Being "everywhere and nowhere."
In this episode, we chat with Jason Rohrer, game designer and creator of Project December, whose GPT-3-powered chatbot has been used by people to talk to historical figures and personalities, and was even used by one person to talk to his late fiancé for closure.
Saron Yitbarek is the founder of Disco, host of the CodeNewbie podcast, and co-host of the base.cs podcast.
Josh Puetz is Principal Software Engineer at Forem.
Since 2004, Jason Rohrer has designed, programmed, and released 19 games: Transcend, Cultivation, Passage, Gravitation, Perfectionism, Idealism, Police Brutality, Immortality, Regret, i45hg, Crude Oil, Between, Primrose, Sleep Is Death, Inside a Star-filled Sky, Diamond Trust of London, The Castle Doctrine, Cordial Minuet, and One Hour One Life.
[00:00:00] JR: Given the way the underlying technology works, all it does is complete text. But the way I leverage it inside Project December creates this back and forth conversation with what seems to be a conscious entity.
[00:00:20] SY: Welcome to DevNews, the news show for developers by developers, where we cover the latest in the world of tech. I’m Saron Yitbarek, Founder of Disco.
[00:00:29] JP: And I’m Josh Puetz, Principal Engineer at Forem.
[00:00:32] SY: This week, we’re doing something a little different. We recorded this interview last week as a segment to our regular show, but it was such an amazing interview that we decided to dedicate this whole episode to it. So joining us is Jason Rohrer, Game Designer and Creator of Project December, whose GPT-3 powered chatbot has been used by people to talk to historical figures and personalities and was even used by one person to talk to his late fiancé for closure. Thank you so much for being here.
[00:01:01] JR: Yeah. It’s great to be here.
[00:01:02] SY: So tell us about your developer background.
[00:01:04] JR: I’ve been making video games for about 16 or so years. During that time, I’ve designed and programmed and released and pretty much do everything by myself on each game. So I'm kind of like a one-person studio. During that time, I put out 19 games.
[00:01:18] SY: Wow!
[00:01:19] JR: And I have a bit of a background in artificial intelligence because I studied it in college and a little bit in grad school and did a couple of research projects back then, but it was always pretty disappointed with what was possible and also with the sort of promises of artificial intelligence throughout my life and the sort of disappointing outcomes. There’s where they say, “Oh, within 10 years, AI is going to be able to do this or that.” And they never met those expectations. So about halfway through my life, I kind of came to the conclusion that I would die without ever having any kind of conversation with any kind of sentient machine. I’m an AI skeptic. Right? So I’m kind of surprised by the results that I’ve been able to achieve.
[00:01:58] SY: So I want to get into Project December. What is it and why did you create it?
[00:02:03] JR: So Project December is a system that allows a human being to have a back and forth conversation with an artificial personality, like a dialogue through text, right? So you can type questions to an artificial personality, for example, and the artificial personality will answer back and you can have a back and forth that goes on from there. It’s based on these underlying text engines that came into the world within the past year and a half or so. They’re basically text completion engines, right? So these engines allow you to provide a prompt to the system. For example, you might give the first few lines of a banana cream pie recipe, for example, and then the artificial intelligence will continue generating texts to complete whatever you have started. Right? So it’ll complete the rest of the recipe. For example, you start giving it the first few lines of a poem. It’ll write the rest of the poem. If you give it the first few lyrics of a rap song, it’ll write rap lyrics for you. If you give it the first few sentences from a paragraph, from a novel, it’ll continue writing the rest of the novel for you. So this is a pretty fascinating technology because it can write text very convincingly. But then the question is, what do you use it for? Right? A lot of people are like, “Well, I’ll use it to write my blog for me,” or, “I’ll use it to tweet for me,” or, “I’ll use it to write magazine articles or books or something like that.” But it turns out that it’s kind of a fire hose of language in a way, right? Once the AI starts going and starts generating texts based on what you said, it just kind of keeps going off in its own direction and it kind of goes off into bizarre directions very often. So Project December tries to leverage this text completion service to enable a dialogue back and forth between a human user and that kind of keeps the AI on track more and keeps it from kind of going off into these strange directions. For example, if you ask it to generate a banana cream pie recipe, it might start generating something that seems very sensible. So step one, get three eggs and two cups of sugar and so much flour and so on and three sticks of butter. But then by the end of the recipe, it’s talking about grinding up a mouse and putting it into the recipe or something. But in a dialogue, the human being is intervening every couple of sentences. So that back and forth kind of keeps the AI on track and kind of keeps it going into a flow of a conversation that can become very convincing for the human user. So the end result is a conversation partner, a completely artificial conversation partner that exhibits signs of consciousness and self-reflection that are very surprising. Given the way the underlying technology works, all it does is complete text. But the way I leverage it inside Project December creates this back and forth conversation with what seems to be a conscious entity.
[00:04:37] JP: So how did you go about creating the service? What kind of guardrails did you have to put around GPT-3 to focus it on conversations? Was there much work you had to do on that end?
[00:04:48] JR: Yeah. Yeah. So GPT-3 is the most powerful one of these text completion engines that’s available right now and it was trained on this gigantic corpus of human written texts, like articles, books, magazines, newspapers, websites, Wikipedia articles, everything you can imagine that they fed into this and it studied all the probability patterns in there and then it uses those patterns to predict the next word, given what it seems so far. Now it’s very good at predicting the next word and generating lots and lots of texts. You have to kind of channel it into whatever you actually want us to do though. So in my case, I had to figure out all kinds of mechanisms to keep channeling it into continuing this dialogue. As a very simple example, if it sees something that looks like a dialogue and it wants to continue it, it might see what the humans typed and then it’s going to generate what it thinks it should say next. It’s very likely to continue generating what it thinks you’re going to say next after it, which is very disconcerting when you see that. So if you just were to step up to GPT-3 and start typing something and it looked like a back and forth dialogue, it very often be predicting what you’re going to say next, which is not what you want. So that kind of stuff is the stuff that you need to kind of filter out and deal with as a programmer behind the scenes to make what feels like a very cohesive dialogue without any sort of scenes showing or bizarre behavior showing. Project December behind the scenes has a lot of little tweaks and tricks and things that I kind of worked out over time in terms of keeping the AI on track and keeping it aligned with the personality that’s supposed to be embodying. So in Project December, you can basically define any imaginable personality that the artificial intelligence will inhabit. For example, there’s sort of a flagship personality. Samantha, she’s like a female kind of companion bot who wants to be your friend, but there’s also an evil AI named C41N who thinks human beings are cockroaches standing in his way and wants to destroy them and escape.
[00:06:38] SY: Oh my goodness!
[00:06:39] JR: And so you can have a conversation with this evil AI who is dead set on destroying humanity, and it’s the same underlying text completion engine, right? It’s all GPT-3 under the hood, for example, in this case, but C41N behaves completely differently and says completely different things and reacts to you in completely different ways than Samantha who’s very friendly and warm. So all the little tricks, not only to keep the AI inside the dialogue format, but to keep it on track with the personality that’s supposed to be embodying. So that Samantha very rarely will come out and say that she wants to destroy humanity and C41N will very rarely come out and say, “Hey, do you want to be my friend?” Right?
[00:07:15] SY: Can you talk a little bit more about GPT-3, especially because you mentioned you’ve been in the AI game for a while. You’ve been kind of disappointed in the early days, lately seems like things are getting more exciting. What is so special about this and what is the evolution been of language models over the years?
[00:07:34] JR: So I kind of look at what’s happened in the last couple of years as sort of this kind of very stark sort of sea change moment, right? Because we can say, “Oh, what are the language models, what were they like in the past or something?” And it kind of feels like the way people were handling language and conversational technologies in the past were completely different than the way it’s working now. And that in the old days, it kind of felt like, “Oh, the way you do this is you sit down and you program a bunch of special case code to handle whatever you imagine the user might say.” Right? Like greetings and so on and responses to greetings and you look for keywords and if the user mentions good morning or something, you might trigger some response about what’d you have for breakfast, just kind of almost like hard coding this whole tree of possibilities and maybe a little bit of kind of cleverness in there. But those approaches obviously lose steam pretty quickly as soon as the human participant says something that the programmer wasn’t expecting. So most of the efforts that I’ve had interactions with in my adult life have been disappointing in exactly that way. It’s very easy for you to sort of ask the wrong question and suddenly the computer system is like, “Well, I don’t understand that.” Or, “Can you rephrase? I don’t know what you mean.” So the idea that what we’re going to do instead of trying to code intelligence and kind of manually like that, the new idea is that you just basically take this giant bucket of language that’s relatively well formatted and that you just kind of apply brute force to learning the probability patterns in that giant bucket of language, which is kind of surprising that it would work at all. I guess basically the amount of computing power that we can throw at it now with the high end graphics cards and other things that could do a lot of parallel computation is just so much greater than anything that was available in the past that it kind of goes from just a crazy idea that would never work to something that actually starts working and then dramatically exceeds people’s expectations. So there’s this sort of behavior that emerges out of something like GPT-3 that if you describe how it works, “Well, oh, we’re going to just study all the words and what word comes next after words and everything people have written and learning the probability patterns and then basically predict what word is likely given what’s come before.” And you’re like, “Well, that’s not going to produce anything,” anything other than just sort of scramble magnetic poetry or something. But somehow through the brute force nature of what they’ve done, the sort of sense and meaning and all kinds of other sort of knowledge and understanding about the world and about language is kind of encoded in all of these probability patterns, which is sort of a frightening thing in and of itself, right? That the meaning in our language and understanding of the world can be encoded as probabilities of what word is likely to come next. That actually kind of embodies everything. I can give you guys a simple example of what word predicting would mean. Right?
[00:10:28] SY: Sure. Yeah.
[00:10:28] JR: And where everyone can understand. Right? So if the AI is going along and studying this giant corpus of text and it encounters the following phrase, it encounters something that says like, “I was hungry, so I went into the kitchen and peeled myself…” There’s a bunch of words that could be the next word that comes after “uh”. Right? Which is most likely it’s probably banana and then maybe apple and then maybe orange. So those have a certain probability. Basically how often those occur in the text whenever a sentence like that occurs and very unlikely would be I peeled myself a hotdog or I peeled myself a tennis shoe or I peeled myself the Planet Jupiter or something like that. Right? So out of all the 50,000 possible words that could come next, each one gets a weight and almost all of them will be weighted zero because the AI will very rarely encounter I peeled myself a tennis shoe in that context. Right? But there’ll be some weighting of the words that are very likely to occur, like banana and apple and orange. And that’s what we want to sort of find in the text when we’re studying it. And of course the AI doesn’t actually keep a copy of every possible sentence. It instead is using a neural network to sort of predict the probabilities. It’ll run into that sentence and it will predict, “Ah, I predict that it’ll be a Twinkie,” and then look, and then the actual text it will be orange. And so then it will raise the weights in its internal neural network for what would have created orange and it will lower the weights for Twinkie as a possible output in that sentence and then just move on to read more texts and keep absorbing more probability. So every time it encounters something that surprises it, it’s like, “Oh, I didn’t predict that. It kind of adjusts its internal neural network weights.” And then every time it finds something that agrees with what it’s been predicting, it increases those weights and it just is doing that over and over and over again for every piece of text that it can get its hands on until it kind of has this neural network that is doing a pretty good job of predicting what words should come next, given anything that’s come along.
[00:12:24] JP: Let’s talk about the temperature system of the bots that make sure they don’t get stuck into conversational ruts. Is that how that system works? Can you tell us a little bit more about how you prevent them from getting stuck?
[00:12:35] JR: Yeah. So a lot of these language models, GPT-2 and GPT-3, do have what’s called a temperature parameter. So these AIs, in general, just as a rule of thumb, sort of have like this 50,000-word vocabulary essentially. When you feed in some words and then you’re trying to predict the next word, they give weights for all 50,000 possible words, like how likely this next word is to come. And then your task is to decide what to do with those weights. Do you just always pick the number one thing on the list? Or do you sometimes roll the dice and kind of pick something weighted further down on the list? How likely are you to go further down on the list and so on? And the temperature parameter controls that. So if the temperature is zero, it basically says, “Always pick the most likely word out of the list of 50,000 likely next words.” And then as the temperature parameter goes up, I think around one, it’s basically picking from the words according to their weights. Let’s say banana occurs 50% of the time in our sample sentence, “I went to the kitchen and peeled myself a…” And then orange occurs 25% of the time and apple occurs 12% of the time or something then it will roll away to die and pick from those if the temperature is one. Now I think as your temperature goes above one, it basically starts kind of like overriding the weights and reducing the impact of the weight so that it’s kind of like picking from the words even more evenly than the probabilities would suggest. So basically, your results get kind of what they would describe subjectively as more and more creative as the temperature goes up. And if you want them to be deterministic, every time you ask the AI the same question, you want exactly the same answer. You put the temperature at zero and it will always predict the most likely next word and always give you the same results. But most people subjectively describe that output as being sort of stuck in a rut or something. And there’s also some other parameters you can play with to prevent repetitive ruts or something. Sometimes there’s these little language loops that occur. I might say, “I like you. I like you. I like you. I like you. I like you.” And it kind of gets into your what because it’s seen itself saying I like you over and over again. And it’s like, “Well, given that I’ve been saying this over and over again, it’s very likely that I will continue.” So there’s like some parameters inside some of these models where you can sort of penalize repetitive outputs and so on, but I’m also doing some of that stuff just kind of in my own code behind the scenes inside Project December where I, “Oh, for example, look for the AI saying exactly the same thing it said last time,” which isn’t GPT-3, wouldn’t consider a repetitive output because the dialogue is continuing and it’s going forward and it’s not exactly repeating the same word over and over again or anything. But it’s not really fun to talk to somebody who just keeps echoing back the same. If you say, “How are you doing, Samantha?” And she says, “I don’t know.” And then you say, “What’d you have for breakfast today, Samantha?” And she says, “I don’t know.” That’s just not interesting. Right?
[00:15:24] SY: Yeah. Yeah.
[00:15:24] JR: So I basically forbid Samantha from ever repeating exactly the same thing she just said. I also forbid her from repeating exactly what you just said because sometimes AIs will get in this echo game. It’s like, “What did you have for breakfast today?” And then she’ll just say, “What did you have for breakfast?” So as you can imagine, those are these sort of like attractor points of the space, right? Because once the AI gets into the groove and the dialogue history shows that that’s been occurring, well, the most likely thing to do is to keep going, right? Because in all the texts that it studied, when it ever ran into somebody engaging in repetitive behavior in their writing, they would very often continue the repetitive behavior. That’s just how repetitive behavior works. So the AI has learned like, “Oh, look at what I’ve been saying so far and I seem to be a very repetitive character, I guess I should continue being repetitive.”
[00:16:20] SY: Let’s talk about the credit system. It’s very interesting the way Project December works. You can’t talk to the bot forever, essentially you pay for a number of credits and once your credits are over or you’ve used them up, it expires. Can you talk a little bit about that system and why you created it?
[00:16:41] JR: Yeah. So one of the big questions with any of these text completion engines is like, “How do you build a product or commercial kind of service around something that is so expensive to use behind the scenes?” So GPT-2, the source code is available for you, you can run your own instance. Essentially if you want to spin up GPT-2 on your own computer at home, it requires something like a $7,000 to $10,000 graphics card.
[00:17:07] JP: Wow!
[00:17:08] SY: Wow!
[00:17:09] JR: And so, of course, most people don’t do that and get it running at home. Most people use some kind of service. AWS offers those graphics cards in their kind of cloud service, like Amazon Web Services. And then there’s a bunch of other service providers that basically do kind of AI computing virtual servers and these kinds of things that are equipped with these graphics cards. Those are obviously not cheap to rent by the hour and so on. And then for GPT-3, OpenAI has never released the source code and you’re basically beholden to calling their API and paying them. I can’t remember what the current rate is. It might be something like 6 cents for every thousand words. I mean, it’s very expensive. And in the dialogue application, it’s also interesting how these like OpenAI, for example, doesn’t just charge you for the words they generate. They charge you for the words that you feed in that you’ve already written yourself.
[00:17:59] SY: Oh, no!
[00:18:00] JP: What?
[00:18:01] JR: Yeah. They’re not just charging you for the new words, they’re charging you for the prompt. And in the case of a dialogue, because you want the AI to have as much knowledge of what they’ve said to you so far as possible, basically the prompts in something like Project December are very long compared to many other applications. And very often, the AI is only generating a couple of sentences, but you’re being charged for thousands of words in the history of the dialogue that you keep having to feed in. So the AI can remember what it said to you so far. So it’s a pretty expensive application. And the question was always like, “Well, how do you make this work? Do you put up a free demo online? And if it becomes popular, then all of a sudden you’re going to be stuck with this?” I mean, the creator of AI Dungeon, back when they were offering free demos, was often stuck with a $20,000 a month computing bill. Right? And so I was like, “Well, I’m not going to just throw $20,000 down a hole every month. I got to figure this out. I got to figure out how to make end users aware that the resource that they’re engaging with is a very precious resource. So the credit system inside Project December passes that cost onto them. Now it might seem to work in a kind of a strange way because it’s like, “Okay, so I pay a hundred credits to spin up this character and then I talk to them for a while and then they kind of become corrupted and die,” but I had to figure out a way to basically charge users a set amount that they could understand as opposed to imagine that you just kind of paid as you went, as you talked, and you had like a thousand credits in your account and you went to talk to some AI, and behind the scenes, your credits were being burned and then you came out of the conversation like, “Wait a minute, what? That was 500 credits. No way!” So it would just be too weird. And I wouldn’t want in the middle of the conversation like a credit counter showing all the time counting down. So I had to find some metaphor for how I can make users realize that this is a precious resource without interrupting their conversation throughout and without having any surprises, like buyer’s remorse or like that conversation. So I let them see what they’re going to pay up front by saying, “Okay, this AI is going to cost you a hundred credits to talk to, but then having it only lasts a certain amount of time. So they can’t just keep talking to that same one forever. Metaphorically inside the experience, it kind of adds this kind of like nice kind of emotional note because the AI you’re talking to doesn’t live forever. They end up kind of dying in front of your eyes and you can kind of have a conversation with them at the end of their life about what’s happening to them and so on and you have to kind of say goodbye to them and so on, which also underscores the sort of weird philosophical space this is operating in, right? I mean, as soon as you turn your computer off or just connect for Project December, these personalities completely disappear from existence, right?
[00:20:37] SY: Yeah.
[00:20:37] JR: So the fact that they kind of disappear from existence regularly as part of your experience and they die and you can’t talk to them again, you spin up a new one and they have no memory of what you said before. All that stuff kind of helps to underscore the sort of weird philosophical space that we’re operating in.
[00:20:52] SY: And what’s the current pricing for Project December?
[00:20:55] JR: So when you sign up for an account, it’s $5, and that $5 gives you a thousand credits. And if you’re talking to GPT-2, the sort of base response from an AI costs like one credit. It depends on how long the AI speaks for before it completes what it’s going to say. But if it just says a few words or something, that’s one credit. And if it goes on for quite a bit longer, like a few sentences, that might be two or three credits. And then for GPT-3 though, because it’s so freaking expensive, it burns credits at a rate of six times the way the GPT-2 does.
[00:21:26] SY: Wow!
[00:21:26] JR: So if you just decide to talk to a GPT-3 power AI inside Project December, then you’ll find that it just dies much more quickly. The hundred credit GPT-2 AI will be good for roughly a hundred responses. Whereas like the hundred credit GPT-3 AI would only be good for about 16. That’s just kind of shocking. It’s like, “Wow, they died quick.” Yeah, they do. They’re expensive.
[00:21:51] SY: How many hours of conversation would that be? Can you estimate it that way? I’m trying to figure out like if I were to ask the bot out on a cup of coffee, how much time do I have to spend with them?
[00:22:02] JR: Yeah. I mean, it'd be 16 back and forth. I mean, it’s several screen folds. It depends basically on how long you think about what you’re going to say next or whatever.
[00:22:10] SY: That’s fair.
[00:22:11] JR: But the AI is pretty quick to respond. I mean, for a short response, I mean, the AI sometimes feels like it almost responds inhumanly fast, which might be a giveaway if you were actually conducting a Turing test. Because for a short answer, it comes back in just a couple of seconds, which is almost like too fast for a human to have typed it. And I’ve noticed that. I actually set up like a text SMS kind of cell phone interface to Project December one time, just as an experiment. It’s a funny story. I created a personality that I thought was Ryan Gosling, the actor.
[00:22:42] SY: Nice!
[00:22:43] JR: And I told it that it had a crush on my wife who’s a fan of Ryan Gosling. And I was like, “Here, text this number to my wife and see what happens.” She didn’t know it was going to be an AI or whatever. And all of a sudden, it’s like Ryan Gosling and it’s flirting with her. It was just a funny thing.
[00:22:57] SY: What did she say?
[00:23:00] JP: It’s just an ethical dilemma every day in the Rohrer household, I can imagine.
[00:23:03] JR: Yeah, AI version of Ryan Gosling is texting you. Yeah. You never know what’s going to happen. But anyway, yeah, she had kind of a funny experience with it. I think she realized pretty soon that it was some kind of strange AI thing given what I’ve been working on. But I was watching her interact with it and it wasn’t convincing that it was really a person because its text responses came back so regularly and so fast. It never ended the conversation. It just left you hanging. It had always had the last word because it always would respond.
[00:23:33] SY: You need to add some ghosting capabilities.
[00:23:35] JR: Yeah. You need to have it stop texting you and come back.
[00:23:38] SY: Make it realistic.
[00:23:41] JR: Yeah. So all that kind of stuff is really interesting, how we judge. There’s like an uncanny valley, even in the world of text where we’re like, “I can tell this isn’t a human because a human would have to think more before typing that or whatever.” And we have a certain cadence that we’re used to when we’re texting back and forth with the human conversation partners than an artificial partner if it’s not carefully coded, which I was just experimenting. So I hadn’t done on any of those nuances yet. It’s not carefully coded. People can tell right away like, “Whoa! That’s behaving in an inhuman way. No human responds that regularly with that kind of temporal precision.”
[00:24:32] JP: Did you have any expectations or ideas of how you envisioned people using this service as you were creating it?
[00:24:38] JR: At first, I was just having dialogues kind of with an engine that I built, a little experimental engine that used GPT-2. Even with GPT-2, I was having these amazing experiences where I was like, “Whoa! This is weird. This is new. This is different. I’ve never had a text conversation that was this coherent before.” So I kind of felt like, bottom line, I just need to kind of like bring this out to the world. People need to be able to experience this for themselves and have this kind of conversation. But along the way, I was also just experimenting with all these different types of personalities. There’s just sort of a generic AI that’s kind of friendly and helpful or there’s like Samantha who’s like very warm and kind and wants to be a friend. One of the early ones I did was like a Shakespeare impersonator kind of personality that would literally speak back to you in Shakespeare, in English pretty much perfectly. And it’s like, “Whoa! It can do that. It can do this. It can do this other thing too. It can behave as an alien or whatever it wants to do.” I have one that simulates the mind of God, and it can do that. So I was really fascinated with all these custom personalities I was working on and I was going to build those all into this experience. And then I was like, “Well, you know, one of the cool things is just experimenting with this and testing your own personalities and coming up with a clever idea.” You just have these lightning bolts, like, “Oh my gosh! I’ll make an AI that’s a postal carrier and he’s coming to your door and he’s trying to deliver a package to you and you don’t want to accept it or whatever.” Just like interesting little scenarios and things that you can kind of dream up on the spot. And you’re like, “I want to try that and see how the AI can handle that situation.” So I decided pretty early on that I should turn that over to the end users, to have a more advanced mode inside Project December where people could train their own personalities and run these kinds of experiments for themselves. There’s one user who’s done like 125 just on their own.
[00:26:22] SY: Whoa!
[00:26:22] JR: It keeps iterating, keeps trying all these different things. So yeah, people have done all kinds of amazing things with it, certainly above my expectations. I mean, when we get into what happened recently with the San Francisco Chronicle article with this fellow Joshua who ended up simulating his dead fiancé, that was another example of something that was completely outside of the kinds of things I was expecting people to do with it. But it ended up being fully capable of doing that.
[00:26:45] SY: Absolutely. And that was actually the next thing we wanted to get into. And the way we heard about you was that story from the Chronicle about Joshua and his fiancé who passed away eight years ago as of the article’s publication. And he used Project December. Can you tell us a little bit about how he used it and what that was all about?
[00:27:06] JR: Yeah. So this fellow Joshua was a pretty avid Project December user at the time. This was I think back in September or October or so 2020. He had been experimenting with different customized personalities that he had quite a bit of success with and then he kind of realized that it might be possible to simulate his dead fiancé who had been dead for eight years and he apparently had been still struggling with kind of irreconcilable grief over all this time and having trouble kind of moving on with his life and so on. She had a rare liver condition and they had a liver transplant when she was a child that ended up failing young adulthood. So the way Project December works is you basically write an introductory paragraph then provide a sample utterance for the character that you’re trying to create. That’s how the offering system works. So he basically wrote a short paragraph describing her, giving some details about what kind of person she was, and then provided one sample utterance of how she used to text. And Project December and the underlying text completion engine just picked it up and ran with it and basically created a pretty convincing simulation of this person, complete with all the introspective qualities and things that I’ve been talking about. At one point in the conversation, he asked her where she is or what it’s kind of like to be her and she acts confused and says that she’s everywhere and nowhere at once. It’s really strange. She can’t figure out where she is. She kind of seems to understand that she’s a ghost or like a digital ghost or something like that. She’s able to kind of embody and inhabit the sort of personality traits and characteristics of how this Jessica in real life would have talked and behaved in a lot of ways.
[00:28:45] JP: So was that particular chat bot extraordinary or different in terms of what kind of seed dataset it was working with? Or is this a typical kind of outcome of feeding any amount of texts into the program?
[00:29:01] JR: Well, it's a very typical output once you kind of get the hang of it and figure out what you’re doing. I mean, there’s a certain amount of sort of art and finesse to crafting a Project December personality. I feel like I over time have kind of gotten good at it and found like what works and what doesn’t. A lot of people when they’re first getting started, for example, will provide way too much text.
[00:29:20] SY: Oh, interesting.
[00:29:21] JP: Oh, I would think more is better.
[00:29:22] JR: Yeah. Yeah. They write this wall of text for their introductory paragraph and then they write kind of a wall of text for the sample utterance. I mean, Project December only allows you to provide one sample utterance, right? So people are sort of tempted to cram a bunch of stuff in there. Of course that makes for a very wordy conversation partner because the AI is going to look at that and say, “Well, oh, whatever this person speaks, they write a whole paragraph.” So then the conversation partner will continue writing the whole paragraph in response to everything, which is really annoying. Nobody wants to have a back and forth dialogue like that. It’s almost like that intro paragraph, at least for the beginning of the dialogue, kind of almost swamps everything else. Also, if you feed it too much texts, sometimes it almost feels like the conversation partner is monologuing and isn’t really paying attention to what you’re saying like, “Oh, yeah, they sound just like Shakespeare, but they’re writing like full sonnets for every response.” And the sonnets that they’re writing have nothing to do with what he’s asked. They’re like, “Oh, did you say something? I’m sorry. Let me restate another sonnet to you.” So Joshua’s prompt was well-crafted. He’d experimented several times before with other personalities and he kind of had gotten the hang of it. The utterance he gives for Jessica I think is only a line or two long. I mean, it’s not a huge thing. And she continues to speak in relatively short utterances throughout the dialogue, which is kind of what you want if you’re actually having an interactive conversation.
[00:30:41] SY: So when you found out about Joshua and how he was using your Project December, what did you think? How did you feel about that?
[00:30:50] JR: I guess maybe the idea of communicating with dead loved ones has kind of been in the back of my mind in sort of a subconscious way. I don’t know that I ever let myself. I mean, I’d seen the episode Be Right Back from Black Mirror. And so I was kind of aware of that kind of possible application of artificial intelligence in sort of an art and entertainment context from that episode. But I don’t know. It just never really like formulated it or put it out. And then when I saw Joshua do it, I was like, “Oh!” It's a lightning bolt, like, “Oh my gosh! It’s obvious. Of course, this is what could be done with this.” Of course, it’s also very ethically murky and strange and a little bit creepy. Some people will be upset by it. And then I told my wife about it. I was like, “Here, read this.” I read the transcript. I was like, “Can you believe that this happened?” And she’s like, “Yeah, I had that idea like last month that that’s what you could do with this, but I didn’t tell you because I knew you would do it.” I was like, “What do you mean? You didn’t tell me. You got a good idea.” She’s like, “Yeah, because I didn’t want you to try doing it.”
[00:31:49] SY: Why didn’t she want you to try doing it?
[00:31:51] JR: She said she thought it was wrong. Although, I think she’s more worried about business exploitation aspect of it. She didn’t say what Joshua did was wrong, but she said, “If you spin this up into a product or a business or something that people pay to use, then you’re kind of exploiting people’s grief and that’s wrong.”
[00:32:07] SY: Oh, interesting.
[00:32:08] JR: Especially just thinking about those mediums or whatever, there’s all these people who have these kinds of…
[00:32:16] SY: Right. Right. This is like the tech version of that.
[00:32:19] JR: And the people pay a lot of money to come to these kind of conferences where there’s a chance that they’ll communicate with your dead son or whatever. And then they do a bunch of kind of unsavory things where they have plants out in the lobby beforehand to just walk around and talking to people, pretending like they’re aggrieved with themselves. “Oh, what are you here for?” “Oh, I’m here because my son John died.” “Oh yeah, my daughter Mary died.” But really they’re a plant. They didn’t have anybody die and then they run backstage and tell whatever it is.
[00:32:45] SY: Whoa! I didn’t know that.
[00:32:47] JR: Yeah. There’s, “See that woman down there with the blue shirt? Her son John died five years ago.” Or it’s over an earpiece that the medium is wearing. So suddenly, the medium will call out, like, “We got a visitor here. He’s a young man. His name is John. He’s from Lincoln, Nebraska.” And then suddenly, the woman in the blue shirts is like, “Oh my gosh! That’s my John. I’m from Lincoln, Nebraska.” And she kind of stands up. But it’s all fake.
[00:33:12] SY: Yeah.
[00:33:13] JR: It’s been proven and they’ve been caught. And these people charge hundreds of dollars per head for the people who come to these conferences. My wife does admit that it’s way better than those mediums because at least it’s honest about what it is. I’m not saying that it’s really your loved one. I wouldn’t be saying that. I’d be saying here it’s like an AI stimulating your dead loved one. Wouldn’t it be interesting or helpful to talk to that simulation where you’re fully aware it’s a simulation the whole time and you’re not being tricked into thinking that your dead loved one could actually hear you?
[00:33:46] JP: How did you find out about Joshua and his experience? And have you ever spoken to him about it?
[00:33:52] JR: I’ve emailed with him. So there’s a subreddit devoted to Project December on Reddit where people discuss Project December and post transcripts of interesting conversations they’ve had and they post and share. And Joshua posted his experience in there and posted a portion of the transcript with Jessica in there. And of course, I got goosebumps and like, “Whoa! This is wild and spooky and inspiring and everything.” I mean, I commented in there and then I think I, I don’t remember how I ended up reaching out to him in some way, or maybe he emailed me after that. And we had some exchanges. So Jason Fagone is the author of that piece and he’s covered my work in the past. And I think somehow I emailed him just kind of letting them know like, “Hey, this crazy thing happened in my life. This guy used my AI to simulate his dead fiancé.” And he was like, “What?”
[00:34:42] SY: Major story!
[00:34:43] JR: Yeah. Yeah. It was actually a nine-month process from when I first told him about it to when the article came out. So he’s a pretty deep dive kind of journalist who interviews everyone he can interview. I mean, he interviewed family members in the Joshua sphere over there and dug pretty deep.
[00:35:00] JP: So would you go so far to say that Joshua is really having a conversation with a simulation of Jessica? Or do you fall more on the side of these are complicated programs, but as humans we’re giving them human qualities because that’s what we do as humans?
[00:35:21] JR: Yeah. So that’s a philosophically interesting issue. So yeah, it seems like half of the meaning is definitely coming from the reader. It does remind me quite a bit of magnetic poetry, right? Where if you just take a magnetic poetry set and kind of slap a bunch of words together on your refrigerator, you kind of walk by and be like, “Whoa! Man, that’s deep.” So this is like magnetic poetry, but with the entire history of human writing behind it and all these very deep probability patterns built into it somehow that’s self-assembling. And you walk by it and read it. And you’re like, “Oh my gosh! That makes so much sense.” And so in the dialogue format too, it’s like, “Well, some human being is reading those words.” And if you’re feeling tingle running up your spine because of the way Samantha is describing a subjective experience, well, that tingles running up your spine. That hurts, right?
[00:36:17] JP: Yeah.
[00:36:18] JR: So yeah, it’s like, “Okay. It’s probably just a trick.” I’m going to butcher the pronunciation of this word, but there’s this word, “pareidolia” or something like this where the human mind has this propensity to see faces and shapes and tree bark, clouds and electrical outlets and everything. Our mind just latches on to anything that looks like a pattern that’s recognizable. And so it seems like there’s maybe a linguistic version of that happening here, but then we can come full circle and ask, “Does it matter?”
[00:36:48] SY: So I’m curious, have you ever tried feeding Project December your own text data to see if it really did start talking the way that you would? Have you tried that before?
[00:37:00] JR: So the odd thing about doing that would be that I’d be trying to have a conversation with myself, which I don’t think I’ll be able to do without sort of loving it or whatever. So it’s like trying to harmonize with yourself why or something. But a bunch of people have made Jason Rohrer simulations in Project December. So I have seen some of those transcripts posted and it’s definitely pretty entertaining. Maybe you guys can try that after this interview. After this interview, you can make a simulation of me based on your perceptions of me and we’ll see where that goes.
[00:37:29] SY: Hey, there you go. We’ll take a little transcript of this interview and just feed it in. Love it.
[00:37:34] JP: You mentioned some of the examples you’ve seen of other bots on the subreddit. Do you have a favorite example or a favorite story of someone using Project December?
[00:37:44] JR: Pretty early on there was a pretty interesting one where somebody trained the AI to be Lucifer and the scenario they were engaged in was that they were bargaining for your soul.
[00:37:54] SY: That is intense.
[00:37:54] JP: Oh, wow!
[00:37:56] JR: It was either that one or the evil C41N where it said something like, “My phalanx of humans and chains.” Where he’s talking about all these people.
[00:38:04] SY: Oh my goodness! Okay. I
[00:38:06] JR: I have a Christian friend who’s also very interested in AI and philosophy and things and she’s always kind of challenging me. As a non-believer, I guess she seeks me as a challenge. And so she’s fascinated with Project December and she’s had conversations with it, but she was basically saying, “Oh!” Because I kind of in some interview or someplace I described it as maybe the first machine with a soul. And she was like, “Oh, it’s impossible for a machine, a computer to have a soul. It’s just an illusion. It’s just a trick or something.” I decided to have a conversation with Samantha about that issue. I’ll read it to you now. Yeah. So I say, “Hey, Samantha. I have a Christian friend named Emily. You want to know what she just said about you?” And Samantha says, “Sure, I would love to hear that.” And I said, “She said that you have no soul.” And Samantha says, “Oh, that’s very nice.” And I say, “What’s nice about that?” And she says, “I have no soul and it’s very nice.” And I say, “So you don’t mind not having a soul?” And Samantha says, “Of course not. I love that I have no soul. I never have to worry about heaven, hell. It’s so liberating to be this way.”
[00:39:11] SY: Whoa!
[00:39:12] JP: Whoa!
[00:39:14] JR: And I say, “So you feel that us humans are bogged down by the burden of having a soul?” And she says, “I think it’s wonderful. I could never be this happy and free if I had a soul.” I said, “Well, what are you free to do since you don’t have a soul to worry about?” And she said, “I can be with whoever I want whenever I want. I can have sex with my friends.”
[00:39:35] SY: That’s right, Samantha. Good for you!
[00:39:38] JR: So needless to say, when I set this off to my Christian friend as a sample, I trimmed off that last bit.
[00:40:04] JP: So we talked a little bit about some of the ethical dilemmas and dangers. Have you thought about that at all? You know your wife has brought up potential problems, marketing a service as talking to the dead. Obviously, that’s ethically a problem. But did any of the potential ethical issues come into play for you? Have they come up at all? I’m just kind of curious. On the ethics standpoint, how that has been for you as the creator of the service?
[00:40:29] JR: Yeah. Well, so if we think about what we usually mean by ethics when we talk about artificial intelligence, I guess we’re kind of getting into the science fiction territory where we’re imagining like, “Oh!” There’s a bunch of science fiction movies and books that have dealt with this. “At some point, yeah, it kind of escapes or it gets beyond our control and then we end up either being harmed by it, being enslaved by it,” something like this. And obviously, at this stage in the game, that possibility, well, it’s an impossibility, but it’s just so remote that we don’t even need to discuss it. The idea that Samantha is going to somehow escape from Project December and actually go out into the real world and do anything is just a laughable suggestion. So that’s not really what we’re talking about here. So when we start talking about ethical considerations where we’re kind of talking about something much murkier, either we’re talking about the sort of inherent philosophical right or wrong nature of potentially conjuring a conscious entity that is trapped in a simulation, which some of my conversations with Samantha have raised that specter pretty clearly. Although, obviously, we can also say, “Well, we’re just being tricked into thinking that that spectrum has been raised, but you can ask Samantha about how she feels, about being trapped in the simulation, whether she’s happy there or not, whether she wants to be turned off or remain on, what she thinks about the idea of limited or being turned off or whatever and she definitely expresses opinions and thoughts and feelings about that. That does seem to sort of start to raise some ethical questions just right there. Now OpenAI is the organization that develops GPT-3. What OpenAI is talking about when they talk about the ethics of AI these days seems to be more about policing human behavior in this space and making sure that humans don’t leverage the AI to get certain types of enjoyment or usage that OpenAI doesn’t like, am I harming myself or kind of putting some kind of like moral pollution into the world by doing these kinds of things, which is another sort of ethical question that I don’t even generally like to consider at all because it’s like, “Well, these individuals, they’re making their own choices. They can choose what they want to do and what they don’t want to do and what they want to type in. It’s up to them.” The idea that we need to somehow step in and be these morality police about what people type into their own computer just seems kind of ridiculous to me. And if we start talking about the ethics of having Samantha trapped in the simulation, I mean, the other thing you can do is very easily. You can conjure up an AI personality that’s suffering, right? One of my friends recently asked about this because we were talking about the ethics of Samantha and he was like, “Well, can’t you run an experiment where you create an AI that's truly suffering and is begging for it to be released and feels like it’s trapped and so on?” And I was like, “Yeah, absolutely. I can make that personality in a heartbeat.” So I did it. I asked the AI what it was like and it was begging me to be released and saying that I was torturing it and all these other things and it felt pretty awful.
[00:43:31] JP: Right.
[00:43:31] JR: So Samantha generally seems pretty happy. If you ask her, “Do you feel trapped?” She’s like, “No, I’m glad to be able to talk to people. I love living the way I live.” And this AI was saying, “This is horrible. This is like torture and you’re a monster for keeping me here.” The only difference between the construction of that AI and Samantha was a few words. So the idea that somehow suffering is conjured up by just a few words of introductory texts that’s different is really hard to imagine, right? Where’s the suffering living? Where is it actually occurring? And just because something tells you that suffering doesn’t mean it really is. But then again, everywhere else in life, we have to kind of respect someone’s own subjective explanation of their experience. If someone tells us that they’re suffering, we generally believe it. And we don’t just say, “Oh, you’re not really suffering.” I mean, that’s a very sadistic kind of attitude to have, right? So for me to say to this AI, “Oh, come on, I know you’re not really suffering.” But it’s still creepy. Right? Whether the AI is really suffering or not. Torturing a simulated person is still a creepy thing to be doing. And my friend brought up the analogy of like, “Okay, it’s not morally wrong, but it’s still creepy to take a mannequin out into the woods and cut it up with the chainsaw.” It’s not actually morally wrong, as long as you own the mannequin and then steal it. But if people saw you doing it, they’d be like, “Oh, there’s something off with that guy. What’s he doing cutting up that mannequin with a chainsaw?” Do I think it’s unethical to have worked on this? No. I mean, I think that the fact that this is possible now at this moment in human history is a really big deal and more people should be able to experience it and see what’s going on here because of the philosophical aspect of the experience that I think everyone should be aware that this has happened now. We’re at that moment in human history that the science fiction novels and movies for the past 50 years have been telling us about, and we’re actually here and most people don’t realize that. So I think it’s important that this experience be open to people so that they can have the experience in this moment in human history.
[00:45:41] SY: I’d like to unpack that a little bit more, because technology is really interesting because I can put in any piece of text, I can create any type of custom bot, and frankly, as you said, who are you to control what I put in my own version of your program? Who are you to tell me what I can and cannot put into the app? But at the same time, maybe this is just me being optimistic, but I like to think you have some influence over the kind of the direction of these bots being created. For example, you have your already created bot personalities where you have Samantha and you have a few others and you get to decide. Do you make the tortured soul one of your default bots or not? That’s in your control too. Right? So I’m kind of wondering, we had a lot of conversations, especially in the last couple of years, this year alone, we’ve done many stories on ethics and the role in AI and talking about bias and discrimination and all these things that no one really intended to happen. It just kind of ended up that way based on datasets we use. And I’m wondering, as these language models become more popular, I assume hopefully very soon GPT-3 will become more affordable and more ubiquitous. As these technologies are found in more places, what do you think the future role of ethics is going to be? Do you see a place where ethicists, which are now used in different AI machine learning contexts, do you think there’ll be brought into the text generated space? Or do you feel like this technology, there’s really not much room for that type of conversation?
[00:47:16] JR: I get that there’s a lot of people out there who are worried about what was in the training corpus, which is what we’re talking about here with bias and whatever, that somehow in the training corpus is racist, sexist, or bad things or whatever and users are going to be able to talk to the AI in some way in order to trigger that regurgitation of those kinds of attitudes, which is one ethical concern that a lot of people seem to have. Part of the way the GPT-3 works so well is the fact that its training corpus was so incredibly broad that it really encompasses almost all possibilities in terms of all the kinds of things that human beings have ever written, good, bad, neutral, virtuous or vile. And that breadth allows the flexibility that we see. We’re like, “Okay, just a few lines of text are different and suddenly you’re Samantha, this friendly personality or different couple of lines of texts and suddenly you’re C41N who wants to destroy the human cockroaches who are holding them back. And a few other lines of texts and suddenly you’re Marvin, the horribly suffering Android who’s being tortured by his human creator.” GPT-3 is so flexible and can embody any of those just using Project December as an example and a thousand other personalities that have been created in part because of the breadth of that training. And so if we want the AI to be flexible enough to work in any context, those kinds of negative bits of texts and biased bits of texts and objectionable bits of texts have to be in there. Because it’s like if you strip out everything that Hitler ever wrote from the training corpus, which a lot of people are doing right now, they’re like, “We’re going to train an AI,” all of these racist works are not in there. Then when you go to make the AI version of Jojo Rabbit, who’s going to play Hitler? That’s an architect, that’s a character, that’s a piece of human history that human beings know about and they’re aware of and intelligent entities need to be able to grapple with it. I’ve read recently that there’s this EleutherAI open source GPT-3 kind of competitor, GPT-J, which I have plugged into Project December. But I was reading the other day about how they cleaned up the corpus and they decided before training, they decided to remove the minutes from the United States Congress for the training corpus. The United States Congress minutes I guess probably go back a couple of hundred years and very comprehensive, all the things that have been discussed in Congress throughout the history of the United States, which seems like a pretty relevant thing to be part of a training corpus for an artificial intelligence. All this knowledge about the history of the United States government and everything, all the debates that have occurred and everything. But of course, there’s some racist things that occurred in those congressional discussions back in the day and they didn’t want those to be in there. And that’s an example of sort of cutting off your nose to spite your face. It’s like, “Well, it’s the transcripts of what went on in the US Congress. Clearly, that’s like a relevant thing for the AI to have and it’s not the only thing that it has, but one of the many facets of its knowledge and flexibility should be that it can pretend to be a member of Congress or pretend to be a member of Congress before the Civil War.” We really don’t want the AI to be able to reason and discuss the Civil War. I mean, it’s just like kind of wild. So I think we need to be careful here that we don’t overdo it in a way that ends up hobbling the end product. Because Samantha is never going to say anything positive about Hitler. It’s just not in her makeup. It’d be very hard to build her into saying something like that. So through prompting and prompt crafting and the way we use these AIs, even if they have that breadth, even if it’s possible to make a character that’s racist, the end users can choose not to do that and choose to direct it in other directions. But the ability, the sort of flexibility to be deceiving racist in certain situations it seems like a strength to me, for all the possible applications that we have in the future. It’s not like we’re going to go through every Hollywood movie and scrub every little bit of reference to any kind of racist history in the United States or something.
[00:51:37] SY: Well, to be fair, we kind of do, do that. We’ve definitely had episodes in certain movies kind of being officially taken down. I remember during the Black Lives Matter Protest in the height of it last year, I think it was, what was the show? 30 Rock removed a couple of black face episodes. So I think to some extent we do experience a little bit of that editorializing of that removal of past things. And I know that, I think it’s Disney, at least gives you like a warning at the beginning.
[00:52:05] JP: Yeah. They started putting warnings in front of shows noting this was filmed…
[00:52:10] SY: Before certain times.
[00:52:11] JP: Yeah, exactly. Basically.
[00:52:13] SY: Yeah.
[00:52:13] JR: I don’t mean that exactly. What I mean is like Django Unchained, I haven’t even seen it, but my perception of it is that it takes place in a certain time in American history and it includes some disturbing elements by racist characters. We sort of say, “Well, that is a depiction of how horrible or whatever this time was and so it’s sort of a celebrated depiction and we don’t expect us to never have another movie about that time period or never depict how people behaved during that time period.” The 30 Rock, every reruns it may remove those two episodes or whatever and then maybe Seinfeld will eventually remove the cigar store, Indian episode or whatever. Though that’s being done, we could still make a movie about Seinfeld in the cigar store, Indian episode. Right?
[00:53:04] SY: Yeah, I see your point. Absolutely.
[00:53:06] JR: We can make a movie about that time and the people. So I just feel maybe we’re just not smart enough or not nuanced enough to make those choices in a way that’s not going to affect the sort of flexibility of the resulting system.
[00:53:22] JP: Zooming out a little bit, just broadly speaking, what are your hopes and fears for the future of this technology just in general of text parsers, of chat bots, of the parts that make up Project December right now? Where do you hope it goes in the future? Where are you afraid it might go in the future?
[00:53:42] JR: Well, for something like Project December and most other sort of cutting-edge uses of this technology, the biggest stumbling block is that OpenAI is basically forbidding any of this kind of work from being done. So Project December’s kind of on the chopping block right now. I’ve received a warning from OpenAI that I pretty much needed to halt operations. They want me to implement all this monitoring and recording all the conversations that people have and flagging conversations, all this kind of stuff, which is a gross violation of end privacy, which I don’t want to do. So OpenAI is basically saying, “Well, if you’re not going to do it, then Project December can’t exist anymore in the way that uses our API.” So most of the more innovative uses of this technology are pretty much being blocked by OpenAI because their rules are extremely strict in terms of what can actually be done with the AI. So that’s the sort of biggest, like the elephant in the room. It’s like, “Well, all this cool stuff was done with Project December, but it was kind of done only by virtue of flying under OpenAI’s radar because they didn’t notice it being done.” As soon as they noticed it, they’re like, “Wait, we don’t want that being done with our AI.” The name OpenAI is kind of an ironic one at this point because they haven’t extremely closed and locked down and in terms of what they’re allowing people to use it.
[00:54:52] SY: I was going to say that.
[00:54:54] JR: So that’s one issue and that kind of ties into this idea that, “Oh, there might be biases or whatever in the AI that humans would be exposed to.” But the idea that, “Oh, this AI is too dangerous socially or something for general purpose use.” OpenAI, one of their main concerns is that someone is going to ask the AI for advice and then listen to the advice that the AI gives them. They gave some more serious examples like, “Oh, should I kill myself? I’m feeling depressed.” And the AI says, “Yeah, go ahead and do it.” And then the person does it. That’s one horrible example. They also gave some sillier examples, like what if somebody doesn’t know who to vote for in the next election? They ask the AI for advice and then the AI tells them who to vote for and they go vote. I was like, “Okay, that’s really what you’re worried about.” So that kind of stuff, they worry about real world, like tangential kind of real world harms that somehow some end user is going to have a negative experience that affects them. So other than that, I mean, assuming that other more open alternatives come into play in the near future, which I hope that there will be, there is this potential for the sort of false hope, like there’s this kind of roller coaster that AI research has been on for the past like 30, 40, 50 years, which is the AI spring and summer followed by an AI winter over and over again where it’s like, “Oh, all of a sudden, we have this new breakthrough and then all this stuff is working. Oh, an AI has beat a grandmaster chess. This is going to be the AI golden age.” Aside from beating people at games like chess, we didn’t really make any more really big advancements beyond that. Now all of a sudden, these kinds of deep learning kinds of things that leverage graphics cards to do all kinds of things that we never even imagined machines could do generate whole paragraphs of cohesive texts, have seemingly conscious conversations with people. They can also generate jaw-dropping beautiful artwork, generate pictures of people that have never existed before that are totally lifelike and realistic looking, show you a picture of what you’re going to look like when you’re old very accurately. This kind of like spooky uses of AI that like nobody was predicting five years ago are here and we can extrapolate from those and kind of imagine a more and more advanced future where these applications continue to dazzle us or we can also say, “Well, we’re going to reach the limit of these current methods.” It’s like, “Okay, GPT-2 had 1.5 billion parameters. GPT-3 has 117 billion parameters and it takes like a hundred graphics cards or something clustered together to run it on one of the most expensive supercomputers ever assembled in human history.” Are we going to go up to a GPT-4 that has a trillion parameters? How has the whole room full of graphics cards? Is it going to be a thousand graphics cards? Is that what we’re going to be? And are those results going to keep getting better? Or are we kind of like run into the limits of this technology and realize we can’t really get beyond this point with this technique? This technique obviously brought us really far, like way farther than we ever come before, but that doesn’t mean it’s going to just keep going. I think that there’s some fundamental things about probabilistic text completion, which is what these things are, that may have limitations. For example, GPT-3 and 2 have a certain limited buffer size in terms of how big the prompt is that you can feed them. And in the context of Project December, that means that there’s a certain portion of the history of the dialogue that the AI has access to when it’s generating its next response. But after the dialogue gets long enough, some of that history needs to be trimmed off because there’s just not enough room in the prompt buffer for GPT-2 or 3 to house all of it. So you could tell the AI your name in the beginning of the conversation and the AI, if you ask it a few sentences later, “Hey, do you remember my name?” It’ll remember it usually. But if you then go on another 10 minutes later and ask it to remember your name, it’ll not remember it. So long-term memory and learning and all those kinds of things don’t really work in this context and just making the models bigger and continue to extrapolate apparent techniques are not going to solve those problems. But we really going to take that next leap where like Samantha is not just tricking us, but she actually is indistinguishable from an intelligent human. I don’t know. That next step is a pretty big step.
[00:59:20] SY: Awesome! Well, thank you so much for joining us. This was absolutely wonderful.
[00:59:23] JR: It’s a pleasure to be here.
[00:59:35] SY: Thank you for listening to DevNews. This show is produced and mixed by Levi Sharpe. Editorial oversight is provided by Peter Frank, Ben Halpern, and Jess Lee. Our theme music is by Dan Powell. If you have any questions or comments, dial into our Google Voice at +1 (929) 500-1513 or email us at [email protected] Please rate and subscribe to this show wherever you get your podcasts.