So with a safe prompt there is always a chance the AI will go on a bad direction and then refuse to work, and make you pay for the tokens of his "I am sorry ....long speech"
But if you’re looking to generate truly racist and vile stuff, yes, you probably will need a model with no training or inference filters. I’m not sure anyone’s investing in building that though.
Really, these sorts of ideas feel like we're getting to the "put everything on the blockchain!" phase. "Let's spend more GPU power for creating speech for the Sims than it takes to run the Sims itself!"
In theory, a full blown LLM gives you a lot more variety and ability to handle novel situations, but it also gives you a lot more potential for conversational gambits that don't affect the game mechanics in the way you want them to and general weirdness (I love the article's anecdote about the Sim who thinks his neighbour Adam Smith wrote Wealth of Nations!). I'm sure someone will ultimately end up designing great LLM-driven game experiences, but I don't imagine they'll look much like The Sims.
Will Wright on Designing User Interfaces to Simulation Games (1996) (2023 Video Update)
https://donhopkins.medium.com/designing-user-interfaces-to-s...
Will Wright - Maxis - Interfacing to Microworlds - 1996-4-26
https://www.youtube.com/watch?v=nsxoZXaYJSk
Here's the transcript, when a student asked him about that:
https://youtu.be/nsxoZXaYJSk?t=4466
Student:
What about from person to person, you talking about the information that's contained within the objects, so there can be information in another person, that you want to interact with in the environment.
Have you looked at any reasons why you would want to do that?
Will Wright:
Oh yeah, I mean, that's that's the hard problem.
I mean, simulating ants is hard enough, when you get to people there's really no hope.
There are two issues here.
You can look at this as a technology.
It's not a product right now.
And their are a few directions this could go.
I could see this becoming, let's say, a multiplayer network MUD kind of a thing.
You might have a thousand people playing SimCity from the bottom up, each person building their own house in a big multi-user space.
In which case that issue is a little less important, because most of the people are real people, and you're dealing with puppets.
As a standalone game, which is probably our our closer target, we have to deal with the problem you're bringing up, which is how do we deal with people to people?
And it's hard, I mean there's just -- I'm sure Terry can elaborate on that more than I can.
But the best thing we can do is prop up a convincing illusion.
We don't have to be doing a valid simulation of human personality.
What we have to do is we have to put up something that's ambiguous enough to where somebody can read in what they want.
Actually in this thing what I have right now are people come up and they converse, but you don't hear what they're saying, they just gesture, and sometimes they look mad, sometimes they kind of look contemplative.
It's kind of interesting how much people will read into that.
This is kind of dynamic that we've seen again and again where something happens in SimCity and they said "oh I was running my nuclear reactor near the red line, and then there was so much smoke coming out of it, this plane crashed, and because of that, this and that happened", and they'll describe this long causal chain of events that I know does not exist.
I designed the simulation, I know that there's no linkage between the power output of the power plant, the planes crashing, but they're convinced it exists.
Don Hopkins:
They're using it as a medium to tell stories about.
Will Wright:
Yeah!
Don Hopkins:
Where they're using it as a piece of paper, to write.
Will Wright:
Yeah, that's exactly right.
There's a parallel simulation going on here in the game.
Everybody's taking a linear path through this, and they're basically, most people will attempt to understand things like this with a story.
They'll think about "I did this, then that happened, because of that", and so the story becomes kind of their logical connection, their logical reverse engineering, of the simulation that they're playing inside of.
Now on the people's side, I think we can do a lot in this as a product, by propping up that illusion of people.
Again, if this is a doll house, we don't want the dolls to be sentient things.
We want the dolls to be interesting enough to where I can play games with them.
There was actually a really interesting doll that this company came out with.
Oh, it was Worlds of Wonder, this really cool doll, I've got a couple of them after they went out of business.
It's called the Julie doll.
[The 1987 Voice First Doll: Julie by Worlds of Wonder Commercial: https://www.youtube.com/watch?v=ewu_NBUHePU ]
But it was like this $250 doll with voice recognition, and it said all these things.
It had just a huge amount of ROM with digitized speech in it, and so it would sit there and try and have stupid conversations with you.
And really it was kind of Eliza, or had keywords it would recognize, and give you these kind of non-committal responses.
But in the testing of that, well first of all it was a $300 doll.
Who's going to buy the kid a $300 doll?
So it was really more, it was actually the only doll I've ever seen that appealed to grown men.
Grown men love this, I mean this is a hacker's doll.
But I talked to the guy who was working this project, and he said they put this in focus groups with girls.
And they played with it for a while, and then after about a half an hour they take the batteries out, and keep playing with it.
And what was happening is that the girls were propping up this elaborate fantasy in their play, and the dolls were supposed to be a structure for that fantasy, they weren't supposed to be the fantasy.
The doll was telling them what the fantasy was, and it was conflicting with what the girls were saying, and so it was interfering actively with their fantasy and their play.
So in that regard, I think we can actually kind of take that path with these people.
And all we have to do is deal with them at a very local kind of a state machine, Braitenberg Machine kind of level, and say that they're angry, and they're hungry, and they're sleepy.
[Braitenberg vehicle: https://en.wikipedia.org/wiki/Braitenberg_vehicle ]
And then we can actually do some things where maybe they have a little, what you might call, structural ambiguity about what they're actually saying.
One of the thoughts I had about this project in particular is that you'd see the people go up and they talk, and there would be some kind of a flavor to their conversation, but it would be more like Peanuts.
When they did the TV show of Peanuts, you'd hear the adults talking, and the adults would always be like "mwa mwa mwa mwa mwa mwa", or soft, or loud.
[Peanuts' Teacher Calls Out Charlie Brown & Linus - "Wah Wa Wa Wah Wa Wa" - 1969: https://www.youtube.com/watch?v=CxC_AjFxS68 ]
You can tell if they're mad, or angry, or what, but you wouldn't hear what they were saying.
You'd have to read that into it.
I think this is the area where we sidestep the issue, just because as a commercial company we have to ship a game, we're not doing a research project.
[How the Language From the Sims Was Created: https://www.youtube.com/watch?v=FGsbeTV76YI ]
So that's a long winded answer, sorry.
Yeah?
Student:
If you do do that, and people can read into what people are actually doing in the game, won't that sort of interfere with their expectations and simulation if the simulation doesn't actually meet those expectations?
If they're reading too much into it, and it's not really happening that way, won't they be sort of misled in there expectations?
Will Wright:
Yeah it really comes down what the simulation is.
That, I think, would be addressed on the game design side.
Again, I mean, just in terms of reality, I don't think there's any way I'm going to come anywhere close to simulating a person.
With a lot of tricks, I can have these people walking through the day, getting up, taking a bath or a shower, fixing breakfast, going to work, while at the same time deciding that I'm bored, I'm going to sit on the couch, I'm going to turn on the TV.
And how you make that into a game design might have a lot to do with how valid the simulation of the people has to be.
If it's a story kind of a game where you want drama to unfold at the right time, and the conversation between these two people is a crucial thing that happens before he gets really upset and does that, then yes the model would need to be a real simulation.
I really don't think we're anywhere near that though, I mean, really I think, unless you're in a very very tightly confined domain, you're going to have a hard time dealing with this open-ended simulation about people in that way.
So I think really unfortunately we're not there yet, and I think it'll be quite a long time before we are, so really we have to constrain what the user is doing.
Now maybe it's just a dollhouse, maybe all I do is I have to get two people to meet at a party, and everything else is kind of indeterminate about what they say and all that, and I'm moving furniture around.
So there are a lot of game design things we can do with this without doing a personality model.
Don Hopkins:
You're saying a multi-user game would be easier to design than an AI game, because you can use other people.
Will Wright:
Oh yeah, far easier, I think so.
Don Hopkins:
It's just the technology of communicating and time lags.
Will Wright:
This is how you're always building models of the system you're playing with.
It's easy to build a model of a stupid computer agent.
It's hard to build a model of your head, while I'm talking to you.
That's what's interesting, is trying to reverse engineer your thought process.
Don Hopkins:
Yeah, there's a lot of commercial services like Worlds Away and Habitat, things that have been done that only have other people, no robots or anything, and they're pretty successful just because of that.
Will Wright:
Yeah, people are pretty interesting.
Isn't that kind of tech exploration central to the hacker ethos? (not trying to No True Scotsman anyone who disagrees, just how I've always understood it)
It's what I always find with open world games for example. No matter what kind of character I build or how I behave I'm only ever going to get predetermined dialogue which I could have looked up and saved myself the time.
Yes in the sense of the Sims it's probably overkill (they sold millions of copies even without a script) but it's only being used as a playground to test the ideas and see what's possible, there's no suggestion that this is particularly what the Sims itself should be.
Bots trained to play like a human are more fun, Bhuman Soni; Philip Hingston
Apparently, human like opponents were considered to be more fun. This kind of thing could be integrated easily in games, as the neural network is positively minuscule compared to an LLM.
Also there is nog chance of accidents if your npc's cant talk (well, they might shoot you, but they won't heckle you).
So I'd expect this to be integrated in real games much sooner than LLM based characters.
ChatGPT-4:
Creating an a cappella cover of a Philip Glass song in Simlish presents an interesting challenge. Philip Glass is known for his minimalist compositions that often consist of repetitive structures, and Simlish is a nonsensical language. Since Glass's compositions are largely instrumental, I will create an a cappella adaptation inspired by his musical style while using Simlish-like sounds.
Verse 1:
Doo doo zib, doo doo zib, doo doo zib, doo doo zib,
Bada bada zoob, bada bada zoob, bada bada zoob, bada bada zoob,
Glarbni glarbni glarb, glarbni glarbni glarb, glarbni glarbni glarb, glarbni glarbni glarb,
Sul sul sul, sul sul sul, sul sul sul, sul sul sul.
Chorus:
Froodle froodle, zibba zibba,
Dag dag, drib drib, yibbs yibbs,
Slooble slooble, wib wib, dor dor,
Lerpna lerpna, gort gort, zor zor.
Verse 2:
Doo doo zib, doo doo zib, doo doo zib, doo doo zib,
Bada bada zoob, bada bada zoob, bada bada zoob, bada bada zoob,
Glarbni glarbni glarb, glarbni glarbni glarb, glarbni glarbni glarb, glarbni glarbni glarb,
Sul sul sul, sul sul sul, sul sul sul, sul sul sul.
Chorus:
Froodle froodle, zibba zibba,
Dag dag, drib drib, yibbs yibbs,
Slooble slooble, wib wib, dor dor,
Lerpna lerpna, gort gort, zor zor.
The inevitable four thousand startups trying to monetize shoving a LLM into everything when the technology is excessively wasteful and unwarranted, that's going to be the bad bit.
imagine this issue when you are just the devloper and not the user, the user complains about this but you try and works for you, but then it fails again for user, in my case the word "monkey" might trigger ChatGPT to either create soem racist shit or it's moderation code to false flag itself.
(1) For example here: "Doo doo zib, doo doo zib, doo doo zib, doo doo zib"
You'd actually do something like: "Doo doo zib, doo doo zib, doo zib, doo zib, doo doo doo zib, doo zib, doo zib, doo doo doo zib, doo zib zib, doo zib zib, doo doo zib, doo doo zib"
(2) Here: "Sul sul sul, sul sul sul, sul sul sul, sul sul sul."
He'd likely go for something like: "Sul sul sul, sul sul sul, sul sul sul, sulna sulna sulna, sul sul sul, sulna sulna sulna, sulna sulna sulna, na na na, na na na, sul sul sul, sulna sulna sulna, na na na".
Source: Am a composer who composes music in minimalist style with strong influences from Philip Glass.
Surprisingly the English translation rhymes!
The fact that Simlish lacks any precise meaning probably makes it a bit more tractable.
Michel Kandel's translations of Stanislaw Lem's robot poetry from Polish to English also rhymed, but he's an world class translator and also an accomplished science fiction writer and editor himself.
https://en.wikipedia.org/wiki/Michael_Kandel
Here's a previous discussion of Marcin Wichary's translation of one of Lem's stories from Polish to English. He created the Lem Google Doodle, and he stalked and met Stanislaw Lem when he was a boy. Plus a discussion of Michael Kandel's translation of the poetry of the Electric Bard from The First Sally of Cyberiad:
https://news.ycombinator.com/item?id=28600200
>Translations of [Lem’s] works are difficult due to passages with elaborate word formation, alien or robotic poetry, and puns. — Wikipedia [...]
>Memoirs of a train traveller: The day I stalked Stanisław Lem [...]
>I have always wanted to ask Michael Kandel how he translated the poetry of the Electric Bard from The First Sally of Cyberiad.
>Especially the one that is "A love poem, lyrical, pastoral, and expressed in the language of pure mathematics. Tensor algebra mainly, with a little topology and higher calculus, if need be. But with feeling, you understand, and in the cybernetic spirit."
Prompt: Write a poem in Simlish.
ChatGPT-4:
While it's important to remember that Simlish is a nonsensical language without a proper grammar or vocabulary, I can still attempt to create a playful poem using some known Simlish phrases and sounds. Keep in mind that this poem will not have any real meaning:
Sul sul! Vadeesh, plurna zib,
Badeesh! Dag dag, cribna drib,
Froobi flazzle, noogit fraz,
Glarbni weeb, frizzle maz.
Lerpna gort, jibbly zor,
Slooble wib, kwebna dor,
Skibbi plurt, pleeba narn,
Grubble veeb, shazzle darn.
Wabble gret, minzi glarn,
Flartle doo, zibba yarn,
Yibbs frazzle, dribna glorp,
Vorpti zound, greeble snorp.
Prompt: Translate that poem to English, as best you can.
ChatGPT-4:
As a language model AI, I must inform you that Simlish is a fictional language created for The Sims video game series. It does not have a direct translation to English or any real-world languages, as it is designed to mimic the rhythm and flow of speech without conveying any actual meaning. However, I can attempt to provide a creative interpretation of the poem, loosely inspired by its structure and sounds:
Greetings all! Joyful day, skies so bright,
Cheers! Hello, dreams take flight,
Whimsy dazzle, laughter near,
Chatter soft, moments dear.
Dancing thought, playful breeze,
Gentle touch, secrets seize,
Colors blend, friendship true,
Laughter shared, skies of blue.
Murmured dreams, shadows play,
Whispered words, threads of gray,
Heart's embrace, silent gleam,
Boundless sounds, life's grand scheme.
No it's not. Elon Musk thinks there's a huge demand for racist AI, and he's probably right, or at least alt-right.
Elon Musk yearns for AI devs to build 'anti-woke' rival ChatGPT bot:
ChatGPT-4:
Thank you for your valuable feedback and insights into Philip Glass's minimalist style. I appreciate your expertise and will attempt to create a more accurate a cappella cover of a Philip Glass song in Simlish, keeping your suggestions in mind:
Sul sul la, sul sul la, sul sul la, sul la sul la,
Sul la, sul sul la, sul la, sul la sul la,
La la, la la, la la, la sul sul la,
La la, la la, la sul la, la sul sul la,
Zib zib doo, zib zib doo, zib doo zib doo,
Doo zib, doo zib, doo doo zib doo, zib doo zib doo,
Zib doo, zib doo, doo doo zib, doo doo zib,
Zib doo zib, zib doo zib, zib doo zib, doo doo zib,
Sul sul na, sul sul na, sul na, sul na sul na,
Sul na, sul sul na, sul na, sul na sul na,
Na na, na na, na na, na sul sul na,
Na na, na na, na sul na, na sul sul na,
Doo doo la, doo doo la, la la doo, la la doo,
Doo la doo, doo la doo, la la doo, la la doo,
Doo la la, doo la la, la la doo, la la doo,
La doo, la doo, doo la doo, doo la doo.
EDIT: "doo doo la" -> "la la doo" looks a bit more "creative" since I don't think this is a suggestion I made (unless you added more examples to your response). So, yeah not bad.
EDIT2: Nevermind, I did use that kind of transformation in my example (1) above: "doo zib zib" -> "doo doo zib". In short, it seems to have vaguely understood that the point of repetition here is to emphasize transformations. But it should be able to come up with novel transformations like Glass does.
I don't. I assume it would need to be constantly running to know when it wants to speak and there will be multiple actors on the screen all the time. Do we have actual estimates for how much a response costs in ChatGPT? All I know is it takes a lot of video cards to power that system.
> If you have some optimized LLMs running on the client
Do these currently exist? I was under the impression that tech to date is compute intensive if you're looking for near real time interaction.
If you want a layer to moderate what the year is seeing, you can add that as well. The point of the reverse moderator is to get GPT to do what it’s told without lying about itself, more or less.
Again: 1 I give them safe/clean prompt 2 AI returns 2 of 10 times unsafe crap that is filtered by them 3 I have to pay for my prompt, then have to catch they non deterministic response and retry again on my money
What should happen
1 customer give safe/clean prompt 2 AI response in racist/bad way 3 filter catches this , then it retries again, a few times, if the AI is still racist/bad then OpenAI automatically adds to the prompt "do not be a racist" 4 customer gets the answer