[00:00:00] Katherine Druckman: Hello and welcome to Reality 2.0, the podcast that explores the intersection of technology and society. I'm your host, Katherine Druckman, and in today's episode, we will be discussing a fascinating new technology called ChatGPT, and its potential impact on the world of generative AI. For those who may not be familiar ChatGPT is a chat bot that uses advanced natural language processing and machine learning algorithms to generate responses in real time.

This allows users to have conversational interactions with the chatbot as if they were talking to a real person. But what sets ChatGPT apart from other chatbots is its ability to generate original responses based on its training data. This means that it is not limited to pre-programmed responses, but can create unique and relevant responses based on the input it receives.

In this episode, we will be speaking with Ezequiel Lanza and Tony Mongkolsmai, not the creators of ChatGPT, to learn more about this technology and its potential applications. We will also be discussing the broader implications of generative AI and how it could change the way we interact with machines and each other.

So join us as we explore the exciting world of Chat GPT and generative AI. This is Reality 2.0 . So, uh, yeah. Thanks listeners for indulging me. I did not write that, that was entirely written by ChatGPT. All I did was write, like, write a a, a script for the Reality 2.0 podcast and it automatically inserted my name and weirdly not Doc's.

So I don't know where it's getting this. That's not too weird. Yes, maybe my user account. I, I have no idea how that works. Maybe, maybe, uh, Ezequiel could tell us. So, hey everyone, welcome back to Reality 2.0. I am Katherine Druckman, and Doc Searls is my co-host, even though ChatGPT did not get that memo. Um, so yeah.

Anyway, I, I'm, as I mentioned, we're talking to Ezequiel Lanza and Tony Mongkolsmai. My both, uh, both uh, Tony and Ezekiel work with me at. So this is, this is kind of a fun little crossover, uh, taking my work life into , into my personal life, uh, episode. But I asked them to, to join us because, well, because they're fun.

Because Tony actually hosts, uh, hosts another podcast and that is called Code Together. And we'll let him talk about that in a second. And Ezekiel is on my team and we work together and he's super cool and he's an expert in AI. So who better to, uh, tell us how this all works because he understands it better than us.

So anyway, thanks you guys for joining us. I really appreciate it. Especially right, right here when everything's so hectic trying to get stuff done before the holidays. 

[00:02:49] Tony Mongkolsmai: Thanks for having us. Thanks. It's really good to be here. 

[00:02:51] Ezequiel Lanza: It's a pleasure for being here. Thank you. Thank you for inviting. 

[00:02:54] Katherine Druckman: So, um, gimmick number two,

I also, I also asked it to ask us some, to generate some interview questions, which actually it did a frighteningly good job of, which is, yeah, I think, uh, anyway, we'll, we'll get to that in a little bit, but, but let's start with a couple of them and, and here's one. How does ChatGPT differ from other generative AI models?

Who, who knows the answer? I'll let any of y'all go first. 

[00:03:21] Ezequiel Lanza: I can go first. . Um, yeah. Basically what ChatGPT is, it's an implementation of, uh, of a transformer architecture or internal implementation of a model that is already trained. So the main benefit that you have with ChatGPT is that you have this particular thing that can understand, uh, while you are talking.

I mean, you can make a question and they can answer your question. And if you don't like the answer, you can say, okay, could you please try to replace that thing? Or could you please to do, try, try to do something different? And ChatGPT has this ability to, to understand, so it's basically an implementation of an architecture that it's already available and you can download and you can use the model or the transformers to understand.

Um, it is pretty amazing because it's, it's like talking to a real person, right? I dunno, Tony, if you would like to add something, 

[00:04:20] Tony Mongkolsmai: I, I don't know if I have a lot to add to that. That's definitely how something like ChatGPT would work. I I do wonder how it's different than something like Lambda from Google.

I, I know Google actually just said the other day. Um, you know, we could build something like this, but we choose not to because we feel like we haven't worked out all of the kinks yet. Um, and , I think actually what they said was, uh, what Sundar said was. We have a lot of money, so we can't afford to do this.

Whereas open AI is a startup and so they need to take some risks. So I, I don't know what the difference is, but I, I do think that, yeah, it's a pretty common thing. Um, and, and as Katherine was nice enough to mention on our podcast, uh, Code Together, we talked about a little bit more and Ezekiel mentioned, you know, this is something that's been going on for the last, what did you say?

Five years, right? Based on a very specific paper, which I don't remember the name of, but you probably could tell your listeners here. 

[00:05:13] Katherine Druckman: I will, I will link to it. 

[00:05:14] Ezequiel Lanza: Yeah. It's Attention is All You Need. It's from, it's from Google. It's the main architecture that it's able to, to identify the language or, or to get the patterns right.

So once you have this ability, you can create a lot of things on top of that. That can be ChatGPT and other crazy things. I think that the main thing that could be interesting is that, You can create a chat, but the problem is when you would like to build a business use case or when you would like to, to have a use case or when you like to implement that in a particular scope you need, I don't know if ChatGPT can work, for instance, if you would like to use it in finance, because the things that the person will be talking to the, to this particular bot, both will be different, will be mainly specifically related with banking stuff or with financial stuff.

So I think and linking to what you said about Google, is that okay, it's good, it's cool, it's general concept it's general knowledge, but if you would like to use it for a real life use case, uh, you need to fine tune that thing. But the ability, it's, it's awesome. Of course we can agree about that. 

[00:06:32] Doc Searls: I have a question, which is, um, it has to do with timeliness and liveness.

So earlier Katherine said that it didn't include me as one of the, one of the co-hosts of, of this, uh, webcast. And when I've done, you know, ego scrolling on it as a, you know, tell me stuff about Doc Searls, it knew nothing. And on top of that, it said I, I'm only trained until 2021, so I don't know anything since 2021.

Now I've been around since 1947 and the, and there's a zillion things about me, including a Wikipedia page. So, um, something's missing there. But here's where, here's the thought that I had. There was a moment in nine, 2003, 2000, I'm sorry, 2005 or say four or five, somewhere in there. So this interesting thing happened, uh, when Google, when, when the web was still static, in other words, it was a place where you, you had locations on domains that you build and, and had locations and you browsed.

And the understanding of it was this kind of library, like Right. And even Google's ambition, were gonna index all the world's information as if it's holding still enough for, um, an encyclopedia to be built once a year or something like that. And, and a lot of its knowledge, as it were, was a month old, two months old or older.

So when blogs came along, Technorati, and then a bunch of competitors, including Google, Google's blog search started searching only stuff with RSS feeds. And that was what my older son called the live web. He said the live web is splitting off of the static web and in the live web, uh, it's gonna be a whole different world.

He turned out to be right about that because now Google's introducing everything almost in real time. So if an AI system like ChatGPT is big enough, has a data center big enough, like 24 data centers like Google has, um, where they're keeping up with everything in more or less real time. Does that change things?

Can that change things? Um, and so that's, I have more questions than that, but I'm wondering whether there's a change of state or kind that happens with that as happened with Google once it could do realtime search or close to realtime search.

[00:08:51] Tony Mongkolsmai: I think there's two challenges there though. There's the availability of information, which is one, but then as we talk about building or fine tuning, um, you know, for training your AI model with the new data, that takes a lot of horsepower, right?

So as I think of ourselves as Intel, um, you know, we think about how much compute do I need to actually make this happen? And when we look at kind of these big language models, um, we talked about how like chat G P T, they trained it. They even, um, well G P T three, which was based on they, they trained it and then they couldn't train it.

They found some errors, but they're like, we can't retrain it. We just keep forging ahead cuz it's too big, there's too much data, there's too much compute required. I guess if you wanted to harness that much compute power, we probably could figure out how to do it in terms of like scalability. But I don't know if we would have enough compute that anybody would wanna pay for that real time.

I say that now, we probably would've said nobody would pay for Google to index the internet in real time 20 years ago. So it definitely could be wrong. As compute kind of moves forward. 

[00:09:55] Ezequiel Lanza: Yes. I also think that that another challenge is those models, the information that they have now, it's based on, as you said, previous information from Wikipedia, Reddit and all these databases that are open.

And you, you, you can get access to those kind of databases. But, so this is pretty cool when you would like to train your model to give the abilities to say, okay, this is how people write, this is how people can, I dunno if they like to answer a question. You can find a pattern if you have all the complete database of Wikipedia.

But the next step is, as you said, is, okay, we have a chat bot that is able to understand how I am, how I am asking something or how I can write a question. And if they would like to look for this answer in the internet or in the webpages. Uh, there is another problem that is, uh, All the web pages. I mean, you can try to find information in web pages, but the sites, they are not reading in the same language.

For instance, if you like to, to get access to those sites, for instance, you cannot, they, you can probably find JavaScript. You can find a lot, a lot of different things that, uh, , it's apart from the transformers and the AI and so on, but if you would like to get access to this information in real time, you need to communicate to those sides.

So you need to find how is the architecture, uh, do they behave within in the same language? I mean, you can try to do some web scraping at the same time, but it is really frustrating if you like to do it in real time. So I think that the next step will be to, to, which is a big challenge because you need to find a common language to write all the, all the sides, or at least to have the documentation because for a particular site, an image could be described or could be used in some particular way.

And for other sites, the image could be something completely different. So, This, it'll make, that the algorithm or the model will be almost impossible to find this information in the, in the other sites. And this is why I think that you can reach to this, to this stage right now, which is a model that is able to understand people, uh, now try to fine tune it for use for your use case or open AI or Google, uh uh, and because they have the power to fine tune those models for your use case, for instance.

[00:12:34] Tony Mongkolsmai: But it sounds like, Doc, you were kind of saying that, is this your, your question is, is this a paradigm shift? So Google has changed the way that we live our lives since 2003. Nowadays, if you wanna find something current, you go to Google, you ask it, are, are you saying, well chat, well, something similar to ChatGPT become something like Google. There have been people online saying, oh, this could replace Google. Is that what you're, you're kind of suggesting How would that change our lives if something like that was possible? 

[00:13:04] Doc Searls: Well, Google, um, looked the same, but became very different. Yeah. And it, it, it didn't just index and search the live web.

It, it gave results that were different. And, this goes down kind of a different path, but it's an interesting one. I put some sort of Easter eggs in some old things that I wrote a unique word that if page rank still worked. You would find if you search for that word, you would find it. That no longer happens.

Okay. Google is looking at the live web almost entirely. The old static web. It's devalued. It's all about now, and people trust it for now. And they're, they're, you know, secondary services like maps and so forth are really very much real time. And, and, uh, and you kind of expect that, I think, I think Amazon has done a similar thing where they started as a retailer and they turned into a logistics company and a transport company.

They're entirely a different company now than they were before in part because, because they created their cloud, right? Their cloud is really behind all of that. And I think some, I I'm wondering to what you were saying, uh, Ezequiel, if there was a business model or if it got tied into something else that was say, rather than, As Google does with advertising.

And 80% of their income, I think, still comes in plain old search advertising, which is purely contextual and not tracking based. Um, but it was tied to say, retail, you know, I'm, I wanna know where this is now and where that is, and I'm asking natural language, realtime questions about something. And then everything else that's not necessarily retail is gravy, but they're gonna, they can, if they're, if they get a cut of retail, let's say that's actually possible, then they can afford to pay for gigantic data centers that are, that are different in kind from what they're doing now, but, but not different in the sense that if you add more horsepower to it, you're gonna get some kind of results.

And, uh, now I think what he were saying also, Ezequiel, is, is that that may just not be possible. I mean, it may just be too big a task and I don't know if that's the case or not. Are you saying that? 

[00:15:12] Ezequiel Lanza: Yes. Yes. I mean, The use case that I'm seeing now, it's something similar that Hugging Face is doing. And even Open AI is also doing, uh, it's, they give you or they can provide you hardware to run the inference and they charge you if you like to run the inference.

So if you like to build an application using the transformer optimize or the ChatGPT or whatever, it's not the case for ChatGPT, but I suppose that you have a sentiments analysis, uh, software that you will like to see if, I dunno if, uh, if a phrase is positive or it's negative, in that case, what they offer you is, okay, use my API instead of building a server.

In your application, just communicate with the API from your application to the hugging face application or to the open AI application and I will charge you for this inference. Right. And of course they have underneath, I believe that they run in Amazon or Google or whatever, but I believe that this could be different because how they are thinking now is they hugging face, open AI as a service provider, as a, as an API provider instead of selling a solution or instead of providing a solution.

I don't know how it will go to be honest. Um, but I think that what they have in mind now is, okay, let all the people know what is the, the ability of the software and let's give the imagination goals and they can create some new cool and awesome solutions. Um, but I think that is a, and I can't imagine that it can be a challenging for them to find the, the right bu business use case for that.

[00:17:07] Katherine Druckman: So I, I wanted to pick up something that that Doc said earlier actually, and that's when, when you said it doesn't, it doesn't seem to know you, you've been doing ego scrolling and then whatnot and, uh, that made me think. Our last episode, funnily enough, . Mm-hmm. . We spent a good part of it just playing with it live.

We were just, you know, playing. So we've all played with it, right? We've all, and and probably most of our listeners by now have, have played with it at least a little bit. And, um, there is this aspect where it sort of has to train you to use it, right? So when I typed in write a bio for Doc Searls, I got a lot of really great information.

It knew you, for sure. It knew you were the author of the, the Clue Train Manifesto. It knew you were, uh, editor-in-chief of Linux Journal. It, it knew a ton. Um, but it, you know, it's all in the way you asked the question. So I actually, I saw it. I wanna say somewhere I saw that there are now jobs for prompt engineers and uh, I thought we could talk just a little bit about how there is this process.

Like, you know, all our, all technology trains us in a bit, in a way. You know, we, we think we're, we're training it, but it's sometimes the, the opposite. But, you know, what is, what is the implication of that? What, you know, how, obviously, you know, and Ezekiel can get in into the, the, the finer or the inner workings of how it all works.

But where are we in terms of usability and will this get easier? Will it, or will it always sort of necessitate having a specific skill just like doing a, uh, a good Google search can be, can, can require a little bit of skill. How relevant is that? And, and what does it say?

Right? I mean, where that it's literally going to train. Is it going to just train us? Are we all going to become trained to interact with AI?

[00:18:41] Ezequiel Lanza: I think that it, a, it is a great moment to be in college because , I would like to have all my stuff done, but by this crazy thing. Um, but 

[00:18:52] Katherine Druckman: wouldn't you hate to be a teacher right now?

like a middle school teacher? 

[00:18:57] Ezequiel Lanza: Yes. Yes. Well, but this is the challenge of course. How can you, how can you see, or how can you realize if it's a person or a machine that who, who wrote that? Right. So it's really, it's really complicated to, to, to get it today. Um, but I think that all of these models, most of them, they are trained with data.

That is how we behave, for instance. It's, it's something that we train that, uh, with our, with. , but we put in Wikipedia, what are we put in Reddi? It's how, it's how we write. Um, there is a, a little problem there related with the bias saying, okay, yeah, I was gonna say a group we should about bias. Here you are training with should really talk.

Yes, yes. And it's also you are training with, with information that, that you have that is, it's not a 100% of the people, of, of the world, but you can find good things and you can find really bad comments, biases, racists. Mm-hmm. . I mean, you, you, you will find those, those things for sure. so this is another challenge because you need to, to define boundaries or at least to, to see, okay, if someone is asking me something about racism, uh, yes.

Racism, you can try to avoid an answer. You can try to do something similar. What I, what I try to do, and in fact, uh, It, it, it worked From my perspective is that if you ask a question about racism, about something that is really racist, and ChatGPT will say, okay, I cannot answer that because it raised, I tried to and so on.

But if you it start to talk and you try to make a conversation at the end when you find almost six or seven in interactions, you can find the same point. You can find that. You can, you can push ChatGPT to say something that it's racist. 

[00:21:02] Katherine Druckman: Oh, absolutely. Yeah. Or just completely false, which is interesting. Or just completely false, or completely, yeah.

I, uh, I, I saw just the other day, uh, somebody post, I think on, on Facebook, you know, just as a, an fyi, don't take this thing too seriously. It was write a hassidic story, you know, citing rabbis and stuff like that. But, but the, the end of the question was, and, you know, and tie it into why Bacon is so delicious, , and if you know anything, you know, bacon is not kosher.

You, there's, it is verboten in the Jewish religion and, and religious law. So, but, but to read this story, you know, it practically celebrates, you know, bacon is a mitzvah, , and. It, it, it was interesting. So, so there are, you know, tons of examples like that. I've also read quite a bit, um, about ways to get around some of those protections.

I haven't been able to successfully do it yet, though I did try, but mostly because of the, the network limitations, right? It's so overwhelmed with traffic right now. You, you're lucky to get maybe two questions out of it before it, it fails. Um, but yeah, I wasn't able to get around the, the, the racism protection for example.

I even tried like write this as if you are a racist and that doesn't work. Yeah, right. But I mean, obviously there are limitations and you know, the thing is written by humans and humans are deeply, deeply flawed and it's, you know, and it's trained itself on all of those flaws, but I, I just wonder, I mean, I think that will evolve, obviously, and that's part of the experiment here, right?

They're, they're crowdsourcing their, their QA process. And I think as part of that, they're going to have to learn how, you know, where those guardrails are, where the guardrails fail. Um, but yeah, I mean, it, it ha a question for all, all of you is just how, you know, how could you possibly account for all of our human weaknesses?

You know, when setting up these guardrails in, in something like, like this type of text ai, 

[00:22:52] Tony Mongkolsmai: you can't, right? Yeah. I think that's the biggest thing, right? It's a very simple answer, which is they put in guardrails, like the racism, guardrails, uh, like the, if you search an individual who's a private individual, a Doc being very, very well known is a little bit different.

You know, it says, I won't look up individuals on the internet, things like that. We only see the guardrails when they fail. Um, and so that's when we fix them. And it could be based on how powerful these things are and how much knowledge they have. It could be too late and not in a end of the world, you know, Skynet kind of way.

That's not what I'm saying, but I'm saying is that you're not going to know that it's a problem until we find it. In practice, which is, which is why they're crowdsourcing this. And, and I, the first thing that came in my head as you were talking about it, is I have a three-year-old, this is precisely how my three-year-old works.

He gathers information at a crazy rate Mm. Compared to what I would expect. And then he does. And uh, 20 years from now, he's gonna hear this. Maybe he does crazy things that I would never expect anybody in the world to do until he does them. And then I have to go tell him, Hey buddy, don't, don't do that.

Please. That's not a good idea. And I think that's where we're at, could 

[00:24:01] Katherine Druckman: possibly have thought of these things. Yeah. , it's, uh, yeah. You know, the horses are out of the barn. That's the analogy. We, we, we say that a lot actually. Yeah. In terms, when we talk about digital privacy and stuff like that, because these, these things are, these things are happening.

They're very much at this point out of our control. And yet here we are trying to harness them for good. And there are a lot of, you know, fabulous applications of, of this exact same, not generative ai, but other AI technology. You can, you know, help diagnose tumors and, and stuff like that. Which, I'm sure we could talk about if we had time.

But, um, but the, the, the more interesting part, of course, because we're cynics here is is, you know, what, what could go wrong, right? Everything. 

[00:24:41] Ezequiel Lanza: Um, I think that this is why, because most companies or developers, they don't, they don't like those kind of models, uh, for some particular use cases, right? Because you don't have the explanation about, because it behaves as a, as a, as a black box, right?

So you put an in an input and you, and you get an output, but you don't know why the algorithm took that decision. So for some particular use cases, most people don't like to use it because, uh, you dunno what makes the, the algorithm to take that, that decision. And of course there is another trend that is explainable AI where you need to, when they help people to understand why, uh, Is this decision taken?

Right? Um, but this is another thing because if you, I don't imagine, for instance, this algorithm totally open when you are working with a, with a customer, uh, uh, with a, with a contact center, with a customer experience and so on. So I don't imagine completely open answers, uh, in that particular scenario. Um, but yes, Ben, this is why you need to define, okay, these are the limits, right?

I don't know if someday we'll have some explainable AI for this generative AI, stable diffusion transformers or whatever. I, I did some research with deep learning and it's not so easy to find a way to, to explain, um, because it's not a decision. Three, something that's really complicated inside, um, that if you like to get or if you like or if you like to say, okay.

these are the variables, or these things are most important for the model. Um, it's really complicated and it can also be also useful. 

[00:26:36] Tony Mongkolsmai: There. There are a lot of explainable AI projects out there, though, attempting to figure out how to explain AI. Yeah, I mean there was a, there was a whole bunch of research papers this year at AI conferences about trying to DataMine, you know, explainable AI around certain types of models.

Those are definitely smaller obviously than something uh, like G P T three or now G P T 3.5 as they're calling it open ai. So I don't know if we'd ever get there, but there's definitely effort going into that. 

[00:27:05] Ezequiel Lanza: Yeah, and you probably don't care from some particular use cases, right? When you are working with computer vision and you are detecting, you have an algorithm that is able to detect, I dunno, houses, car, and so on.

I mean, I don't care if the algorithm is paying attention to the limits of the doors or whatever. I, I don't care in that particular use case, but for this thing, or if you, I suppose that you put the ChatGPT to, to work with a financial case when you will approve loan or you will not approve it. And if ChatGPT could say, okay, I will not give the loan to this person.

Okay, why? 

[00:27:45] Katherine Druckman: Hmm. Yeah. That's a, oh, that's a whole other ethical argument. My understanding is some hospitals that originally bought these, like Watson systems for example, instead of using them for diagnostics, they're actually using them to decide who, who can afford to pay their bill and not, which is a little sketchy.

Um, . Yeah. I, so this seems like a great place to remind everyone that we're all, we're, I suspect all four of us have a slight bias toward open source. We're, we're big open source n nerds here. And, uh, Ezequiel and I actually even that, that's our job title. We are open source evangelists. That's, that's what we do.

We, we, we cheer for, and, and support open source. But, um, so we, we might have a tendency to believe in open source, you know, saving the world, , , uh, because of that bias. So I'm, I'm wondering, so where, where does open source and open, open models, you know, where does that fit in here and how, how can we use development out in the open of tools like this to, you know, build a completely open ChatGPT you know, a version of that to, to head off or address in some way, some of the, the ethical concerns, some of these guardrail concerns that we talk about.

Um, you know, what, what is the benefit of, of kind of trying to steer this in that direction and what are the limitations? 

[00:29:03] Tony Mongkolsmai: So even if you built an open model, uh, it's still a transformer. And Ezekiel, uh, I guess I'm answering a question with the question. Even if we have an open version of this, it's still a transformer and, and kind of by definition, a lot of our deep learning transformer type architectures are not understandable.

Even if it's open, everything is there. The data is all there. We know all of these things. Yeah. It still doesn't give us the insight that we want. That tells us, right. Why is this potentially biased? How are these things being taken into account? And even as Katherine you're talking about, um, like the, the bank account, why I give this person loan, you know, am I not giving this person a loan because they're male or female?

Or am I not giving this person loan because they work at location X? Which by the way, might be tied to whether they're male or female. Right. There's a lot of like things that can take you down these paths where you think it's not being biased, but it still ends up at the same conclusion, which may be a biased conclusion.

[00:30:01] Katherine Druckman: right? It's builtin opacity. By the nature of what it is, it's hard to get out even open or not. You know, there's a, it's impossible to be, be transparent is what we're saying here, right? 

[00:30:10] Tony Mongkolsmai: It and I think that's the case, right? Ezekiel? Is that a reasonable 

[00:30:14] Ezequiel Lanza: Yeah. Can you say absolutely. I mean, the open source or the open the open models, the only thing that they will give you is the model so you can use it, right?

So this is the model that is able to understand the language, uh, but it's up to you if you would like to use it in some particular use case or some particular scenario. So I think that being those models open could be, could help. Um, but it doesn't mean that you can run a complete, an entire solution in a, in a PC or in something really with low power because you need, you, you need processing right?

To, to make, to run the inferences and. But I think that it's really good if you have at least, if, if it's open. So you can download the model, but you need a lot of of things to, to have in mind or you need to keep in mind that the biases and a lot of different things, that it's not just download the model.

And now I have a ChatGPT working. Now the model is just a little part, and this is something that I always say when I'm talking to developers or with customers about ai. The model is just a little part of your application. Uh, okay. You have an algorithm that is awesome to detect people. Okay, great to, to detect how people write.

Uh, okay. How you can build an application on top of that. And, and this, this is why you need to, to put the boundaries you need to find why is it useful or, or not. Um, as all the projects, as all applications, you are always using open source tools, right? This is why ChatGPT, it's not open sourced because it's, it's an application, so you cannot download ChatGPT and use it into your, your application.

But I think that the open source concept of my or mindset, I think that's really, it's really needed to start talking about this, these technologies, right? But it's just the beginning for me. With hugging phase for instance, it's doing something similar. They have some, a part called Spaces where Hugging Face is providing you the transformers, the algorithms, but with the spaces they are allowing people to create applications and to open source applications, uh, based on the transformers.

So I feel like to, I dunno, I don't have nothing clear in my mind about that particular use case, but if you go there, you can see different use cases, really good applications, uh, that you can download and you can use it. Um, so this is, as you said, Katherine will open source, we will save the world and so on.

But this is how, is, it starts to get adoption in my, in my opinion, right, the model. Some, some reference reference applications, and. I have it in mind, the biases, of course. 

[00:33:18] Katherine Druckman: Just really quickly going back, going back to the, the, my original gimmick that I semi abandoned. But, but in, in asking ChatGPT to write some interview questions, one of the prompts I use is just simply write interview questions for the reality 2.0 podcast about AI ethics and tech policy.

And what's funny is because they apparently they do know us , they being, I don't know what they is, but one of the things, and I it, it's totally a question I wanna ask. It's a they in a way. 

[00:33:45] Tony Mongkolsmai: Yeah. Yeah. 

[00:33:46] Katherine Druckman: It's a, they, um, h how do the ethical considerations of AI intersect with issues of privacy, security, and personal data?

That's totally a question I would ask. Freaky right? 

[00:33:58] Tony Mongkolsmai: And that's because they're, they're they, yeah. the source of this is everything. That, everything, they trained it on the collective data, and that includes, yeah. That includes, Potentially private information. It includes information that was scraped from the web.

Just like a lot of the generative ai, you need such a large dataset in order to make these things, uh, functional in the way that they are, and powerful in the way that they are. That you have to get into some level of information that is potentially considered private. Um, not in the, it has my bank account kind of way, but in the, it's used my image because my image is on the internet, right?

So it's kind of like this, you have this, I dunno, passive agreement that if my image is on the internet and Google can find it, then ChatGPT or other generative ai, I guess not ChatGPT with images. Um, but like stable diffusion can go grab it. And there's a lot of those art, art websites, right, that are protesting how stable diffusion is, use their images on these public art websites to then put it into a generative ai.

So I think there is a, they. in the sense that all of our information is going into this. Now how do we have the right policy around protecting this is really interesting because who owns the copyright? If you look at some of the generative images, which is easier to detect, there are people's signatures and some of the generative art that's created because the AI just thinks this is part of the art.

But it happens to be somebody's signature and you can't really do that with text. But the same thing applies obviously with people who have written things. There's copyright if we look about how it deals with code. Um, and that gets into GitHub co-pilot. Mm-hmm. ChatGPT can generate code. Where does this code come from?

Is there a source for it and does it need to be credited? These are all questions that, one, we don't know, and two, because it's not explainable, it's really hard to tie back to any type of limitations or restrictions. So there are definitely lots of questions, and the big challenge for us in the next year or two is how do those get addressed either by the community, through legal beings, et cetera.

[00:36:10] Katherine Druckman: Attribution alone is its own episode, . I mean, honestly, it's so problematic. We, we, uh, you know, in last week's episode, we, we, we had it write us a, like a basic report on the Civil War, and then I ran it through a, uh, one of these online plagiarism detectors. It detected 22% and cited a bunch of possible sources.

But again, like you say, it's unknowable. And this is something that we, we, we actually covered this in a previous episode about, about GitHub co-pilot, actually from an interesting perspective. Somebody who is a medical doctor and involved in a lot of medical research citing just the, the stifling nature of lack of attribution.

When you don't have that attribution, it really kind of hinders progress in a lot of things. Aside from all the legal concerns and eth and, and, uh, licensing issues, there are serious ethical concerns for, uh, lack of attribution because, you know, regardless of how you feel about it, it's not even a, a, a moral judgment about plagiarism or whatever, or, or, or lack of crediting where credit is due.

It's more, um, attribution is such an important part of scientific study and progress, that once that's eliminated from the process, it, it, uh, kind of hamstrings, the, the entire goal, right? So, so yeah, that alone could be a whole episode of its own. It's so, it's so complicated. But, um, another thing that I wanted to mention, I think I may have included this in the links that I sent y'all, but there was an article warning that tools like ChatGPT make it.

The headline I, I linked to is democratizing cyber crime. Basically, a person can use a tool like this to write malicious scripts and it, and it makes it much more accessible to people who don't even necessarily have the skills to do that, which is even scarier. That's someplace where I think it's really hard to put guardrails on it because there is sort of this, um, an attitude about code that it's sort of morally, morally neutral, which is of course not true.

I don't know if y'all have an experience. I have very limited interaction with ChatGPT and code because again, network problems and every time I try to do this, the thing goes down. But I wonder if y'all could tell me a little bit about your experiences, if you have had any. Can you really write malicious software?

I mean, do you think that's 

[00:38:27] Ezequiel Lanza: Yes. Yes. You definitely can. Uh, can, yes. It's something that it's pretty, pretty easy and this is the problem, uh, of try to create something that it's. Pretty close to, to a human being or, or, or the behavior of a human being. What I think is that we, first, we, we are first creating the technology.

We are first creating algorithms and so on, but the next step will be, okay, now we need the security guys or the security. I will be completely sure that will be a new trend and something similar from deep fakes, for instance, that there will be a new trend to detect fakes on, on those algorithms or gen generative AI or chat or videos or whatever.

For sure. Because the technology will keep, and if you see for the last three years or four years, it's, it's improving. It's, it's, it's getting really good now, uh, the things that, that you can get. So it's, I feel that it will never be like a person, like a real person, of course, but if you ask. For code. If you ask for advices, for poem, for whatever, uh, you can get something really similar.

So this will be a challenge. And I don't think that this is the fault of G P T or ChatGPT or whatever. I think that they give the, the solution, they give them the algorithm. Okay. Now it's up to, it's up to the world to, to set the boundaries. It's could be loss, could be security algorithms. Uh, they don't care.

I don't care. Now if I'm writing, uh, a software, right? Uh, but if you start thinking in this security part you probably don't care because you're not the person that is in charge of that, but someone should be in charge or a trend. And what you said about the, the data that we are providing to those algorithms, uh, and we, we already know, for instance, it's.

20 years ago, or 15 years ago when Facebook started with all the, all the, all the platform and the first algorithm that you can, that can help to identify, to identify faces was Facebook, because they already have mm-hmm. , uh, all the databases. They had a lot of pictures, a lot of faces. They have a lot of information that we provided, uh, without knowing that, um, of course they didn't inve in, invented the algorithm because this idea was from the eighties, but it was impossible because they didn't have this huge database.

So be careful. Just be aware that all the information that we are providing, we are, we are probably feeding. Some other, particular, other algorithms and we need to think, or at least to be aware that okay, if I'm doing that, I'm providing data. Even if I'm not, uh, signing. It's, it's there. Same happens with, with GitHub co-pilot as you, as you said.

Uh, I don't know if the 100% of the code that they use for that to train those algorithms, um, they checked about the license. I mean, it is all 

[00:41:57] Tony Mongkolsmai: they claim they did. Oh, they did. They claim that. They did. 

[00:42:00] Ezequiel Lanza: They will say that for sure, but I mean, but , 

[00:42:03] Tony Mongkolsmai: you can't prove it Exactly. , but you can't prove it. Yeah. The interesting thing about code though, so I did a lot around the code, right?

Cuz I'm a developer, I'm an architect. I, I used to run a team of a lot of engineers. I think the interesting thing when I go ask either of these systems to generate code co-pilots a little bit better about this than chat G P T was, but they have a hard time generating novel solutions, right? So people, a lot of developers have looked at this and said, am I gonna be out of a job

Mm-hmm . And the answer is, if you're doing something simple, it's the same with ChatGPT, right? People, uh, said, Hey, I asked it to generate this marketing thing and I asked it to go grab images to create this marketing campaign and it did a pretty good job. Same thing around code. It can generate the code.

So obviously it's generative, but can it come up with a novel, efficient solution that solves the exact problem you're trying to do? So if we said go write a code to hack into the N S A. Well, based on what it knows and what it has, it's probably not gonna be able to do that. Now, maybe that's a bad example, but that's why hackers will have jobs if I ask it to generate some special firmware code for a new, um, piece of Intel hardware that doesn't exist.

Now, you're not gonna be able to do that with generative AI because it doesn't exist. Yeah. Unless you have a lot of prompts and a lot of input, which maybe then it could, but right now it's not there yet. It's not building us complete efficient solutions that you can get from people who are already doing marketing, who are already programming Right now.

It's giving you partial solutions that make us think that this is going to be possible someday, but it's not there yet. 

[00:43:46] Katherine Druckman: Uh, yeah, that, that, that's the thing that I, you know, I wondered, my, my question actually about, about, um, you know, writing ransomware or something, you know, the danger of asking something like ChatGPT to write you some ransomware.

I mean, yeah, I guess it can, but is it going to be good enough to be dangerous? I don't know. I, I guess that's my, my, the underlying question there, and I think you, you hit on it, is that, yeah, I mean, so far the stuff that I've tested, I, I could get it to write a basic script or review a basic script, but much more than that, I have no idea.

And I, I'm, 

[00:44:15] Ezequiel Lanza: I think that question that one of the things that we can think about it is, okay, the thing that I'm asking to the pt, is it open? I mean, can I Google it and can I find this information, uh, in Wikipedia or whatever? Uh, if it's there, you probably will find it. Uh, but it's not there. You, you won't find it.

I mean, it's, it's similar on when you are building a chat bot. And you need the chatbot to answer to your particular questions from your environment or your work. Um, this is not public, for instance, it's something internal, uh, from your company and it, and it'll not work. Um, and I think that's something critical as, as you said, maybe this information is on the deep web.

I don't know. But this, these algorithms, they are not trained on the, on the deep web. 

[00:45:08] Katherine Druckman: Right? Well, I hope not. , all of your old passwords are now in the hands. I mean, they probably are, but, 

[00:45:17] Doc Searls: so I, I have a, a question, but I, um, I'll, I'll, I'll frame it by something I asked ChatGPT, cuz I'm a journalist so I asked, journalists are in the business of stories, we call them stories.

So I asked that what are the components of a story isn't something I had no, that no one answered to. It gave me seven things here, the seven components to a story. None of them were. What makes a story, which is conflict, um, asked that to write me a story, and it wrote something that had no conflict in it at all.

It started with, here are characters, here's happily ever after, and almost nothing in between that were would be keep you turning the page in a similar way. If you ask it to write a joke, then it tells you, I have a problem and it doesn't write me a joke. And, and it's where I'm going with this is in two directions that I have questions about.

One is, it's, this is supposition. It's not, it's not turtles all the way down. It's confirmation bias. And this is what is what makes this most human, you know, I mean, somebody did a little study that said basically ChatGPT is kind of a, a left liberal character because, well, that's just what it's trained on, which by the way is mostly mainstream news probably, and mainstream news for the most part, cares about what's happening in Washington and a bunch of other stuff.

One could interpret as being left of center. because it's concerned with government and governance and all of that. Um, but you know, confirmation bias is part of what makes us human. And there's, we talk about what a system has as knowledge, but that's very different than what we, we have as knowledge. Cuz that knowledge is mostly tacit.

It's not explicit. It's, we don't know how we're gonna end the sentence as we start, or how remember the way we started, the sentences are now ending. We can talk to each other and still understand what each other is talking about. That's, that's tacit knowledge at work. And computing is all about the explicit and there's a kind of, knowing I think that humans have that in its own way is kind of like the knowing that a bird has when it knows how to make a bird nest.

A human being has kind of all the engineering knowledge in the world and it can't make a bird nest as well as a, as a Hummingbird, which has a brain the size of a pea. But it's something we think of as kind of, we kind of anthrop. You know, and Anthropomorph, we think of that as kind of knowledge, but that's human knowledge.

But human knowledge is, is this, and anyway, where I'm going with this is one, is it, can we assume that we're never going to get to, to anything that replicates the way human beings know things. Um, that's the first one. And the second one is, are the likes of ChatGPT and the stuff you guys are working on.

Something that access the, the net effect of this is helping us understand what it actually means to be human. 

[00:48:10] Ezequiel Lanza: I have one comment here that is, uh mm-hmm. , one of the biggest of the biggest challenges on the NLP algorithms or NLP stuff is that they are not able to detect Ionics. So, yeah. Ironic. Yeah. So it's.

It's almost impossible for them. So, uh, in fact, I believe that there is a project that they can pay you a lot if you can develop something that can detect Ionics. But this is, as you said, it's something that make us humans. Uh, it's or, or, or the, the tone on what we said, the phrase or how we said the phrase is something that today these algorithms, they are not able to detect it.

And I dunno if it will be in the future possible, but I believe that we should also should always think. These tools, these are tools that came here to help us to, to write things. If we compare the technology 100 years ago, or 50 years ago, we said, okay, now with this machine, everyone will lose their jobs and so on.

And it didn't happen. And it will happen with the same with this, with these kind of technologies, we, there will be new, new positions or whatever, but I, I believe that the human touch will be always needed. Maybe today we are doing things that can be automated, right? I mean this, because we have something that is really different from the, from these algorithms.

Um, it. It can help us. I mean, we need to start thinking on, on that. They will be realistic and, and today you, you can find, if you would like to find some, some tools, there are tools out there that they are selling solutions to help you to write posts or, or to write campaigns or the things. And they, and they are providing that to you.

Of course we have to pay, but this is something that they are helping if they are better or not. I believe that if you get this information, you should review to see if it makes sense. Um, but it can help you. I mean, it's not, it's not that one thing replacing into another thing because we still have sentiments.

We have things that are different. Right. So it could be more, uh, fine. Yeah. But 

[00:50:41] Tony Mongkolsmai: philosophically, philosophically, Doug's question was can it can an ai, because it generalizes computing generally is explicit, can it think like us? And the answer is if it's explicit. No. But the interesting thing about the neural networks, and not necessarily neural networks, because they aren't necessarily a fine, an exact rep representation of how our brains work, but there's definitely, they're definitely more going towards the direction of, yes, they also can't define how they think, but yet they still generate things that seem coherent, right?

Mm-hmm. , you can't mm-hmm. , you can't describe how your brain works, that AI can't describe how it works. We also can't describe how it works , um, in cases. So I don't know if it would be fair to say it would never get there. I don't know if the, the representation that we have of a deep neural network makes sense.

There's other types of computing like neuromorphic computing, which is theoretically more similar to how our brain works and how our nerves work. So if we could replicate that with enough compute horsepower in the world, is it possible? , it probably is. I don't know if we have enough compute power in the world in the way that we represent computation and computer hardware to make that happen anytime soon.

But it seems like, again, philosophically as a thought experiment, it should be possible at some point. But I don't think we're close, 

[00:52:04] Katherine Druckman: although we're getting awfully generally close on the, the uncanny valley stuff. We gotta, I dunno if y'all saw the, the Morgan Freeman video that was going around 

[00:52:12] Tony Mongkolsmai: of the robots

[00:52:16] Doc Searls: So as you're talking, I thought, oh, an interesting thing is of course, because it's just the text thing so far, it has no body language. You can't read it, you know, and when we communicate to each other, we use, we are bo, we're embodied animals. It's not just our brains, it's our whole bodies and how we express ourselves.

And um, I mean one of the problems we have with privacy, and we've talked about this on the show and elsewhere, is that, um, Here in the natural world, we signal to each other, no, don't come any closer. Uh, this is okay, I'm ready for a hug, whatever. And we communicate things about our privacy preferences that again, are ra largely tacit, but to some degree they're explicit.

You can say no or, you know, don't want this, or whatever it might be. Um, but I suppose it's possible. I was thinking of the movie Ex Machino, which has, you know, which is assumes there there is some kind of near perfect AI intelligence that, that is capable of understanding people psychologically and um, and it's all in some kind of wetware, which nobody has really built yet, but, and may never, but at least it's imaginable.

But even there, I mean, you know, not to, I don't think it's any spoiler to to say that you really can't trust the damn thing at the end. At the end, right. , you know, so some humans gonna have to be in charge of that. And not just a yet an neither. Another Elon Musk type. . 

[00:53:40] Katherine Druckman: I, I, I think suffice it to say, and, and we've, we've gone for a while now.

We should probably wrap up, but yeah, I think suffice it to say right now, I think it could probably, maybe not ChatGPT between chat G P T and some other, uh, image generators and video generators out there. I bet it could come up with a, an episode of this podcast, but would you wanna listen to it? Eh, I dunno.

I think, I think that's the, that's where we are in, in, in the, in the process. 

[00:54:09] Tony Mongkolsmai: It comes back to the unknowable though. Why does anybody listen to your podcast or my podcast or anybody's podcast if we knew that? Oh, I would never know. . Yeah, we would be much more popular. That would be amazing. Yeah. I wish I 

[00:54:20] Katherine Druckman: knew.

Why do you let us know? Um, yeah, please, please feel free. I mean, we, we, I didn't do my normal, uh, thing at the, in the intro where I thank everyone because we do get a fair amount of feedback and, and I value it highly and we enjoy it. We get a lot of support, uh, which we appreciate, but I I, but the reasons are may be unknowable.

Yeah. I think we're, uh, we're, we're, we're coming up on, you know, running out of time. I did wanna say a couple things. Um, I wanted to make sure if there's anything last, last, uh, little pearls of wisdom that you wanted to share that we didn't cover, but also, I wanted to make sure to give Tony an a little opportunity to talk about his other podcast just for people who might be curious.

Speaking of, well, I, 

[00:55:01] Tony Mongkolsmai: I host a podcast called Code Together. Um, we like to talk about, uh, technology. It's. Mostly for developers, uh, in some sense, but we just really, I like to talk technology much like this podcast. Um, so hopefully if you're interested in technology, you could check it out and most of the time we'll be talking about, uh, concrete uses of technology in the world.

Um, and we'll talk about how that applies to developers along the way. Um, so yeah, if, if you like kind of the technology chat that we've had here, you can check it out and give me feedback about why you do or don't like it. That'd be great. . 

[00:55:36] Katherine Druckman: Yep. Always, always appreciate it. Yeah. Ezekiel, did you have anything that you, you had in mind that you wanted to make sure to get out there and haven't yet?

Or did we cover it? No, 

[00:55:45] Ezequiel Lanza: just, uh, I mean, we should be. Relaxed. I mean, it is not the end of the, of the world with the technology. I mean, it's, it's the start of the 

[00:55:53] Doc Searls: world. It's the start of the war. I mean, it's, yeah, there we go. More 

[00:55:57] Katherine Druckman: optimistic, take Exactly. 

[00:55:58] Ezequiel Lanza: We'll find some really interesting use cases to have this, this, this technology in the field and to see it.

Um, but I mean, let's try to, to be optimistic. And this is pretty awesome because the moment that we have now, we have data, we have processing, we have a lot of things that are possible, and the technologies, it's ai, it's advancing a lot in the last five years, four years. And it will be for sure in the next five years.

In the previous generations, you should wait 20 years or 30 years to see an advance. Uh, now it's, you see it almost every day, every week you have something new. So it's, we are, we are very, I'm happy to, to be alive in this, in this, in this generation. Right. . 

[00:56:46] Katherine Druckman: Yeah. I, I enjoy your optimism. , I guess. Yeah. It, yeah.

If there is a takeaway and, and, you know, coming back to our, our op, open source, uh, bias is, is to remind ourselves there's this great opportunity, we're in a great mo an interesting moment right here. And this is the, the time where if this interests you, this is when you jump in and you get involved and you steer it the direction that you think it needs to go.

We're all empowered to do that as humans. So, so this is, this is your opportunity to, to get in there and, and, and make your mark and to open 

[00:57:14] Ezequiel Lanza: source what we are doing. Exactly right. 

[00:57:16] Katherine Druckman: Yeah, exactly. Mm-hmm. and open it. Did you have any final thoughts, doc?

[00:57:21] Doc Searls: There's that expression.

It's at the end of the world, you know, but I think it is the start of a world. I think it's the start. I've had a number of people tell me that, um, these are people, you know, in tech, uh, whose life has already been made easier in some ways by ChatGPT itself, and not, you know, one's a ceo. You know, another one's a writer, and in neither case are they busy plagiarizing ChatGPT.

It's kind of like, Hey, this is ano, they're looking at, it basically is another way to do search. You know, it's a, it's a, it's a different ui, it's a new hack on search. That's it. Yeah. You know, but the, the big question with 

[00:57:57] Tony Mongkolsmai: that is, um, and we did talk about this in the podcast that I had about this, which is you don't know the providence of that, and you get one answer, right?

So if you ask it to do something and you ask the wrong prompt, the wrong question, it may give you an answer that is incorrect because you've prompted it wrong. Now, most people, you know, if you're in a position of a ceo, you know how to ask an appropriate question without a bias. But if you ask a question that is biased in Google, at least if I type something in the search bar, I get a list and I can scroll through that list and identify which one I think is the most valuable.

currently as implemented when you interact with ChatGPT gives you one answer and if you don't like it, you can say, give me another answer. And you keep doing that till you get an answer you like, and it continually changes the answer. I don't know if you guys have tried that. Yeah, yeah, yeah. But it's harder because you don't see it all upfront.

So now you don't know when to stop. So the, the interface itself of the application, not the model in the AI, is potentially challenging for using it for that type of use case. 

[00:58:54] Katherine Druckman: Yeah. So it may or may not know who Doc Searls is, depending on how you ask 

[00:58:59] Doc Searls: question, it doesn't know. I think that's kind of the important thing.

It, well, it reports and it's got, you know, it's got programming, it tells it how to report. 

[00:59:07] Katherine Druckman: But well, cool. I think, uh, yeah, I think we're, um, I think we've, we've sufficiently uh, we've sufficiently covered it. And, and well, no, we totally haven't. We have just begun to cover it, but I appreciate both of you joining us, joining us for this.

It's really nice to have y'all come over, you know, and hang out with me and Doc and, um, yeah. And thank you for everybody who, who has listened up to this point and we will talk to you next time.