CJ Avilla

ChatGPT, LLMs, OpenAI, doom or boon?

Show Notes

In this episode, Colin and CJ discuss using the new GPT tools for work and play.
  • Open AI
  • Large language models
  • Generative AI
  • How do we use it, and is it going to be doom & gloom?
How We Use It
  • Brainstorming & riffing on ideas 
    • I was building something in React and hadn’t touched React in a few years. I had all this iteration going on in my components to generate dynamic content. Still, I wanted the topmost component to be able to have references to all these components to download them all as images in a zip file.
  • Assistance with math-based things like animation or rendering graphics 
    • Example: I had to display a bunch of circles in a 2nd plane in a random pattern
    • Computers are pretty good at that
  • Learning how to do something new faster 
    • Explain this to me like I’m 5
  • Summarizing
  • We use Descript for editing this podcast
  • Shifting the tone of writing, building a writing style
  • Writing video scripts
  • Finding obscure information
  • Story arcs
  • Brushes to write tailwind classes
  • Jokes
  • DnD character sheets
  • Bing to find esoteric star wars character that’s buried on Wookiepedia
Doom and gloom or boon?
  • What does this mean for the future?
  • Privacy & security concerns 
    • Voice clones and scams (create a safe word!)
  • Productivity + time 
    • Are we just going to be expected to be more productive with more output because cApItaLIsM
  • Constitutional AI (Anthropic) 
ChatGPT coauthored a PR to a major open-source project
  • via Ruby Weekly: How long is it until ChatGPT makes a pull request to Rails? Oh, it (sort of) just happened. Ultimately it's still under the name of Rails core team member Guillermo Iguaran, however.
Everything in this pull request, including the code, tests, changelog, commit message, pull request title and description has been created by ChatGPT with some guidance. If you believe ChatGPT is infringing your copyright please let me know.
People to follow
(Written by Notion AI)
Here are 5 social media influencers to follow to stay up to date about AI:
  • Rachel Woods (@the.rachel.woods on TikTok, @rachel_l_woods on Twitter)
  • Logan.GPT (dev advocate for OpenAI)
  • Swyx (shares a lot of thoughts on AI)
  • Lex Fridman (MIT researcher and AI podcaster)
  • Siraj Raval (AI educator and YouTuber)
Tools
Github Copilot
  • Copilot X for VS Code & VS Extension
  • Copilot Docs
  • Copilot CLI
  • Copilot Voice
  • Copilot for PRs
Resources
Full Transcripts

Colin: Welcome to Build and Learn. My name is Colin.

CJ: And m cj, and today we're talking about large language models and all this AI stuff that has happened in the last four months and. We were just talking about how the episode we just released, we recorded four months ago, and so it'll be like entering a time warp and we're gonna just fast forward four months and an incredible amount has changed,

Colin: Yeah, little how the sausage is made. I was doing the editing on that last episode and, I was listening to it. I'm like, oh my goodness. I know this was one of our pre-canned episodes that we recorded so that we had some buffer. And I'm just listening to, this was in November. Things like Twitter getting bought by Elon and ChatGPT came out and we're like, yeah, we use it to summarize stuff. So what it's, it's cool, but so much has happened in four months and I think there's a big conversation going on about whether or not this replaces jobs or does it just increase productivity? And, we've seen these similar cycles, back when the calculator came out and the computer and the digital camera. And so I think having a chat today about how we use these tools, will be pretty fun to do. We can dive in.

CJ: Totally. GPT-3 was pretty good at generating text, but it is nowhere near as good as GPT-4 and I think when we recorded, maybe chat, GPT hadn't been released or, I can't remember if it had or hadn't, but

Colin: was like the release

CJ: Okay.

Colin: November.

CJ: Got it. Got it. so yeah, I think the leap between three and four has just been like mind blowing to everyone who touches it. and yeah, I think also like when you asked. GPT-3 things. Sometimes the content you got back was like, okay, whatever. Maybe that was like a, a response that you would expect from a third grader, fourth grader. And the stuff that you get back from GPT-4, I feel like is, senior in high school with a professor editing the responses. And it just makes it so much more useful.

Colin: And the difference between them being that four has been trained on more data, right? More language, more data. And so you can see what happens when you add more layers and More just data sets to it, that you get a more complete answer. we'll leave a link to, the Wolf from Alpha website about how ChatGPT works under the hood, and it's pretty amazing that it's literally like statistically detecting the next most probabilistic word. And so it's just guessing word by word. And when you have more words to pull from and more examples of those words in sentences, Papers and things like that, it gets really interesting. The thing that I'm most excited to see is what happens as this gets trained on proprietary data sets or like specific verticals like legal docs, constitution, all these like different things like court precedents and stuff like that.

CJ: I think companies are starting to just think about how they can use it to do weird things that you wouldn't expect. It's not just generating texts, but it's, yeah, making informed decisions about things and. I don't know. It's crazy, but I'm, it's, I'm excited to hear how you're using it cuz every time I talk to someone who is in tech, who is using it for their job or just using it to improve their life, I pick up something, I'm like, okay, I'm gonna do that for sure. So yeah. what maybe what is like a recent thing that you're like, okay, this is useful and.

Colin: Yeah, so I've mostly been using it for brainstorming and kind of riffing on ideas. if I have, I don't know, like I think when you name something, you always have to go check a bunch of domains to see if they're available. I don't think ChatGPT can go check like whether or. something's available for domain registration yet, but that should be a thing it should be able to do. we need a plugin for that. But if you're coming up with a name for a project, you're always like looking for what's a good name, what's available as it as it.com, dot dev, whatever. and then you, there just might be words like, I'll go to the thesaurus or dictionary, but now I can just be like, all right, ChatGPT. Let's have a little back and forth whiteboard session on this is the app idea. These are a bunch of root words. Give me a bunch of tangential things. and it comes back with some stuff that you just wouldn't have found. In Code I, I use GitHub co-pilot, so I've that's become a little too normal for me. And I'm really excited to see what comes in the new, like when it's backed by a four. Cuz I think right now it's just three, 3.5. but the most specific thing that I was blown away by was in React when I was trying to do something. It was like a component with a bunch of child components. And I write in Ruby most days. So I'm not just like I don. Know all the ways that you would use react off the top of my head. And so I just explained what I was trying to do. I was like, I want that thing to know about all these other things. And it gave me such a good example where I was like, okay, that makes sense. Why didn't I think of that? And I was like, you know what, we're just gonna do something very similar to that. And it was like one of those Spotify style, Spotify wrapped, we generate a bunch of images and we want to be able to download those images as a zip file. And so I needed a reference to all of the images in order to loop over them and download them and put them into a zip. And it was able to do a lot of that for me. So it was pretty, pretty cool.

CJ: That's super cool. Brainstorming ideas has definitely been a huge part for me too. one, a project I was tasked with recently was like, help prepare for a presentation that's gonna happen at a conference and tell a story that fits in lots of different Stripe products into the story. And so I was able to just give it a bunch of context. The presentation and say, it should highlight X, Y, and Z features, create a company that will illustrate how you can use all these different features to do these things. And it spit out these story arcs with company names and, character names of the, the employees and the CEO and how they came up with the story. And like all of this stuff that you're able to really just like flesh out into this beautiful story. And you could even tell it, or I was. I don't know what the common story arcs are. So it start out by saying what are the five common story arcs? And then say okay, what is a good common, what is a good story arc for a presentation? And then it tells you, one, you're like, okay, write a story about x where it includes these different characteristics. And it just, yeah, it helped so much. The brainstorming piece is really cool. When approaching a blog post that I need to write, I might come up with an outline that has three or four steps and. I'll say okay, this is what I think I'm gonna write for the outline, and then I'll just go ask ChatGPT, write the same outline. And it will sometimes think of other steps or like it will have different arguments for or against something or different pros and cons that I hadn't thought of that you might come up with if you're brainstorming with a group of 5, 6, 7 people with lots of diverse experience, from different areas and like really interesting diverse perspectives. And Yeah, it definitely helps fill in the gaps as just a single person who's creating content,

Colin: Yeah, and I think that's the thing I keep hearing is, Especially now that companies are trying to do more with less and things like that. Is that it? It's the Ironman suit. that you're, that I think GitHub co-pilot was designed and why they called it GitHub co-pilot was that they still want the human in the seat. And that, that you're a director of sorts. So you're producing what the story is you're doing the Hero's journey of the Stripe Conference, right? what does that look. And, that you're giving it inputs and it's able to go look at all those kinds of things without replacing a person. And I think there's a lot of people, I think the media has mostly picked up on this of I think my mom called me and was like, oh, aren't you worried you're gonna get, your job's gonna get replaced by ChatGPT? And it's whew, we got a long way to go, thankfully.

CJ: I think the fact that it can generate code is pretty, impressive. And one trend that I've noticed that's really interesting, if you go in discord, in the Stripe discord in particular, people will drop in there and say I don't understand why my code isn't working. Can you help me like figure out how to integrate this API or whatever. And then we'll say okay, can you show us your code? And they'll show it to us. And we're like, That is super weird. Like, where did you find this? That is like not correct. or it's like odd in a certain weird way, and they're like, oh yeah. I was just like using GPT like chat, GPT to write all of this. We're like, oh my God. like a lot of people are already just using it to write a ton of code and they don't fully understand all the details, but. That's like proof though that I, that you still need humans involved to check, check to make sure that it's behaving as you expect,

Colin: Yeah, I am excited to see what the new GitHub co-pilot X stuff does, because it seems to take that, like the idea that we could point ChatGPT at the Stripe Docs intentionally. And say, answer these things for us means that it's not making things up from the rest of the internet. It's making it up based on what's in the docs. and it has context like even now, like when you talk, when you go to the ChatGPT website, it doesn't have the context of your code unless you copy and paste it in there. And so being able to highlight something and say, explain this specific thing, or, Iterate over all of these things and add them up together, things like that becomes really powerful. but you still need to know what you want to do. Like that part is still important. It's not go build 10, dog walking apps in 10 different languages. Because there's gonna be a whole lot of inputs that, I'd be curious to see what those apps do, generate from scratch. It's okay, do you assume user login? Do you know what a dog is? Do you

CJ: Yeah, there's still like leaps that it needs to make in understanding in order to connect dots between different systems and things like that. one of the things I was using it for early on was helping to write video scripts for videos where we're teaching people how to use stripe basically. My normal process was start from the Stripe documentation and then go through, build out a demo, and then walk through and write out by hand what I, or like roughly what I plan to say. Now I can just take the whole thing, give it to ChatGPT and say write me a video script that helps teach this in X, y, and Z ways. But I just saw this really mind blowing demo by Siraj Raval on YouTube. So we'll link to it. But he basically built an entire YouTube channel with an AI influencer and so it like will go on Twitter and find trending AI books, then it will. write a script that will pitch the book. Then it will like generate video. Of a fake person saying the script, and it will u it'll generate audio. It like generates their voice, it generates animations of the concepts in the video. And, all of this isn't just like a one. He couldn't just say ChatGPT make me a automated YouTube channel. Instead, he needed to know okay, here are the like 25 tools throughout the video, he uses like tons and tons of different tools. And Python libraries and, APIs. So he is using like the shot stack API to compose all these different pieces so that they show up the right way. But at the end of the day, he was able to create these. These, videos that are being automatically published to YouTube and are based on trending books that are on Twitter. And the idea was like, okay, now I want to take my affiliate link for those books on Amazon. Put them in the description so that you can like maybe build up this passive thing. So from my little idea of help me write a script for a video. It's like people are able to use these tools in different ways to combine them and build like these really wild automations and systems and, yeah, I don't know, like basically like the, I feel like the way that I'm using it is still very much just scratching the surface of what you can do with even just the stuff that's available today.

Colin: I'll say, I haven't seen the one that you just talked about, but I will say that the difference it sounds is that you use the tool like a research assistant, and then you're still the person who's doing the video. And I have to wonder, we talk about this a lot with our show, like people wanna follow people and I feel like computers and AI are not gonna make this. It's actually gonna make it more valuable. Real conversations by real people. Do I. And trust a book review from an AI robot that's just meant to be an affiliate link farm. Or do I want to hear what c? Thanks about this book. And you can still use this to do summaries and to do research and, come up with a bunch of other books that we should read. But, it's similar to when, wizards of the Coast and Dungeons and Dragons was talking about, they're already seeing people using their IP and NFTs. And AI dungeon mastering and things. And a lot of people were worried that they're gonna focus on creating some sort of AI DM and replace dms when I think people forget why we played Dungeons and Dragons in the first place. I think there's been a meme going around, we're doing all this work and we're giving AI the ability to do art and music and things like that, right? And it's we should be doing the art and music and let the AI do the work. we should flip that around. I think there's really cool things that, there was a really good blog post about someone who had chat, GPT Run a Dungeons and Dragons game for them, and it was cool, but I still don't think I would necessarily be threatened by that or feel like I'm replacing my hobby with an AI for that.

CJ: I guess we're like slipping into doom and gloom territory a little bit, but like I, yeah, I'm so conflicted. Okay, so this is coming from, my perspective as a parent. my kids, they're super into D&D and I am really conflicted about letting them use ChatGPT, because I'm worried it will take away their. And they'll start to lean on it as like entertainment and they'll lose their creativity like, before seeing ChatGPT, they would sit down and have piles of books across the dining room table that they borrowed from the library and were building their own character sheets. And they have like dozens of these. They're floating around the house. Many of them make it to the recycle bin. But, we sat down and then, I was like, okay, let's, let me just show you some of this ChatGPT stuff. We did jokes, we did poems, lots of fart stuff, lots of Mario Brothers stuff. But then we like, we generated a D&D character sheet and we like gave it some really specific things and my son was like, oh yeah, I wanted to have these cool weapons and I want to have this and that, whatever. And it generated this really impressive character sheet with like custom. Weapons that have, like cus I dunno, it's like damage or something. I don't remember like all the details of what they do, but it was like super, super customized and really comprehensive and complete and so I think that's cool. But I also am like, okay, I want them to be able to use it and get ahead by using it, but I also don't want it to steal their creativity and I don't know what the answer is to that. I also like when it comes to art and all like the creative pieces of my work. at work, right? Like when I'm doing my day job, I actually don't mind if it's doing the creative work. Like I don't wanna have to think up images and I don't wanna have to think up like these story arcs and have to have all the details. Like I'd rather the, AI do that for me. And then on the weekends or like when I'm not. Working for money, then I can spend my time thinking about my creative stuff or drawing, or how do I want to have a different creative outlet. And so just the tension between those two things right now is really, yeah. It's something that I know like I, I'm having an internal battle about.

Colin: Yeah, I think this is the, old man yells at cloud segment of the podcast, right? Where, we both are like, we can see how it can be used for good and see how it can be used for bad. I think that the other side of that would be, it could make you more creative because you get. To iterate through lots of different versions really quickly. I do feel for artists who have their work has been trained. Into these systems where it's if someone was gonna hire someone for a portrait before, the odds that is gonna happen in the future, go down. and I, there was an article going around from someone who couldn't, the time to create a character in a game went from weeks to days and they don't feel as creative anymore. But also now that means that you can. A whole bunch, like AI can create a whole bunch of crappy characters, but it still needs some input like we're talking about. And so you might not get, have enough time or skill to pull off something 10 different ways, but with this you might be able to, and then pick the best one, improve on it. So you get a lot of that iteration on it. I think the thing that I've been using it for the most, other than the brainstorming has been like having it teach me And the challenge there is the hallucinations are real in ai. So you can't take it at its word all the time, but if you keep your curiosity and you keep asking questions, it's very interesting how it behaves. So like I gave it, a prompt in 3.5 and in four just to see the difference. And I highly recommend sometimes running the exact same prompt through to see what the difference is. But I had it create an API in Ruby in 3.5 and in four, and with 3.5 you could see how it got to where it got, you could fill in the missing gaps. But there were definitely gaps. if you ran this, it would not work by, just by the steps it gave me. but there were a few things missing. Four. it. Like first try, it had a bunch of gems that it didn't need and wasn't using. So like I would ask it like, oh, what is, why'd you include this? And they're like, oh good. my bad. Obviously that's not being used. And it was like, okay. So you're clearly looking at, most apps that include. this gem also included this gem, so that was the next probabilistic word, and it was fine. It was able to figure out, and that, I think that's the thing that still messes with my brain is that even though I know it's probabilistically detecting words, how does it unders it doesn't understand, is the point. But like, when you ask it about a thing, it knows what the thing is because I'll tell it to remove that gym and it will rewrite the code without that.

CJ: Yeah. I. can I take a swag at trying my best to explain stuff that I don't actually understand yet?

Colin: do it.

CJ: I th yeah, I th when we were in college, one of the projects that we did was building markoff chains, which was cool. You take all this Shakespearean text and you dump it in, and then you just break up the text into lots of, Element, tuples, basically thrus, I dunno if it's actually a throuple. and then you like, look at, you look at a throuple and you say okay, the last two words in this three word array. Now randomly grab one of the next arrays that also start with those two words. And that gives you your third word. So then you build this chain. So that's like whatever, a dumb markoff chain. But you could have it spit out like what looked like Shakespeare and this was like 2006 or something, 2008. If you imagine every single word or concept. Being mapped onto like a 2D grid, right? So we have the X coordinate and the Y coordinate, and we're just gonna throw every word and they're gonna splatter somewhere on that 2D grid. Now we can figure out exactly where each word lands on that 2D grid, and each word will have a vector that has like the, the length and the x y coordinates, or like the x and y that will get us to that point in 2d. each of these models has a certain size, so there's like a certain number of dimensions on the model, and I think GPT-3 was like 2048 or something like that. And if you look at these new, the newer models, they're much bigger. I think they go up to 30,000 or something. So if you take that concept of having every word mapped out in this space and you exploded out to. 2000 D basically, or a thousand D, then you have a lot more different dimensions. So it's not just X and Y, it's not X, y, Z. It's like X, Y, Z, a, B, abc, all the way up to 2000 of those. And in that space, words that are semantically similar. Will be close together. So like hotdog and burger are gonna be like, if you look at this like giant space, those are gonna be close together in terms of like where their vector lands. And so like you can use that piece of information to get a better idea of what like the next word in the sequence is. And so that's I don't know if that makes any sense. And explaining over audio is tough, but my understanding is that's like you can use this semantic similarity between these concepts and even between like sentences or paragraphs or like entire blobs of text to figure out what is this thing like? And then use that as part of your chain.

Colin: Yeah, and then we can ask it. The question of is a hot dog a sandwich with

CJ: Yeah, exactly. So one of the things that I think is in interesting is using these, part of these tools called embeddings, where you can say here is a concept or a sentence or a paragraph, and you can send that to open AI in their api and it will give you back a vector. That is just literally like a giant array of numbers that represents that concept for that given model. So say you're using Da Vinci, whatever the Da Vinci model for GPT-3 and you give it hotdog, it'll give you back something that's like an array that's 2048 elements long. That's a bunch of numbers. And those numbers, if you like, map them out through 2048 space. That's like where the hotdog lands. And there are, yeah. one of the things I was trying to do was take all my blog posts on my site and then run them through and build up these vectors and stick them into Pine Cone, which is like a vector database, and then use pine cones like querying language to have it give me back like search results for my website. I think this is like the bits and pieces that were, that I know, like I'm trying to like, learn about for sure. And build, I don't know. Th there's so much that I think will be built on top of this that, it definitely feels like, early, early days of the internet in terms of accessibility, tooling. like all these crazy new concepts they have to pick up.

Colin: Yeah. And you have to eventually learn how, like I, I've been learning how to give it prompts and that's like a skill in of itself. And some of them, especially these image ones, you can, the prompts become these paragraphs of describing the mood, the scene, the reflections. I'll put it in the show notes too, but a friend of mine is a photographer and he is using these. As well, but he generated something that I would not have guessed was not shot on his camera, up in Tahoe Fair. I'll have to find out what his prompt was to see what it was, but it was like he's a nature photographer and he's over here generating from scratch the same thing and isn't really threatened by it as much as trying to explore and push and see what he can do there. I think the challenge is going to be whether or not. for some of the applications that I've seen, they're not things that I was gonna do today anyway. it's cool, but like we went through a bot, a hype cycle a few years ago. then we went through the crypto hype cycle. And I have to wonder, I think that this feels like magic because it understands what, it remembers. if you're in a ChatGPT session, you can say oh, what does that mean? And it knows what the thing above was that you were talking about. So that shared memory is super interesting. But I think people have to remember that these are still. Models like you just gave us the whole pitch on how they work and not another person on the other end. Cuz there's starting to be some very scary stories about people falling for these, chat sessions as real people. you have some notes here about some of the doom and gloom, or Boone type of things where, there have been, scams where someone will call. Record your voice and then use that voice to do a voice clone, and use these models to be able to then call. A significant other or a child, and pose as you, and, they won't, especially on a phone where it doesn't have high fidelity, it can sound a little goofy or distressed or, say I've been, kidnapped as was the example that I hurt and please send money to this location type of thing. And it's whew. Like how do we get to a world where that exists and how do we protect our.

CJ: And the same tools that these scammers are using, we can, we're also already using that for this podcast, so one of the tools we use is called Descript, and like right after the podcast, we drop it in there. It transcribes everything and has this feature called Overdub where we can train. Descript and say this is, I think it's four, it needs 45 minutes or something. So I can say here's 45 minutes of Colin talking and here's 45 minutes of me talking and okay, so this was a word that we, one of us, stumbled over, or we said the incorrect word here. Then you can just edit the word as text as if you're just editing a Google doc and say, overdub, and it will replace that word with that person's actual voice. And that becomes really useful in scenarios where you're doing something like podcasting. You don't need to like, jump back on the mic and tr do a bunch of different takes to say the word again and then splice it in or whatever. but yeah, it's definitely a huge risk. And one of the recommendations I heard from, Rachel Woods on TikTok was that to avoid these voice scams is to create some sort of safe word with your family so that if anyone calls and they're like pretending to be someone, you, yeah. You have a way to, I don't know, crack their scam or whatever.

Colin: that's your, safety tip from Feld and learn this

CJ: Yeah,

Colin: go create a safe word for sure. and I would say definitely don't be feeding these things, any of your private or secured data. anything that you like. I think I've been hearing people putting in there like medical records and tests and stuff into this thing, and it is an impressively. Able to understand protein chains and things. And but that's probably because there's a bunch of research out there on protein chains and they show up in a certain way and they show up on the speced, map the same way that you were describing earlier. And so I am hopeful for the things that it's going to be able to teach us. Like it's better at spotting things on in radiology than a human. Awesome. That doesn't mean it replaces the radiologist, it just means that now they can detect and look at things faster. but I have to wonder, and this happens, happened ever since the industrial revolution, but as we get more productive with these tools, does that mean that now we're just trying to churn out more stuff and in 40 hours a week, or do we start to head towards that four day work week, three day work week?

CJ: Oh, I hope that three day work week comes, I think it's right around the corner . So going back to could it be a hype cycle? Is it a fad? I think when you, reflect about search and Google in general and the job to be done there is I want to find an answer to a question, . The way that I search is I'll like type in my search query and then I open like the first nine links , just like command, click all of them and then dig through the tabs as fast as I can. Scanning and my, my brain has been trained over the last 15 years, how to just try to extract maybe a right answer out of these things. And now you can just get the right answer, almost immediately. ha. Have you had a chance to try out the Bing, like the Bing client?

Colin: I haven't.

CJ: So what were we doing? We were looking for a type of Star Wars character. It was like a class of Star Wars character that's like this super esoteric character type that is only documented like on one page on Wikipedia, buried somewhere in the archives. And so we went on Bing and used like the Bing. GPT search or whatever, and it spit out the right answer. It was like, I can't even remember what it was, but it was like some type of, dark sister or something that, that had, yeah, it was like a force, I don't know, force sensitive thing. but yeah, anyways, like there's lots of different interfaces to this and I think the jobs to be done about just getting an answer quickly. If we can start to trust and rely on the models, which, I think one of the co-founders, Greg, one of the co-founders of OpenAI, said that coming soon, the answers that you get back from the, these models will be accurate and not be hallucinating, which, when that starts to happen, I think that's, that becomes a huge game changer. But yeah, until then, we all have to be skeptical and take every answer with a grain.

Colin: Yeah, there's actually a good book on this that's called Invisible Women, data Bias in a World Designed For Men, which I think is also something we have to keep in mind is there's been a bunch of conversation about ethics around all of this. And to be honest, most of the worlD&Data sets are very much skewed for as perspective. It tends to be, people who are building these things tend to be. white men and or, a few different, types of backgrounds. And so that we need to have representation in both the data sets and the people working on this. because otherwise, I even think about simple run of the mill things like recommendations. Sure. Like it's harmless to ask for a pizza recommendation in a certain town. Now, if it's generative AI though, is it based on Yelp reviews? Is it based on just like it knows what a pizza shop is and it gives you the first one? is it the closest one? there's a whole bunch of stuff there that we don't yet know. And so if you're asking for actual answers to things, it does seem to also have this sense of the way that it writes in chat, GPT specifically, it tries to agree with you. like the way that it presents things is very pleasing and agreeable. And what I'd be curious to see, and I've tried a little bit with poking some conspiracy theories and things at it, is wh how it replies because it can also. Reinforce things that you tell it. And so if you have a belief whether or not that belief is true or not, and you give it to it, and you have, you create this little, reality distortion field for yourself. And there's a whole bunch of books in sci-fi that would love to, to take us all for a ride on that. so we only have to look at our whole body of science fiction to, to see like where this could go. But, I'm excited to see where it goes. It is impressive and every time I use it, I'm like amazed and, not concerned for the future as much as, we gotta take a measured approach to safety and ethics here and figure out how we can use it for the better.

CJ: Anthropic is one of the main competitors to open AI and. this is again, gonna be stuff that I learned from Rachel Woods, but the, the team at Anthropic was a bunch of former open AI researchers who wanted to break away and make something less corporate and, they have this blog post that we can link to in the show notes about their views on AI safety, and it's my understanding that open AI is using humans to create the guardrails around the models. And Anthropic is taking a different approach called constitutional ai, where it's like giving it these principles. That are, these like things that basically say, behave or please behave nicely based on these set of values. And then it goes and like trains itself and then tries to make sure that its decisions are consistent with those values. And, I, yeah, it's it's a really interesting approach and it also made me think about, this is gonna take a little bit of a turn, but the way that our current constitution is, resulting in less than ideal. Country like formation, right? Like here we are realizing that the Second Amendment is actually not the best formation of, what the founding fathers expected. And we have all of this gun violence all over the United States. And so there is a chance that you write a Constitution with values that are consistent with what you believe. Correct and good today, but that doesn't stand up to the test of time. So figuring out how to evolve those values and principles in a way that's measured and safe and ah, it's gonna be really interesting. But,

Colin: it's that interpretation and context that's key. And even some of this reminds me of the three laws of. Robotics too, which are very dated and from sci-fi, but they're something we should look at.

CJ: Yeah. it is, sci-fi is like basically here today. It's wild.

Colin: It's very similar to some people have said, with all of this, the last 5%, the last 1% is so hard and hardware is gonna be a limiting factor. But also like when we think about full self-driving, if say we're at 90% of the way there today, like there's still a lot that's gotta happen before you'll let ChatGPT take the wheel literally. It's like obviously large language models are not what's powering full self-driving, but if we look at it similarly to the vectors of what's a person in a crosswalk and what's someone backing up when they're supposed to be going forward, right? Things like that, that a human can make those decisions, but if it doesn't have enough examples and it doesn't have. To reason around it's gonna be hard to get there. And then can you have enough of that on hardware that's in your pocket or in your car? we've got a long ways to go and it might be that hardware is a limiting factor for a little while. I think to take it back to the build and learn side of things, we've got two things that we can leave you with. I think one is a funny, thing I saw come across Ruby Weekly, which was that chat, GPT co-authored at PR to Rails. and this was a Link baity title, but one of the rails core team members, went ahead and gave it a prompt, created a poll, request, code, test, change, log everything, and then submitted it with, Disclaimer that it was created with guidance by ChatGPT. and so we'll leave a link to that in the show notes. You can read the pr, you can read the whole conversation. And, it was interesting to see like if you believe ChatGPT is infringing on your copyright, please let me know type of thing. and then you have a list here. What do we have down below?

CJ: Oh yeah, so this was, I used Notion ai. So we plan all of our shows inside of Notion and I just, you just do slash AI or something. And then I said, tell me five social media influencers that people should follow to stay up to date about ai. And it's spit out these five Rachel Woods. Logan, GPT, swyx, Lex Friedman, and Siraj Raval. So we'll have to do a little bit of, vetting and research about who we're gonna put in the show notes, but it's just I don't know, it's wild that, that it can do this.

Colin: Absolutely, and I think that is where we'll leave it. We've got a whole bunch of links that we'll put in the show notes to a lot of tools, a lot of different, models that we talked about today. And if you want to hear more about this, please tweet at us. @buildandlearn_ on Twitter. you can find all that in the show notes as well. Find us, and we'll see you next week.

CJ: All right. Bye friends.