![](https://images.squarespace-cdn.com/content/v1/5e9884bbf25f9a7326953521/1674623287727-E05AM4C9D1VQUI89VBNZ/ChatGPT.png)
This recent natural language chatbot has been getting ALL the hype.
In this episode we interview Deep Dhillon about the ins and outs of ChatGPT!
Deep is the co-founder and leader of technology development at Xyonix, where his mission is to find novel value in clients’ data. Deep has experience as a technology executive; conceptualizing, architecting and deploying advanced applications, leveraging machine learning, natural language processing and data science to build smarter businesses and more powerful products.
Follow Deep on Twitter @zang0
Listen to Your AI Injection Podcast
If you enjoyed this episode please make sure to subscribe, submit a rating and review, and connect with us on twitter at @radicalaipod.
Relevant Resources Related to This Episode:
Try Out ChatGPT (if you can get through the queue)
Exclusive: OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic
Socrates on the Forgetfulness that Comes with Writing
Ghost Work: How to Stop Silicon Valley From Building a New Global Underclass
Transcript
chatgpt.mp3: Audio automatically transcribed by Sonix
chatgpt.mp3: this mp3 audio file was automatically transcribed by Sonix with the best speech-to-text algorithms. This transcript may contain errors.
Speaker1:
Welcome to Radical Eye, a podcast about technology, power society and what it means to be human in the age of information. We are your hosts, Dylan and Jess. We're two PhD students with different backgrounds researching AI and technology ethics.
Speaker2:
All right. So you've seen the headlines, you've heard the news. You might have even visited the website and tried it out for yourself. Chatgpt The recent natural language chatbot that has been getting all the hype is the focus point of this episode. In this episode, we interview deep Dylan about the ins and the outs of Chatgpt. What is it? How does it work? Should we be excited or should we be scared?
Speaker1:
Deep is the co-founder and leader of technology development at Ionics, where his mission is to find novel value in clients data. Deep has experience as a technology executive, conceptualizing, architecting and deploying advanced applications, leveraging machine learning and natural language processing.
Speaker2:
And Deep also has a podcast of his own, which we recommend you check out. It's called your AI injection and we will link to it in the show notes if you're interested. And with that, we're going to jump right into the interview. All right. We are on the line today with Deep Dhillon Deep. Welcome to the show.
Speaker3:
Thanks so much for having me.
Speaker2:
Absolutely. And today, we are going to talk about the ins and the outs of Chatgpt. So if you wouldn't mind, we'd love if you could just give us a brief tldr an intro, if you will. What is Chatgpt? Why is this all the hype right now?
Speaker3:
Yeah, that. I think the short answer is not to be too hyperbolic, but it feels like a like our community is split the atom. Maybe three weeks ago, four weeks ago, when Chatgpt came out. That's not to say that. Chatgpt was a huge change over like the rudiments in the large language models that were there before, but to be super high level about it. Openai The folks at OpenAI have built a system that takes the vast, vast quantities of publicly available, unstructured, natural language content you know, news articles, blog posts like, you know, everything out on the web in different languages, software code, you know, things from from hundreds of different languages, tons of different programming languages, tons of different just just all of it. Take it, take it all and then train up a large language model. And the easiest way to think of large language models are there are these sort of complicated neural network architectures so similar to the human brain where we have biological neurons that work based on sort of electrochemical signals, you know, like a neural networks, you know, is sort of a mathematical representation that's sort of similar. And this neural network architecture has to be trained on something. So typically, you know, the way a toddler learns about words like learns a word, like furniture from examples, you know, somebody might point to a chair and say, you know, this is furniture, a table, this is furniture. And, you know, and maybe a sofa and say this is furniture.
Speaker3:
And then the is like, you know, points to like an outdoor wicker chair and says, not furniture. You're like, no, no, no. That's actually still furniture. So, so similarly. So we need these sort of examples. And what OpenAI has done and what everyone who trains large language models typically does is you have to train this complicated neural network architecture on some sort of example dataset. Well, it turns out that in order to teach something of this scale, what we do is we basically train it to to predict like future words, so and sequences of words and sequences of characters. So, for example, I have to go to the and then then the model has to like, guess what? The next word is like store bathroom. But it's not going to be, you know, orangutan and it's not going to be, you know, giraffe or something completely different. And so the model basically is then trained. And so typically, like these large language models, one of their features is they're really big. And, you know, like I think I'm not sure exactly. I think the OpenAI one is about 178 billion parameters. So and then there's multiple layers in the architecture of of these neural networks and they're basically have to like kind of formulate in the same in a similar fashion to like a, like an infant when it's born, it doesn't know about anything. So they go through this training process and it takes millions of dollars to actually train it up on all of this language.
Speaker3:
And the model ultimately gets really good at predicting future words and sequences of words and characters, irrespective of language and language and even whether it's, you know, it's human language. And the artifact of that is that the models learn an insane amount, like they basically learn way more than a given human can in some senses. So most of us, if we want to learn French, you know, or, you know, Swahili or something, have to, you know, maybe we download Duolingo, maybe we take classes at a college or university. Eventually we, you know, it takes us years to get good for most of us who aren't polyglots and. But the model, sort of all of that stuff just shakes out by this like mirror sort of process of training it on future sequences of words and text. And it ends up that it understands language. And so that's the kind of base layer that we've had for the last few years. And what OpenAI did that was different when they built Chatgpt was they put an another layer on top that, you know, in the machine learning or AI world we call reinforcement learning. But it's you can think of it as you you can ask these large language models, you can prompt them with something and it can be incredibly broad. Like you can prompt it with, you know, some French like and then the English translation and some French and some English translation.
Speaker3:
And then you give it some French with no translation and the model finishes it. And so that sort of prompt completion is a very general rudiment that you can use for conversation. You can use it for all kinds of things. But then the OpenAI folks went a step further and they took those outputs because sometimes the model does great and sometimes it doesn't do so great. And they trained a layer on top where humans basically said, Don't like that completion, like that completion, don't like that one, something like that. We don't actually know the details because I don't believe they've published it as of yet. But everybody's speculating and stuff's leaking out and people know how this stuff works in general, and that's basically what it is. So now you chatgpt as manifested today. Anyone can just go to the URL. I don't remember what it is, but just Google Chatgpt You wind up there, you log in and it's just a text field and you just talk to it like it's your genius grad school buddy or your, you know, person who just every once in a while just blathers things that don't quite make sense but sound like they do So. But generally what's spectacular about it is it's really good. Like better than anything we've seen before. Before we get into the applications.
Speaker4:
I'm curious about the data that goes into this and the data collection, because we know from some computer vision case studies sometimes that can be a little thorny about what the outputs in the data in and the data out and where that data is coming from in the first place. And I'm wondering what your thoughts are or if you can shed some light on where this data is coming from in the first place.
Speaker3:
Well, it's hard to say exactly what they trained on because I don't think they've published it yet. But most of these models use a ton of what's just publicly available. So, you know, like when Google goes out and crawls, it's that stuff. So it'll include Wikipedia, for example, like, you know, all of Wikipedia and it'll include like public repos of source code, like software code that computer scientists work with. It will most likely include online forum discussions like Reddit. It will include just pretty much anything that's easily accessible where you don't have to argue with a bunch of lawyers to get it. And. I think in the case of OpenAI, they are also probably leaning heavily on it seems like they're doing a lot with expertise and like credibility, because one of the challenges with these systems as Microsoft sort of learned the hard way a few years back, is that, you know, humans aren't that great all the time, like a lot of us are just neo-Nazis, you know, and spew all kinds of horrible stuff. And you don't want the model to just blindly learn all of that too, and start applying it. And so credibility is really important to use in these models.
Speaker3:
So if you're going to talk about health care, then you want to talk about PubMed articles, you know, like and within the world of PubMed, for those who don't know, PubMed is sort of the global repository of of everything medical. So any research paper published by anyone and, and it's freely available in wide open. I think there's last I checked there was like 30 million or so articles in there and so you want to like give more emphasis to that on biomedical stuff or and, and even within that universe, like, not all of that content is created equally. So you might want to graph so that, you know, so that stuff that's coming out of nature, the journal Nature, which is very credible versus something coming out of some like obscure university and some, you know, not so sophisticated university system in a country that's not known for publishing things and having high publication quality. So that stuff will get penalized. But long answer, sorry, do long answers all the time. But like I'll try to. I don't do as well as at summarizing information as Chatgpt does. It's much it's much better than I am.
Speaker2:
So we can we can feed in how we should script a podcast episode about Chatgpt to Chatgpt and we'll just replace this interview with that. It's actually been.
Speaker3:
Done. There's a good one. I can't remember who did it, but it was pretty good.
Speaker2:
Oh, no, of course it's already been done.
Speaker3:
Yeah, but there's plenty of room to do it again.
Speaker2:
I do have a follow up question about this. I guess like diving into the the data part of Chatgpt, which I think is something that's generally like almost obfuscated from a lot of these like open AI platforms. I think people just see like the output and the end product, but not a lot of people get to see like a a close up look at what data is actually being used to make these products. And there was this article that came out really recently, I think a few days ago from time that's titled Exclusive OpenAI used Kenyan workers on less than $2 per hour to make chatgpt less toxic. So this has been like one of the more recent scandals that's come up is like, oh, so Chatgpt has this like extra level of reinforcement learning, as you were saying, where, you know, all these people are being paid to tag what are the good and bad outputs of this article? Or maybe like medical students are being paid, for example, to help train what are like useful and not useful outputs for medical prompts, things like that. And so I'm curious like what what are some of the ethical concerns of, um, paying or underpaying or not paying people to do the hidden labor to create the data that fuels this kind of technology?
Speaker3:
You know, that's that's a really good question that I don't think gets asked enough and I don't think people really understand. But there is a dark underbelly of machine learning, of low paid labor that, um, that powers this stuff. And, and they're not necessarily, you know, the. Um, most fulfilling jobs. Now, in this particular case, um, I think you have to even I don't think it's like, ethically simplistic to even address. Right. I mean, I've, I'm from, from India. I've spent a lot of time in India, and $2 an hour sounds horrible and downright, you know, terrible to somebody sitting in San Francisco or Seattle or, you know, or anywhere in the West, really. But, you know, that's about $100. Um, you know, sorry, $2 an hour is like 16 bucks a day. Like that's that's pretty decent money in some parts of the world. Now, I don't know, in Kenya in particular, um, you know, it's been at least a couple of years since I've been in Africa. But I think that's also part of the equation, you know, and it's also fair to ask whether even if it is high on the spectrum, if if it's the right thing to do to allow this type of work to go to, you know, wherever it wants to go to on the planet.
Speaker3:
And I think those are more complex questions. But, you know, a place to start might be some transparency around it. You know, like saying I don't think it's the necessarily, you know, bad that it's not inherently bad, that a lower wage means that somebody was exploited. Right. If if it's like, you know, in the top, you know, above a couple of standard deviations above the mean in the place, it can still be a good job, you know, So, um, so I think that's kind of part of that question. But I think another part, like at least with respect to the West that we think about, is just the contract only nature of a lot of that work, you know, and frankly, like sometimes it can be dehumanizing because it's just completely algorithm driven. You know, if you're on Mechanical Turk, not to pick on, you know, the Amazon folks because there's many of these platforms out there, but there's definitely a race to minimizing cost. So that and there's a there's like a drive to make them as simplistic as and like easy to come up to speed on how to do these human trained jobs. To give the examples to the toddler. And the example that I gave earlier, you know, so that we can have lots of them and say that's that's kind of part of it.
Speaker3:
But I would also go a little bit further and just say like, don't one of the main like amazing things about these models, like the large language models is the. The. Markedly reduced amount of training data. We need to get them to perform well. So if you rewind like five years ago, you know, it was pretty common. If we had a machine vision problem, like we would immediately be talking about, you know, tens of thousands, hundreds of thousands of training examples. But, you know, with large language models like the right three examples goes really far. You know, like probably is further than ten years ago or five years ago, even with the 10,000 plus examples. So. I don't know how long that particular problem is going to be here. And it's also a bigger deal when you are getting your network off the ground. So OpenAI is like pushing this thing up and trying to give birth to this, you know, this machine learning system that's quite advanced. But once but now it's in everyone's hands and all of us have the thumbs up and thumbs down on the responses. And all of that's going to get used in future incarnations. So.
Speaker5:
Yeah, Yeah. I'm curious about.
Speaker4:
Openai because it's OpenAI. It's not Google, it's not some of these other companies and they've kind of built their brand off of democracy and for all. But when Chatgpt gets used now, it it both belongs to everyone. But then also openai. But then it's also being used by other companies. So guess who owns this and does anyone own this?
Speaker3:
Well, I mean, I think we have to define this not to be too Clintonesque there, but like. I don't think there's I think this is a question that will evolve over the next probably 10 to 30 years in the same way that we ask these questions. When the Internet came out and and started rapidly developing back in the early 90 seconds. You know, at least with respect to the Web and. I don't think they're necessarily straightforward in some senses. Like if so, if you have some text that has, you know, a Creative Commons license on it, then it's fair game according to most readings of the law. And you can do whatever you want and you can, you know, as long as you kind of provide attribution on some level, I think it's like at the paper level, like if you write about it. So OpenAI can still claim ownership over that. But then there's, you know, there's like a bunch of gray area where somebody puts something out. It's accessible on the web, but you're not quite sure. It doesn't necessarily have a license on it. And so depending on what lawyer you talk to, like an aggressive lawyer working with a startup will say, sure, take it and then we'll deal with it when they complain.
Speaker3:
And, you know, a conservative lawyer working for like, you know, a large firm in a conservative industry would say like, no, you can't touch it. So so there's that kind of question. You know, it feels like we probably need some clarity. But most people sort of at this point will default to Internet precedents and say, well, look, pretty much whatever Google gets away with, these guys can probably get away with. Um, and then there's just the question that I think you're asking about, like is openai really does the open part in it mean anything? I probably would say no, not not anymore. Not after. I mean, like, you know, the rumors are that Microsoft's going to put in $10 billion and that they have kind of first rights. And the OpenAI folks are kind of trying to straddle this world between a nonprofit and like an insanely profitable thing. So I think, you know, maybe, you know, the the founders don't get the the hundreds of billion dollar payout that they would if they were. I don't even know. Are they. They're not a C Corp are they a B corp or something like I don't know what their tax status is.
Speaker5:
That's a good.
Speaker4:
Question. I was I was looking at their website a second ago, so maybe we'll put it in the in the footnotes. And I couldn't really tell at the moment.
Speaker3:
So I mean, it feels to me like there we might as well just think about them like a C corp. I don't I don't see like a big difference at least because they kind of just deciding like, okay, first X number of years we're going to, you know, we're just going to run and grow because we're worried that if we don't, you know, clearly C Corp like Google are going to do this anyway. And so there has to be an open a more open alternative out there. And I think that's their positioning. But practically speaking, I don't know. How much it it matters given how much, you know, public, private or how much private companies give back anyway in the form of open source like Google, Facebook, Microsoft, like everyone's giving pretty substantive, freely available models. You know, there's and then there's the hugging face, guys. So I think that's a real thing. You know, like there's a lot of very ethical people in in the sophisticated kind of deep tech companies that are trying to do the right thing and like open up. But you know that for whatever Google releases wide open to the public, they've got at least ten x that behind the curtains. And so they're kind of doing the the like, you know, what the DOD has done for years is like, just stay ahead, you know? So I don't know. Long answer, but the short answer is. We'll see over the next ten years, this stuff will play out.
Speaker2:
Well, on the topic of openness and what's being fed into chatgpt, what's coming out of Chatgpt. It seems like there's sort of two big issues right now, or big concerns or questions surrounding like what? What is allowable to be used. And so on the input side, we have like what text is Chatgpt allowed to use to train it and like you were just describing? And then on the output side, we have, well what text that we gain from chatgpt are we allowed to use? And so this is what we kind of start to see these conversations happening right now about like plagiarism especially and like universities and schools. And what's happening is, you know, student logs on to chat openai.com they input, okay, this is my homework assignment. Chatgpt has this very eloquently worded paper that it writes for them, and then they copy paste into their canvas course module. They get a great grade and teachers are now freaking out because they don't know how to tell whether something was written by Chatgpt or not. Because Chatgpt clearly passes the Turing test. So what? What do we do with this like difficulty of understanding what we're allowed to use from chatgpt? What are the conversations that are happening right now around plagiarism?
Speaker3:
Okay, so let me take your first one first and then the second one I think is the deeper conversation. But on the first one, like what can Chatgpt use? I think, you know, in the same way with the web where we had the robots.txt standard put that in quotations because it's kind of a soft standard emerge, you know, that basically gave people the rights to not let the bot into stuff. The crawlers like Google's crawler and Bing and etcetera. And, and you know, like maybe it wasn't the world's best solution, but anyone who doesn't want their content crawled, you know, if they follow the robot standard, it works out. And I think we'll probably we already have one. I just don't know of it off the top of my head. But I'm guessing if I Google it, there's probably something already exists and maybe somebody can put it in the, you know, in the show notes. But if not it, I'm sure it will. Like where where people who have creative content that they fundamentally don't want used in large language models can do something to to like block it out. And I think that will probably largely address the problem and we'll move on. Um, with respect to the plagiarism question, I mean, that's, that's kind of like a big question, but the, the super short answer I would have is, um, I think. I think these systems are are here to stay. They're only going to get better. They're only going to get better at writing.
Speaker3:
They're only going to get better at writing better than the vast majority of people. They're already better than like 99% of the people I know. Um, and. And so I think we have to not just learn to live with them, but I think we have to learn to embrace them. You know, if you rewind to I don't know when the calculator was invented, but I'm sure we had mechanical ones, you know, before. Like my guess is probably in like the mid 1800s or something, but probably even back to the 1500s, there's been some kind of calculator. But you know, like mass adoption of calculators with electronics maybe in the 50s or something, they became cheap. We learned to live with them. And, you know, educators need to do the same, a similar things with Chatgpt. I think they're asking the wrong question when they're like, you know, Seattle School District, you know, typically in their impulsive way, just flat out banned them, which I thought was completely ignorant. But instead what you really need to ask as an instructor is like, Hey, how do I give more challenging problems and how do I just blatantly embrace this stuff and use it as a learning tool? And I don't think it has to be so. So like, it doesn't have to be like we don't have to only aspire as humans to do the things that we could do before the technology. We can aspire to do more.
Speaker3:
And these systems are so powerful that there's definitely a way to leverage them, like if you think about it. I mean, like the analogy I kind of use with chatgpt is, you know, I can go so, so here's, here's an example that I that I like my favorite use case with Chatgpt. Like, I've been obsessed with union therapy for a while and like, just, you know, Carl Jung and all these ideas of the subconscious and stuff. And so I always wanted a union therapist, but they're, you know, they're like at least a couple hundred bucks an hour and they're usually backed out and it takes like 5 or 10 years to even get ahold of one. And like in Seattle, I think there's only three. So I take Chatgpt and I just all I do is I say act like a union therapist and ask me a question. And then I interact with it every day. I talk about, you know, and then usually every morning I like we analyze my dreams and like, boom, whatever. So, so what I have to do with Google to get the same, which is what I was doing before three weeks ago, was I would wake up in the morning, I'd have my dream and I'd write it down and then I'd go Google like the symbols in the dreams, and I'd get most of it would be complete, like New Age garbage that I'd have to wade through until I could get to like a legit definition of like a discussion around the symbology.
Speaker3:
And then and then it would take me like 20, 25 minutes to analyze the dream. Chatgpt does it like boom in usually somewhere between five seconds and, you know, a couple of minutes of back and forth conversation. That's that's like a case where a lot of value is delivered. Chatgpt got, like, nothing out of it. This nobody who talks about chatgpt talks about it as a great union therapist. In fact, most of the therapy world doesn't even talk about union therapy anymore. They just talk about some other types of therapy. So. I think of it as like. You know, it's like you have 50 people that report to you personally and you can task them with reading a ton of stuff and they can like read it and assimilate it. And now you can just put them at the whiteboard and like talk to them and that's chatgpt. Now, how is it that, that, that we can't figure out how to take advantage of that in an educational context? Someone else is like reading for you, assimilating information for you. So now you can like operate at a higher level and you can just learn so much more. And it just feels like we need to think harder about how to use this stuff to train. And whatever we did with our curriculums and our sort of, you know, industrial era educational system, I don't think we have to keep doing that.
Speaker5:
One thing that's striking me about these different examples is that they're really.
Speaker4:
Different and they're at different parts of human society of plagiarism and schools on on one end, and then you're using it for Jungian psychology on the other end. And I imagine there's a million other use cases in there. And I guess I'm it seems so like it can be used in our everyday lives. And I'm wondering how you see that greater transition of being able to use maybe similar to a calculator, maybe something much bigger or different, but it really being deployed everywhere. Maybe another way to ask it is are there places where it either shouldn't or can't be deployed? Or is this like here to stay in part of our daily lives?
Speaker3:
So. I think there's like two ways to look at that. One way to look at that is given that it's a chat thing and chat back and forth with it, how would I use it? And then the other way to look at it is given that it's infrastructure and it's got programmatic interfaces on it and they have this global ecosystem of technologists that can take it and embed it into a bajillion different projects and products and all of that. So I think the first part of the question is. It's like. Maybe. I don't even know how to answer the first part of it because I don't think that's the part that matters. I think it's the second part that really matters. The first part is like like OpenAI has done an amazing job at like painting the picture so that regular people can grab this thing and then however skeptical they were before they walk away going, Oh my God, this is crazy. And but most of us who've been working with this stuff, when we saw it, we're like, wow, that they went the last mile. But we've been getting this level of quality of results in narrow areas for for a few years now. So we're, we're kind of used to it. Um, but in the so I don't know how the conversational part will, I think people will just, you know, well I guess I'll take a few guesses.
Speaker3:
One is I think we'll be able to talk to it, which will change the dialogue stance a little bit. Um, I think we'll get better at knowing how to communicate with it. So one thing I noticed with my wife, she's a, she's an artist and not a technologist, and I often use her as my kind of guinea pig test for how usable is something. But she had no idea like how to prompt it, even though it's it's really simple to prompt after a few minutes of watching how I was interacting with it and how I was like describing things, she's like, oh, I get it. And then she's gone. And then. And so I think that those little problems, those little socialization techniques will, like, be learned really quickly. The way that we learn to formulate queries on Google, right? Like, you know, we used to have, I think, 1.7 keywords on average 12, 12 years ago, 15 years ago. And now that number has gone up because, you know, we can kind of like write more and we've got we can kind of most people know how to find things on Google. And I think the same kind of thing will happen with Chatgpt where most people will know how to have the conversation about anything to get the kind of knowledge out.
Speaker3:
So I think that will continue to happen. But I don't think let's talk about the other side of it. So actually there's even a middle tier. So the middle tier is even constrained by conversational element. I don't think we're going to mostly be interacting with a totally generalized model like Chatgpt, right? Like I mentioned, union therapy, but like, like somebody like probably there's, I don't know, 20 companies right now, startups that are that are trying to build a therapy driven version of chatgpt trained just on therapy. And within those, they're probably like a bunch of them are just doing like CBT and style therapy. There's there's going to be like tons of narrow chat interfaces, like every chat bot on every website that says, Hey, how can I help you? Every one of those make like whoever's making those systems, they're all figuring out how to how to use Chatgpt because everyone looks at that and says, That thing was dumb. This thing does way better. How do I use it? So all of that's going to continue to happen, but the parts that I think are even more interesting is like, so take a company like like Google. So like, you know, they just pulled the red alarm and they're like, okay, you know, we've been way too conservative. Everybody's been like, you know, hanging out, doing the same thing.
Speaker3:
We haven't been releasing serious stuff probably since Sergey and Larry Page left, and we need to do something big. So what did they do? Well, they're probably going to do what Microsoft did, where they go to every single project manager inside of Let's just take Bing as an example. Like the search engine, every in Bing already competes on keywords. So they're all they're all like machine learning, people working inside of a big machine learning system. So you, you basically define a set of keywords like, oh, I'm going to grab everything to do with, you know, with. I don't know. Let's just stick with the union therapy thing. I'm going to grab every keyword that has to do with union therapy, and I'm going to, boom, pass it through chatgpt and then I'm going to like have a really niche system to assess whether or not that answer is better. And if it is, I'm going to one box it meaning it's going to appear above all the search results, and then that's going to go into the Microsoft system, which is then going to figure out if people engaged better with that answer. And if so, then that project manager inside of Microsoft gets more queries like that. And until eventually they're like the new king of, you know, of that narrow class of query. And so now you roll that out at Google across, you know, 100 or 200 projects.
Speaker3:
You know, all of a sudden, you know, this whole thing emerges in a way that we don't even talk about because it doesn't say chatgpt anymore. It's just an answer to a question. So I think that kind of stuff will happen. But now take it like way outside of, you know, big tech, you know, like if you take it, you know, like, let's see, you know, like we've got a ton of different projects like, you know, because we're an AI consulting firm, like one of our one of our firms, you know, we're building models to like people upload videos, you get these like sort of automated transcripts extracted from there. And then there's just like this automated summarization. So like here's the, like, you know, eight topics that were discussed in a long video like this conversation we're having right here. Right. And those and so so it's like, okay, well, we're using large language models to do that. And then that's like an example. And then I would say like even beyond large language models, just large neural nets that are similar, similar things are happening in the vision arena, in the audio arena, you know, like it's just like the whole world is just radically going to change, you know, in the next ten years.
Speaker2:
This is something that I've been thinking about a lot since Chatgpt came out. Is this this radical shift that it's sort of symbolizing? And in my head, I've kind of been going on this journey as you've been speaking of like what um, information retrieval has been like for society pre and post internet, I guess. And I'm thinking like, I don't know, maybe back in the 1800s somebody would have a question about something and they would go ask their neighbors like, Hey, do you know the answer to this question? And the the capacity of their own knowledge is constrained by what their neighbors are able to tell them to answer that question. Right. So then the Internet comes along and now you can ask Google that question. And now the capacity of your knowledge is constrained by what the world on Google knows to be the answer to that question. And Google now has this duty to, like, coalesce so much information to make sure that you're not overloaded with billions or trillions of lines of text as an answer to your question, and instead now just gives you like, oh, here's the one suggested answer to your question that we have discerned from so many different possible answers we could have given you.
Speaker2:
And so they have this duty to give you the, quote, best, most accurate, correct answer. Right. And there's so many ethical considerations to be had about what that best answer might be. And now we take that even a step further with Chatgpt, which now takes sort of in the same way that Google has taken this like idea of ranking and optimization to find that best answer. And it's doing this again. But I feel like I'm concerned that there's maybe not quite as much ethical consideration that's been had around how chat Chatgpt is doing this. And I'm this is kind of a bit of a rant at this point but I know there's a question in here somewhere. Like where what what do we do with this? Like is how how are we what are we to do with Chatgpt potentially being the future of our, our ability to, um, to answer questions and to come up with answers to questions. And it is in a sense, our access to information. Like what, what considerations should we really be prioritizing in this, in this regard?
Speaker3:
I mean, your comments make me think of a conversation. I can't remember what it was. I thought it was Plato or Socrates or something. But there was this this whole debate in in one of these old, uh, I think it was Plato, but where they were arguing about what would happen to the human mind given that people could write things down. Right. There was this prior to that that that moment. There was this long oral tradition about being able to talk and orate and, you know, and say things eloquently. And and it was kind of considered to be like a fundamental part of being human and being a good human was being like a good orator and having good strong oral skills. And there was this deep fear that writing things down would offload the the in resonant memory required to be a good orator and hence a good person. And I think there's something similar going on now where what we think of as good thought and good thinking capability. It's it's like we sort of associate it with work on some level, like being able to go out and find stuff. So, you know, I'm of the generation that went to college. I mean, you know, I went to a tech school. We had, you know, the Pre-web Internet, and I had, you know, Archie and Finger and all that and all that stuff and Usenet and everything access to it.
Speaker3:
But, you know, we still pretty much went to the library and, you know, it was still a need to go peruse the journals and physically grab a journal or physically find the right topics and content to pursue a project or write a paper. And so that skill sort of mattered. And then, you know, the the web kind of blew up and eventually, you know, a few years later, I have a laptop and I've got access to the world's information on my laptop. And mostly I kept my laptop at work and it was like, you know, eight, ten, 11, 12 hours a day, however long I wanted to work, I had access to it, but I didn't really have access to the world's information on Saturday and Sunday. For the most part, I didn't really care. I was doing stuff. Fast forward to 2007 iPhone. Now it's in my pocket. Fast forward to now. It's like it's still in my pocket, but I can talk to it. It's in my Google home, although check that out into the yard one night when I was sick of it. But. But like. I think proximity to information is going to just shrink like it's going to be like at the speed of thought and maybe in the next 30 years or something, I don't know. So then you have to ask the next question like, well, what matters in that world? And I think this is kind of part of what you're speaking to is the threat to diversity of thought.
Speaker3:
Maybe, you know, like if everybody is. Is like driving too much of their thinking based on this ease of access to two. Two systems. Then maybe we just don't have like the diversity of thought that we used to. And there's a fear there similar to, you know, Plato or I wish I could remember who. Plato or Socrates fear that we wouldn't be able to remember things. And I think I think that that's, that's real. But here's where I think the education system needs to really change. One of the things that I feel like the, you know, my, you know, the education system really fails at is it spectacularly fails in getting students to ask the questions. And, you know, we talk about the fact that it's so easy to get answers on Google and it has been for 20 years or so now. And it'll be even easier with OpenAI. But what I see in, you know, I've got a couple of young adult friends and all of their friends and then, you know, and then just in the kind of youth world is they don't even know how to use Google half the time like they know how to. But it goes into this like academic bucket, but they don't necessarily go there when they need to and they don't. And and part of that reason is they don't have like they don't they're not like like intellectual curiosity.
Speaker3:
It feels to me should be like the fundamental. Thing that matters in an education because we have the ability to answer questions. But I see the opposite. Like my son, at one point, I think he was like 13 or 14. He said to me, he's like, he's like, Dad, school's only good for one thing. And I'm like, What's that? He's like, It's good at like, destroying everything I love. And I'm like, What does that mean? You know, it's a bit hyperbolic and a little melodramatic. He's like, Well, I used to love reading. Now I hate it. I used to love, you know? And he rattled off like five things. And, you know, it basically came down to, you know, like you have to create a love of mental work. Like it takes it takes you have to love asking questions and then you have to, like, create a love around getting those answers. And if you fail at those two things, it doesn't matter that all the machine learning systems can help you out. You won't even use it, You know, like that. Like the the story I always tell here is, you know, one time I was I was like out in the the village in India, like sitting around talking to a bunch of folks. And in this particular area, like, everyone's just kind of like, you know, you know, it's hot.
Speaker3:
So like, people don't necessarily want to think. And there's just like a folk tradition of this talking for the sake of talking. So everybody's like talking about they were talking about like how much an acre of land because they're farmers costs outside of Vancouver, BC in Canada. And they talked about it for like two hours. And I was like just kind of ignoring it like I was, I don't know, reading a book or something. And then finally I jumped in and I'm like, There's like six of you. Can everyone, like, reach in their pocket and pull out their phones and everybody reaches in their pocket, pulls out their phones like. Now I want you to go to Google and I want to I want you to ask a simple question. How much does one acre of land cost outside of Vancouver, BC? And they're like, Oh, whatever. And I was like, okay, well, here's the answer. And I don't remember what it was like. 39,537 CAD was the average. And they were just sort of like, Huh, okay. And then boom, just moved on to the next topic. That's the problem. Like, people don't ask questions and then they don't even try to get answers. And and so you can't address the problems in intellectual diversity and like thought diversity until we get those things kind of anchored in the educational system.
Speaker4:
Well, obviously, we could continue this conversation probably for a long time, but unfortunately, we're out of time for today. So we just want to thank you so much for joining us today. As always, we'll put various links from this conversation and perhaps more resources on this topic in our show notes. But for now, thank you so much for joining us today. Yeah, thank.
Speaker3:
You so much for having me.
Speaker2:
We want to thank Deep again for joining us today for this wonderful conversation. And as always, it's time for our Post-interview debrief session. So today we're starting with Dylan and Dylan. Why don't you just tell us what is top of mind after this interview?
Speaker1:
Well, Jess, what is top of mind to me is I'm thinking a lot about his.
Speaker4:
Example of using chat GPT for as basically his his therapist, because as he said, you.
Speaker1:
Know, Jungian therapists are becoming more and more niche as the the new century rolls on. And it's really interesting to me that he's using GPT chat, GPT to supplant.
Speaker4:
A previous service.
Speaker1:
And it really did.
Speaker4:
Make me wonder more about what he was talking about with the threat to diversity of thought. So if there can be a good enough impersonation.
Speaker1:
Even if it never gets to.
Speaker4:
The level of nuance, although it could, I guess as an in-person psychologist or psychiatrist, the fact that we can impersonate.
Speaker1:
A Jungian therapist.
Speaker4:
Something that can.
Speaker1:
Be so niche.
Speaker4:
So well and so with so much reproducibility.
Speaker1:
It's kind of shocking and.
Speaker4:
Striking. And that case study made me think a lot more about just how. Every day. This could become just how ubiquitous this could become in our everyday life, which.
Speaker1:
I share some of deep.
Speaker4:
Excitement and hope for how this can be used.
Speaker1:
And how this can transform technology.
Speaker4:
In the same way that the Internet.
Speaker1:
Improved connection and technology and communication. But I also still am sitting.
Speaker4:
With that dystopic.
Speaker1:
Fear, which maybe I'm just old.
Speaker4:
Fashioned, I don't know.
Speaker1:
But but yeah, I think those are some of the emotions. And I was really taken with that young Ian case study. What about you? Mhm.
Speaker2:
Yeah. I think that dystopic is a good way to describe it. I agree with that. Um, I wouldn't call chatgpt utopic, at least not by my like, visceral reaction. I would call it like radical, which is kind of cool. Like along the theme of the show, it is the name of the show. It is something we like to talk about on the show, coincidentally. Um, but I actually did write radical in my notes at one point when Deep was talking about education specifically because he was talking about this sort of like radical reframing of the concept of education as a result of this new technology. And I've been thinking about that a lot recently because some of the professors in our department have been discussing how to prevent students from plagiarizing with Chatgpt. And I had some like, interesting, possibly dumb ideas that were to try to like, get students to maybe submit one answer that was prompted by Chatgpt or answered by Chatgpt and then one answer that was not using Chatgpt and they couldn't tell the professor which one was which. And then if the professor guessed or if the professor graded the one that was done by Chatgpt better than theirs, then they got a zero.
Speaker2:
So it's sort of like pitting the the student against the machine. Like, I agree, there's a lot of like innovative, weird ways that we could we could try to work within the new system that that we've found ourselves in now. But something else that came to my mind too, when Deep was talking about this like radical reframing of education is I agree that there's a lot of like positive benefits to be gained that students can gain in terms of their access to information and knowledge now that they have this ability to just ask questions to this chat bot and have those questions be answered pretty quickly and very succinctly in a way that Google hasn't quite gotten to yet. But then I was thinking if students become so reliant on this ability to get these answers so quickly and so efficiently, are they going to lose the ability to critically think and problem solve for themselves? And and that's something that I think is a little bit more dystopic to me.
Speaker6:
Yeah.
Speaker1:
I was thinking that. And then he brought up the example of the calendar, not the calendar, the calculator, but maybe the calendar would work here too.
Speaker4:
Of where it replaces some of this.
Speaker1:
Internal mental.
Speaker4:
Stuff that we've been taught for, you know, centuries.
Speaker1:
And then all of a sudden there's this thing that does it for us. And I remember.
Speaker4:
Growing up, I don't.
Speaker1:
Know if this was your experience just because maybe the calculator was already a thing, but I was not allowed to use a calculator.
Speaker2:
You're saying you grew up with an abacus?
Speaker6:
I did, right?
Speaker1:
Yeah, I.
Speaker4:
I don't want to get into.
Speaker1:
That, but it's a sore spot for me. But yeah, no, I wasn't allowed to use a calculator because.
Speaker4:
Especially in.
Speaker1:
Algebra, because they wanted me.
Speaker4:
To know how to do everything, including like long division within my algebra. And now I don't think that would even be like a second thought. You would just go straight into use your calendar or your God.
Speaker1:
I really can't get those two.
Speaker6:
Words. They're not the same thing.
Speaker1:
They're not the same thing. I'm the problem. Chatgpt would would know how to differentiate these two things. But but it it's like supplemental.
Speaker4:
To the learning. And I think that's one of the arguments that Deep's making here is that it's not replacing, but it.
Speaker1:
Like frees us up to now do different things.
Speaker4:
It's like a transformation for myself.
Speaker1:
I don't.
Speaker4:
Know how much that I totally buy that because I.
Speaker1:
Know that for myself. Like if I can just go Google something and I don't have to actually memorize it, I am sure that it's impacted the necessity for me to have a memory in some ways. And so I.
Speaker4:
Can jump over that. And so I do.
Speaker1:
I wonder about that for.
Speaker4:
Chatgpt in the.
Speaker6:
Future.
Speaker2:
Well, I've actually been thinking about this topic a lot lately and, and I agree that I think there are there are some benefits to be gained about, like, okay, maybe we don't need to know long division anymore as like a society, maybe that doesn't benefit us. Right. And maybe we don't need to actually have debates about like, what the answer to a question is anymore if we have the ability to Google it. But it doesn't mean that I, I, I don't miss those things. Right? Like, I kind of hate when I'm hanging out with friends and somebody asks a question and then instead of discussing it together, somebody just quickly pulls. Out their phone and then answers it on Google. And and then I begin to wonder, like, at what point does that begin to do a disservice to us? Like, I've thought about this a lot in terms of like phones and taking photos. And I think for a while I thought of taking photos as this thing that augmented my memory, like me taking photos and videos, helped me remember the things that I've done in my life so much more because now I can look back and be like, What was I doing on July 1st, 2017? Now I remember. But the more that I've begun to do it, the more that I've started to worry that I'm almost outsourcing my memory to this machine. And the act of me taking more photos and videos has actually made me have a worse memory overall in the long run because I just assume that I no longer need to remember without the help of this technology. So I wonder with Chatgpt, like at what point will this stop augmenting our ability to answer questions and to problem solve and potentially replace that ability?
Speaker1:
Yeah, I agree with all of that. One thing that it chatgpt is.
Speaker4:
Different than some of the other GPT three generative things in that it's a dialogue and I'm wondering how much that that dialogue dialogic I don't know if that's.
Speaker1:
Dialogue.
Speaker6:
I don't actually know how.
Speaker1:
The dialogue nature of of.
Speaker4:
Chat GPT of the chat the chat function. I wonder how much that changes how we relate to it because like Google, I wouldn't qualify as a chat. I would say, okay, we ask a question, we receive an answer, and like that's either a one off or you keep going. So maybe like symbolically it's but it's not like an actual dialogue where it feels like a conversation. And I wonder how that psychologically impacts it.
Speaker1:
Does it make us like think that it's more human.
Speaker4:
Like or.
Speaker1:
Ask different questions because we.
Speaker4:
Think that we're in a dialogue as opposed to just like a, a use value of this thing.
Speaker2:
Um, and I mean taking that even a step further now that if we do start treating it like a human and, and assuming that this thing is human qualities, especially in the use case that you started off with here, the therapist, a chatbot acting as a therapist, then do we assume that the chatbot now has the capacity to emote and to empathize and to do what trained therapists have spent years doing, where in reality it's just, you know, a bunch of inputs and outputs that that might bastardize a tradition that people have, you know, spent their entire livelihood trying to become an expert in. And and I think there's a lot to discuss there as well.
Speaker1:
And we will see. We obviously have a lot of.
Speaker6:
Opinions.
Speaker1:
About this. We weren't kidding at the end of the episode. We could talk about.
Speaker4:
This for a long, long time.
Speaker1:
But for now, for more information on today's show, please visit the episode page at Radical Eyeborg.
Speaker2:
If you enjoyed this episode, we invite you to subscribe, rate and review the show on iTunes or your favorite Podcatcher. And if you haven't tried out Chatgpt yet, we encourage you to give it a go. You can go to chatgpt openai.com and see if you're one of the lucky ones that can get through the queue because it is high traffic right now.
Speaker1:
On the plus side, if you try to get in and you can't, sometimes it.
Speaker4:
Gives you a nice.
Speaker1:
Little chat. Gpt generated acrostic poem about how you can't get in and so that's kind of charming.
Speaker6:
God, the robots are taking over, right?
Speaker2:
You can catch our regularly scheduled episodes the last Wednesday of every month, and you can join our community on Twitter at radical iPod. And as always, stay radical.
Sonix is the world’s most advanced automated transcription, translation, and subtitling platform. Fast, accurate, and affordable.
Automatically convert your mp3 files to text (txt file), Microsoft Word (docx file), and SubRip Subtitle (srt file) in minutes.
Sonix has many features that you'd love including upload many different filetypes, collaboration tools, powerful integrations and APIs, share transcripts, and easily transcribe your Zoom meetings. Try Sonix for free today.