a lot less effort on a similar endeavor before deeming GPT inappropriate for this purpose, with similar takeaways as the ones noted in "What doesn't work" I think the reason behind these shortcomings is because of the nature of GPT. It's designed to predict the most likely words to come after the prompt and as such it's very prone to fall in cliches, stereotypes and tropes. I think the next iteration algorithm is going to feel less like generated content but will still maintain a hack writer feel. I would wager the author could get better results with a system where the substance of the content is generated through another process, possibly something like dwarf fortresses world and lore generation, and having a GPT layer that would put that content in more creative prose.

impalallama 4y ago
I have come to a similar conclusion. I still get better results generating d&d npcs with my usual random charts than using gpt. Although I have a suspicion it’s because openai has heavily clamped down on gpt responses so it will not give wild inappropriate responses that would be bad pr. For example it generated exactly 1 Npc that could be considered rude with a special note how it’s sorry about that but that it’s important some npcs are rude for a more dynamic world (which I find amusing)
- pixl97 4y ago
  The business to user AI alignment issue is an interesting problem to watch folding out. As you state, if your creating a general help AI for things like searching you don't want a user asking "How do I solve all my problems?" and the AI responding "Tie a rope around a tree, the other end around your neck and jump". That's how you get the NYT headlines of 'AI kills users'. But that particular safety approach fails to create interesting 'human' stories. Sometimes the bad guy comes in and performs a genocide setting off the heros story arc. Sometimes the good guy is a terrorist (Luke Skywalker: No, I'm a freedom fighter). The problem I see here is the power to create interesting and believable fiction is the same power to talk people into suicide. It's the same power to talk a mob in to rioting. It's the same power that could drive a nation to war. As the power to run these models goes from something that requires a datacenter to run down to something that runs in the palm of your hand it's something that society and the world will have to deal with. Any of you that run sites that allow posting won't like the idea of a million virtual Nazi's that never sleep (or maybe you do like the idea of?). Authoritarians in power may love the idea of countless voices spreading their message. As they say, may we live in interesting times.
  - SamoyedFurFluff 4y ago
    I think we’re focusing too hard on whether or not AI can produce controversial in the sense of villainous stories. I find chatgpt also has issues producing stories from a minority standpoint. It cannot generate interesting stories about being gay, an immigrant, a refugee, incarcerated, poor, disabled, the list goes on and on. It only repeats the most stereotype of minorities, and I fear it perpetuates bias as a result. Even if I ask it to tell me about someone with say an invisible disability, even that is depicted in a way completely stereotyped and not at all reflective of actual experiences.
  - squarefoot 4y ago
    > Any of you that run sites that allow posting won't like the idea of a million virtual Nazi's that never sleep (or maybe you do like the idea of?) If I ran a social site, I would totally love a virtual world filled with virtual nazis, in which quietly confine the real ones, so that they can't pollute healthy environments with their nonsense.
    - pixl97 4y ago
      Unfortunately that's not how they work. Much like the ones that were in Germany in the 30's they love to perform a lebensraum on your website to spread their toxicity rather than stick to their own.
- kelsolaar 4y ago
  Are those random charts something you shared publicly?
pfdietz 4y ago
I'm feeling that automating the feedback process is going to be important for GPT.
throwaway4aday 4y ago
Did you play around with the temperature and the Top P settings? Adjusting those can vary the output quite a bit. I also think "prompt engineering" is a real art and not just a fancy name. If you can land on the right set of keywords and instructions it seems like you can explore uncommon areas of the model. I haven't tried fine tuning the model yet but from looking at other projects people are working on that use fine tuning it seems like you can get the model to act in a certain way more consistently by doing so. It just seems like a pain to generate all of the example prompts and responses.
ianbicking 4y ago
What's the cliche way to complete: "Here is a list of 10 things that have nothing in common:" ? Ask GPT this and it comes up with a pretty good list (a pineapple, the color blue, the taste of coffee...). Ask yourself this same question and you can come up with a pretty good list... but it's also surprisingly hard. There are ways we can increase our creativity and tangential thinking. Brainstorming lists, exploring opposites, free association, etc. A lot of these also work with GPT. Which is itself a strange thought... is GPT mimicking us or is there an attribute of cognition which we share? Anyway, while it takes work and exploration to figure this out, I've yet to hit a wall, so I am pretty optimistic about GPT generally. But it does involve a mixture of prompt and algorithm and human intervention.
- ianbicking 4y ago
  Also I should add: ask for what you want. You want weird characters? Say "give me weird characters." Be careful about your phrasing, you may be implicitly asking for hack results. Sometimes when you ask for something you'll get a too-literal response, like "more fraught with internal tension" and it will say "Anna is fraught with internal tension." But you can usually change the phrasing to fix that, like "show examples of how the character is fraught with internal tension."

impalallama 4y ago

I have come to a similar conclusion. I still get better results generating d&d npcs with my usual random charts than using gpt. Although I have a suspicion it’s because openai has heavily clamped down on gpt responses so it will not give wild inappropriate responses that would be bad pr. For example it generated exactly 1 Npc that could be considered rude with a special note how it’s sorry about that but that it’s important some npcs are rude for a more dynamic world (which I find amusing)

pixl97 4y ago
The business to user AI alignment issue is an interesting problem to watch folding out. As you state, if your creating a general help AI for things like searching you don't want a user asking "How do I solve all my problems?" and the AI responding "Tie a rope around a tree, the other end around your neck and jump". That's how you get the NYT headlines of 'AI kills users'. But that particular safety approach fails to create interesting 'human' stories. Sometimes the bad guy comes in and performs a genocide setting off the heros story arc. Sometimes the good guy is a terrorist (Luke Skywalker: No, I'm a freedom fighter). The problem I see here is the power to create interesting and believable fiction is the same power to talk people into suicide. It's the same power to talk a mob in to rioting. It's the same power that could drive a nation to war. As the power to run these models goes from something that requires a datacenter to run down to something that runs in the palm of your hand it's something that society and the world will have to deal with. Any of you that run sites that allow posting won't like the idea of a million virtual Nazi's that never sleep (or maybe you do like the idea of?). Authoritarians in power may love the idea of countless voices spreading their message. As they say, may we live in interesting times.
- SamoyedFurFluff 4y ago
  I think we’re focusing too hard on whether or not AI can produce controversial in the sense of villainous stories. I find chatgpt also has issues producing stories from a minority standpoint. It cannot generate interesting stories about being gay, an immigrant, a refugee, incarcerated, poor, disabled, the list goes on and on. It only repeats the most stereotype of minorities, and I fear it perpetuates bias as a result. Even if I ask it to tell me about someone with say an invisible disability, even that is depicted in a way completely stereotyped and not at all reflective of actual experiences.
- squarefoot 4y ago
  > Any of you that run sites that allow posting won't like the idea of a million virtual Nazi's that never sleep (or maybe you do like the idea of?) If I ran a social site, I would totally love a virtual world filled with virtual nazis, in which quietly confine the real ones, so that they can't pollute healthy environments with their nonsense.
  - pixl97 4y ago
    Unfortunately that's not how they work. Much like the ones that were in Germany in the 30's they love to perform a lebensraum on your website to spread their toxicity rather than stick to their own.
kelsolaar 4y ago
Are those random charts something you shared publicly?

pfdietz 4y ago

I'm feeling that automating the feedback process is going to be important for GPT.

throwaway4aday 4y ago

Did you play around with the temperature and the Top P settings? Adjusting those can vary the output quite a bit. I also think "prompt engineering" is a real art and not just a fancy name. If you can land on the right set of keywords and instructions it seems like you can explore uncommon areas of the model. I haven't tried fine tuning the model yet but from looking at other projects people are working on that use fine tuning it seems like you can get the model to act in a certain way more consistently by doing so. It just seems like a pain to generate all of the example prompts and responses.

ianbicking 4y ago

What's the cliche way to complete: "Here is a list of 10 things that have nothing in common:" ? Ask GPT this and it comes up with a pretty good list (a pineapple, the color blue, the taste of coffee...). Ask yourself this same question and you can come up with a pretty good list... but it's also surprisingly hard. There are ways we can increase our creativity and tangential thinking. Brainstorming lists, exploring opposites, free association, etc. A lot of these also work with GPT. Which is itself a strange thought... is GPT mimicking us or is there an attribute of cognition which we share? Anyway, while it takes work and exploration to figure this out, I've yet to hit a wall, so I am pretty optimistic about GPT generally. But it does involve a mixture of prompt and algorithm and human intervention.

ianbicking 4y ago
Also I should add: ask for what you want. You want weird characters? Say "give me weird characters." Be careful about your phrasing, you may be implicitly asking for hack results. Sometimes when you ask for something you'll get a too-literal response, like "more fraught with internal tension" and it will say "Anna is fraught with internal tension." But you can usually change the phrasing to fix that, like "show examples of how the character is fraught with internal tension."