A panel discussion about some of the key questions around AI with Michelle Robbins (LinkedIn), Leah Tharin (Jua.ai), Randeep Sidhu (experienced healthtech CPO), and Fergal Reid (Intercom).
← All talks
Lead Turing Fest 2023
AI panel - Where Are We Taking AI? And Where Is It Taking Us?
Michelle Robbins , Fergal Reid , Leah Tharin , Randeep Sidhu
Auto-generated transcript - may contain errors. Tap a timestamp to jump the video.
Good afternoon, everyone. I'm Michelle Robbins. I lead a decision science team at LinkedIn and super happy to be here with everyone and talk about this, exciting topic, see how spicy it gets. I'll pass it over to Lea. If everyone could just, you know, briefly introduce yourself and kind of what's top of mind for you at your companies right now and what you're doing with AI.
I'll start if that's okay. My name is Lea Dahn. I am the head of product at Chua dot ai, and what we're trying to do is we're trying to simulate the Earth's atmosphere with machine learning and then derive weather predictions from it, which is a fundamental change from how we used to predict the weather in the first place up so far, in history.
The interesting bit about what we do is that we can actually check objectively whether our predictions came true because we have sensor data where we can cross check on what we did. Actually contained errors or, you know, whether what we said is actually true.
And the thing that is top of mind for us is any technology jump that we have, just such as this one, needs some kind of regulation as well. Because whether you believe it or not, even weather data can be used for really insidious things.
Not to scare anyone. I was just going like, yeah. Yeah. But, yeah, this is on the top of mind that we definitely have, you know, regulation, what you do with the data because it is a powerful thing. Thank you. My name is Randeep Sibiu.
I spoke previous on stage three. I was previously the co, the head of product for the NHS COVID nineteen app where we used with the Alan Turing Institute AI to predict where COVID outbreaks were going to occur up to a week before they happened, and I'm a chief product officer working in health tech and AI.
Recently, I've just left a role building health care in Nigeria. So in terms of what's front of mind and what's worrying for me or not, I think working previously places like Babylon, building AI and other systems, it's usually AIs and things like that, proof of concepts that are built.
And whenever they're built, they're always built for the middle middle, for the standard able-bodied cis white forty something year old. And you build it as a proof of concept to see if it's gonna work. Hoziah works. All the investors pile money in, and then you just scale it.
But you forget that that thing was never meant for everyone. It was meant for that group. And then what happens is in health care, you compound lots of systems of different AIs together, like a chatbot that diagnoses your symptoms, a voice recognition bot to kind of speak to that doesn't understand non English accents, to and and and.
And you just compound all these kind of inequalities into, you know, the middle middle ends up being five percent of people at the other end who actually can be served by this system of different things connected. So for me, it's the system itself having issues but then compounding all those inequalities by having other systems put together.
That's probably what keeps me awake at night. Hi, everyone. I'm I'm Fergal. I work for a company called Intercom. I guess, by comparison, our our concerns seem prosaic. On a day to day basis, we're we're building customer support software. So we're trying to build bots to, like, answer people's questions constrained to a knowledge base.
And, our concerns mostly are around, like, hey. Taking these models and putting them in production and making them kinda, like, fast enough and cheap enough and stuff to use. I think there's lots of great stuff to get into outside that, but that that's where we are on a day to day basis at work.
Yeah. So it sounds like there are issues around, risk as well as model bias. How are you how are you approaching, mitigating for those risks and mitigating for the biases? I think from from our side, as I said, I mean, we have the really comfortable position that we can check the bias in the model.
So you usually have three different biases in any model that you're running. First of all, you're introducing some bias through the data. Anything that a machine learning model does, so, like, if you're using CHETCHEPT, all the data that has been fed is the only reality that it can know.
So any system has only access to the real word as you've given it to him. And that means if you have a bias when we're talking about the error in relation to what reality is is what's in that data as well. So if you give it data and you tell them that a specific group of people is successful, then the machine learning model will believe that this is true.
So you have a bias that comes from the model itself, like, from the data itself, then you have a bias that is introduced in the model. So, like, in the processing of this particular data, there can things that can go wrong. And then usually, if you're downscaling data or, like, if you're interpolating anything about this, like, if you're translating into other things, so you also have an output bias.
And all of this together is the kind of error that you're introducing that you see when ChatGPT is, for instance, starting to hallucinate, starts to invent things, and so forth. But the fundamental question is is always, how much bias do you accept as an end user?
And we have proven that we accept a lot of bias for mere convenience because ChatGPT is not known for being exact exact, but it's fast, and it's very, very convenient. So that's an interesting thing, I think, in regards to bias that we need to be aware of.
But then also it's to do with the perception of that bias because as you said, chat GPT, that kind of uncanny valley craziness where someone you read something and you go, it reads like a human, therefore it must be accurate, like lawyers who are submitting chat GPT as kind of petitions with invented cases.
So now people are more aware of how people are perceiving AI. But then if you're doing health care, there's some situations where patients or doctors are more trusting of something if it's come from a robot or an AI. And there's some situations people are much less trusting.
And it's sort of back to the kind of first conversation we've put on how you frame that in terms of how you frame what that information is and what output is. Because people are looking at the same information, do radically different things depending on what they perceive the AI to be doing and they what they perceive how accurate it is to be.
So I think the human perspective, I'm not technical, is what I care more more about because we can make the model as unbiased as we want. But then how do users perceive what they're getting, which is kind of a concern which often doesn't get spoken about.
I I I mean, I guess, like, I I would say is that, you know, we're all still learning how to use these large language models and to use modern AI. And I I think if you look at the tech we have just now, you know, it's possible to use these things in a way that mitigate a lot of risk.
So, like, you know, we've built a chatbot. Our chatbot will just answer questions from a knowledge base. So we're not using the AI to, like, generate sort of any sort of, like, open domain context itself. We're just using it as a building block, you know, to do better information retrieval.
So, like, if you do things like that, like, I I guess my my message would be right now, today, there's a lot of things you can do when you build a product. If you use these things in a creative way, you can end run around a lot of issues.
I'd almost say that, like, a deployment like JWT, which we're all familiar with, is is is a particularly tricky area because it's just it's just so open domain. So I think we're all learning how to use these things more and to put them in ways that they they cause problems less, in practice.
Can I can I maybe expand on this really quick? So I think what's very, very interesting in regards to ChatGPT and what it tries to do is that we're inherently incentivized to think that AI models or machine learning models, these language models that are so prominent right now are trying to replace what we do as humans.
Yeah? So you definitely have to also explain to people that, oh, you're replacing the support people. Right? You're taking away their work. That's not really what's happening here, at least not completely. Yes. We're becoming more efficient, and we're gonna get rid of some jobs.
It's true. Language models can do some work from marketing. Language models can do some stuff from from customer support, but it is not about imitating people. Some of the stuff is genuinely better, like information retrieval, especially what you said. It's not just, like, to plug in a chatbot where you had a human talking.
It is about a different way of navigating information and finding what you were having trouble with. And it's not just about giving an interface that you can talk to. We will have systems that predict that you are in a problem before you even talk to it.
That's the really exciting thing. That's the kind of data that we have because we can objectively maybe not check whether an answer is correct, but we can definitely objectively check whether someone has a problem. And these kind of binary outputs are very, very strong for machine learning where you can say, like, based on this group in here, fifty people had a specific problem.
So there's a correlation to maybe the food from one food stand that we had. Right? I'm not saying this happened. I'm sorry. I'm just saying, like, this is crazy. This is what's going on, so we don't know exactly how we're gonna use them, but do not think that machine learning models are just here to imitate what people do.
They make us do specific things better that sometimes have been done by people. And I think that's that's fundamentally what this is about. Maybe I just come in on the jobs thing. So, like, you know, we work with building automation for customer support.
And so, like, yeah, one thing we think about a lot is, like, does that reduce jobs or replace jobs? And I think what's interesting is it's kind of an open question. Like, anytime you make people more efficient, you run the risk of, like, replacing jobs.
But it can be complicated. And an example that my team mentions a lot is, like, when ATMs came along, everyone thought they're gonna, like, really replace bank tellers. But certainly in the US where you have stats, like, the amount of bank tellers really increases, ATMs got deployed, and they went up the value chain.
And so, you know, it's hard to say. I think we should be honest that, like, AI probably will, you know, replace some jobs that are done today. But maybe the folks doing those jobs will do more high value jobs. Maybe they won't. I don't think we know.
But I I don't think we know for sure on either side. So back to when you sorry. Back to when you were talking about what AI can do, the thing that we're all speaking from our positions of what how it impacts us and our roles.
So I did some work in Nigeria, and it's very interesting. Because being out in Nigeria, it's kind of a hustle economy. A lot of people are trying and hustling. And I worked with colleagues who had a side hustle that started using ChatGPT to build websites.
They didn't have coding backgrounds. They used the tool to build a new business. This is insane. The barrier to entry for someone to be able to build a website. And of course, in the future, more than just a website, to use code to I thought, you've got an idea.
You can now execute it using words and then see if it's gonna work. And for me, that's sometimes quite interesting about how it democratizes a lot of access in that way that people and so we don't know the impact it's gonna have in, say, weather predictions and the impact that that could be perverse.
But, actually, there's quite a lot of excitement about how people could do things that weren't possible before because of barriers to entry. So that's sort of interesting as well, and that's happening today. Can I say one more thing? And we get to say as many things as you'd like.
I don't think the clock agrees. But okay. So the most fascinating thing on why this is now becoming a thing, because we used to have machine learning and AI in our lives all the time, for movement models, for robotics, for all kinds of things.
We have this all the time. The main difference that we have today is that language models can interact with each like, with other systems that have been trained on machine learning as well. So when you're talking about coding a website, it's not that ChatGPT is a tool to code websites.
You can upload a documentation from an API. I did this with our API to give it to ChatGPT and say, like, okay. Now pretend to be something else, a specific system, and create a call in our API to deliver something out there. That's the fascinating thing.
If you wanna have a stupid statement that you can tell your grandparents, then you can say the machines have started to talk into other machines. Everything is starting to be really interconnected. And this is what's so fascinating about it. And I agree with you one hundred percent.
We don't know yet what the end of this will be, but it is absolutely fascinating that I can upload something in English, structure it in a specific way that I want, and have an output in some other language. So we created the intimate Great.
Have we created the API that connects all the different models? That is exactly what it is. English. That is exactly what it is. In the nineties, we started to bring APIs on how to talk with systems, and you have to sit down and define a standard language.
It's almost like you speak English, you speak French, and here's a dictionary now to talk with each other. And language models are extremely good in taking unstructured information from some kind of place and translate it into another medium, and we have not seen the beginning of it.
But this is exciting. It's really exciting. Yeah. I think whenever there's a significant change I mean, this happened. I mean, look at any any point in history, you know, the industrial revolution, cars, you know, everyone, kind of jumps to, it's gonna replace people, it's gonna replace jobs, what are we gonna do, what are we gonna do.
And I think the transition period that we're in right now is is we're just experiencing that. And I think that the ability to make, to make jobs easier, to make things faster, you know, jobs that you do, you can do faster now. And that's that's actually a good thing because if you can leverage a machine to do, you know, essentially the math of all your job, then why wouldn't you?
Right? Because then that leaves you time to focus on things that a machine's not gonna do. A machine's not gonna develop a strategy for your company. A machine's not gonna develop, you know, not gonna tell you a machine can make your website, but it's not gonna make you profitable.
So I think that the concern about what are people gonna do, is is we're just in that weird transition time. So do you disagree? Do you think the machine is gonna make the call? Say anything when you you kinda see the first. Well, okay. I mean yeah.
I I mean, I I guess I'll go. Yeah. I I think that that that that feels true to me at the moment. Obviously, things have been developing extremely fast. And, you know, like, GPT four is a very powerful model. Like, can it not do strategy?
Probably not, but it's getting close, I would say. But if everyone has the same strategy, what's the point of differentiation for any any business? So that's the thing. Right? It never outputs twice the same stuff. So you always have to go in context and like, okay.
So what if you did not have ChatGPT? Let's say you've never attended TuringFest. Shame on you. Well, no one in here in this room is applying this. And you don't know how to create any strategy. What language moms can do is not only, like, teach you how to do strategy from a frameworks perspective because a lot of people have been talking about strategy.
There's some commonalities. But for correlation analysis, if you have some kind of dataset and we're talking about language models that are trained on internal datasets. Right? So, like, you take everything that is in intercom. We have a language model to interact with it.
It's absolutely possible that a machine learning model can predictive on how other companies work to create some kind of relations. So let me take the stupidest example that I can think of. A company that has a high revenue is probably more successful than one with no revenue.
That's the first connection. Right? And then as you feed more data into the system, like, is there a connection between employees? Is there a connection between where a company is located? The amount of taxes that it pays? These are all things that an machine learning model can actually analyze.
And is it will it find obvious mistakes in the things that we do? Absolutely. I think so. Whether it can do better strategy always, I'm not sure. But, yeah, I think it's gonna happen. But then there's things like if you look at if I think about health care now, if you look at what was the biggest impact of AI or what was now, we can start thinking about large language models in COVID.
People like so I the Turing you know, Alan Turing Institute, we kind of say, we could predict outbreaks a week before it happened. People are, oh, that must have been the best use. No. The biggest impact of things like what we'd call algorithmic prediction was in Facebook advertising and YouTube videos doing COVID denials because I got hundreds of emails from people telling me I was implanting four g masks into their brains or something.
Like, this is did you? You can't say it publicly. Call a master. I like that. That guy. But the reality is it's that kind of stuff where, like, large language models and those predictive analytics means that you can actually amazingly help give information to people that they're gonna use to activate on their health care, but also provide massive amounts of misinformation.
Like, using things like these, you know, these map models that can predict the things that will help motivate you to do something is terrifying. Sure. Stupidity is scaled as easily as intelligence. I'm going to steal that. Mic drop. And in terms of some of your predictions, what you're talking about, that correlation, that's the end of insurance.
If I could put all my health care data into something, it would know my risk of getting sick. What's the point of insurance? People are not gonna be using that anymore because they'll have in a prediction of when I'm gonna get sick or not.
Like I I I I would say I don't think they're quite at that level of, like, you know, high level systemic synthesis yet. You know? Like, I I think they're they're really good at, like, you know, acting intelligently given, you know, a couple of pages of, you know, even English language description of a scenario or a situation.
But I don't think they're gonna, like, look at a a whole system and do synthesis at that system level yet. I don't know how long away. But Well, if you put all my doctor's notes into a large language model, could it predict what might be happening just because of patterns of behavior?
I'm I I I guess maybe, but that might be quite a simple prediction. You know, that might like, if if it was something like subtle and edge, Casey, I I would say you're you're on, like, the cutting edge. And, like, Google doing really interesting stuff for, like, MedPalm and so on at the moment, like, really large language models, massive language models trained just on on medical things.
So, you know, that particular domain is being studied at the moment. I I would say, like, general purpose foundation models like g p t four, like, that that that's probably that's probably stretchy thing. Not now. From from a program point of view, if you've got, like, a single class or module to debug, they might be quite helpful at that.
But if you have, like, a system to try and figure out implications of, that that's kinda at the edge of a single prompt. So this is not about binaries things where you just say, like, yes, no, true, false. Right? This is about a likelihood.
Even for a machine learning model, a machine learning model has a specific likelihood of being right, and then objective truth is another thing. But let me just ask a question into the audience right now. If you start to wake up in the morning and your knee hurts and you want to go to a doctor, raise your hand whether it takes one day or longer before you can even talk to the doctor.
That's a lot of hands, and it's just assuming that people just raise, like, one hand instead of both. So the thing is it's not just about prediction quality or, like, accuracy of what you're doing, specifically also when we talk about support. If I am worried for my health, something is happening, the amount of reaction time, the time to value for this, talked about this as well today in my talk, is incredibly valuable in itself, even if it would be less accurate than what the doctor does.
And then the other thing is we have a big problem in society, specifically from the older generation that they're afraid of talking to humans about specific things because they're embarrassed about them. If they have some problems, you know, like with their bodies, like, becoming frail or whatever, why not talk to a machine for that particular thing?
So there is more than just, like, how likely that you're going to predict something. And that's probably the case where I'm actually more actually more excited in terms of the speed and and then just, you know, binary, like, is it true or wrong?
We've actually already seen that. So Babylon, another health tech company, so it's building chatbots for the last nine to ten years. You see embarrassing conditions, conditions, sexual health, rashes, things like that already over index in and when I was working in Rwanda where, you know, we had chatbots and doing work there, LGBT issues became higher because they were more sensitive locally.
So it is a way, and we actually tested using a human like person or a deliberate robot. People wanted to speak to a robot because they knew it wouldn't judge them. So they're more open about stuff back to the perception. But there's definitely still challenges there in terms of how people are using it.
That's the reality. Yeah. I mean, I I would echo all of that. Like, I definitely think there's these are different shapes technologies. They're different shaped solutions to problems. There'll be some things they're worse at, some things they're better at. Yeah. There's a lot of benefit to AI systems.
We build support chatbots. If, you know, you're a simple informational question about an entertainment thing you're working on, You can be quite tolerant to some amount of error, get your answer quickly. It's really good. So, yeah, there's a lot of asymmetric advantages to AI system.
So what are you most what are you most worried about for, for AI or these large language models? AI generally Large language models. Let's be specific. Yeah. So I I mean, like, I I think that, the medium to long term risks of, like, you know, people are talking about, like, AI safety risks.
Oh, I take those seriously. I think those are serious things. They're not what we have today. They're not the class of models we're using today, but, there's a debate around that. And, yeah, that that definitely is something I've been thinking about recently, yeah, about everybody else.
So it seems like there's, recently been a couple of, fields of thought on what we should do in the immediate and then what should happen going forward. So in the immediate, you've got the everyone should take a pause for six months. I'm not sure what everyone's gonna do in that six month time and how that's gonna You you not know what I'm talking about?
You mean holidays. Right? I wish. I don't do Everyone should take six months holidays. Go have fun. No. I I Sam Altman in that group were, you know, saying, we should pause for six months. This is getting out of hand. No no real indication of what that magic six months does for anything in particular.
But, then you have, you know, the the other side of the coin. Marc Andreessen recently did an op ed saying, you know, AI is gonna save the world. Like, so you've got the people, AI is gonna, you know, end the world. AI is gonna save the world.
Where do you fall on that spectrum, and what do you think of those two different, you know, arguments? I think this is the biggest jump and opportunity we have had in the last, I don't know, let's just say a high number, ninety five years, specifically for medicine.
What this will do for medicine is absolutely mind blowing. I don't know where the bad things are going. I really don't know. But I find it very funny when someone is actually suggesting knowing how the markets today work and how countries work, that in six months, you can achieve anything on a legislative level.
Give me one country in this world that has moved anything in in six months, unless it was an actual pandemic that was threatening everybody physically. It's just not realistic. And the other thing is this is a classical thing of if the cat is out of the bag, it's out of the bag.
Specifically with OpenAI and, or, like, open source communities, the stuff that we can do just from what is available, you're not going to contain this anymore. Specifically, I'm not sure whether people are aware of this, but, like, in terms of how fast that machine learning, the technology behind, how you do model training and neural networks, how much this has changed.
I don't know whether you read research papers. But if you do, the stuff that you read two months ago is probably as old as the Matrix movies, the original ones. That's how it feels. It was amazing, but it happened twenty years ago. It's so fast how we came from, like, another breakthrough, another breakthrough on how to correct bias, on how to train models faster, on how to store all this data.
It is absolutely mind blowing in how fast that all of this is actually norm starting to go. And the new problem that we start to have is is also with these models, like, for instance, generative models, like mid journey and all these models that create new pictures and so forth.
They start to train on their own data. Before we had these big models starting to just train on stuff, all the pictures were real. This is a feedback loop right now. Right? We've never dealt with something like this. And we're not marking our content as AI generated.
I don't know whether we have any markers to determine this. I don't know. It's another problem. Right? Yeah. I was I was gonna ask about that, actually. Well, how do you feel about disclosure? Do you feel like if, AI if a product, if a service, if if a chatbot, if anything is leveraging one of these large language models, do you think there should be disclosure about the fact that it is and then the model that it is trained on?
Well, I mean, if you look at other areas, if you're getting g GM crops, you know, you're meant to be told at least in the UK if you're using it. And I wonder whether knowing what you're using can help keep people more honest when they're using it and how they're implementing it.
Because if you don't know where you're looking, how do you know to even look for it? How do you know to be reassured when you're hearing it? The good news about health care is at least there's some level of caution from doctors and using stuff.
I don't know whether that's the same I mean, some. Not sure it's I know you're begging, waiting to give your your death knell of what's gonna happen. But, yeah, I do think that there's some challenges. I mean, I I guess on that last topic specifically, you know, we built a product.
We put a little label in it. It's an AI whenever it generated an AI answer. I think there's there's ways of designing affordances around these things to to manage user expectations, yeah, in the short term. Yeah. I don't know. I think look. If you don't get paid or you don't get punished for something, which is a fundamental incentive that we have in our society to do anything, then it's just a paper tiger.
Should we do it? Yeah. Sure. Will we do it? No. Absolutely not. Nobody will. Because fundamentally, like I mean, if you have children, you know that they started to use ChatGPT. They are smarter than all of us in the room together without ChatCPT than this one child that is now starting to learn to use language models.
School was not ready for this at all. How are you gonna do this? Right? Like, when I give courses, I have to tell my students and also tell them, like, like, hey. Please don't use chat GPT to hand in your assignments. I don't wanna grade chat GPT.
Like, just don't hand in anything. The legislation's up for it. I mean, like, Yeezy said, like, the reason this building is not gonna collapse on us is because someone's gonna get sent to prison if it happens. Right. If your chat GPT or your AI did something, who's gonna be sent to prison?
Who's gonna get prosecuted? It's so diffuse. You just kinda pass it up and down the value chain. The reality is no one's gonna get held to account for it, so no one is being held to account for it. So there's something there about I mean, we're not gonna do it.
On all these language models, the thing that terrifies me is, like, look at the next pandemic. Look at the stuff. Anything that's based in text and some things that aren't can be trained and learned. DNA. Probably, like, the things about medical records. Yeah.
Maybe not my unredacted medical records, but there's loads of systems that pass out and give the basic conditions of my BP and stuff like that. That's well understood. That stuff is shown to start producing predictive outcomes of your likelihood of various things. Like, we've got an information problem where the reality is there's so much information, say, in my medical record.
No one can honestly read through every line of it to give me a sense of information on it. A doctor won't really be able to do that. If that could happen instantly, however accurate or not it is, that's gonna impact how I'm perceived in my risk, my ability to do stuff.
That's what we're getting. All that information being connected, as you said, the information being freely shared on chat g p t, that's the stuff that people aren't realizing that's out there now, and that's gonna start being used probably against them at certain certain points.
Not necessarily maliciously, but I don't I I mean, just to come come in on the on the bigger issue, like, you know, I I I do think that there has been a lot of good faith, and to me, it seems like good faith discussion from, like, major players in this area.
And I I really like, you know, when you start talking about, like, major, major AI risk, I think there's a lot of alignment of incentives. Nobody wants AI to, like, you know, do something really, really bad. I think you're seeing that. I think you're seeing a lot of, you know, leaders of, like, major orgs and major companies, you know, actually, very unusually and very weirdly start a conversation about regulation.
Like, when's the last time you've seen that sort of thing happening? I I know there's regulatory capture, but I think even beyond that, I I think I think there's a lot of good fake discussion about this. So I I I don't know. I'm not I'm not so cynical that I think, like, hey.
People will, like, do something really, really terrible, just for the money even if it negatively affects them and everyone Tragedy of the commons. What's happening in the pollution? Why would you think that people are gonna be any different with AI than they are with pollution and kind of infecting the environment?
I I I I I think because if you actually look at the discourse, what people are talking about, it it it's arguably much more clear and present, than even pollution, I would say. Yeah. Now if you take it seriously, not that's up to you.
But, but that's certainly in the discourse. Yeah. I I don't know. I think, ultimately, these systems, these models are a reflection of us. We put everything into them. It captures everything we're putting out into the world so that we would expect them to behave any differently than we behave, I think, is, false.
I think that we should understand that, the biases that we take to it, it takes to us. It takes back to us. It reflects back. So I do think that there I think there are some legitimate concerns. And I agree with you. The incentive structures, it's all about the incentives.
I don't know how you get to a place where you get countries globally, you know, to agree on how should we manage this. What how how can we anticipate the problems we can't anticipate? And then how would we manage them? How do we build in the safeguards?
I I mean, I I would kinda come back and then say, I I think there's there's major things to to be thoughtful about and be concerned about here. But actually, it's not because they'll reflect our biases. I think, you know, everyone's trying to everyone's worried about alignment.
Like, how do we get them to reflect our values is actually a harder problem and are you really more risky? Yeah. I agree. I'm still waiting for you to give me my tariff. Like, come on. Tell me how we're gonna use, how where the system's gonna be used maliciously.
That's what I'm waiting for it to drop. No. No. No. That's not that interesting. Let's go into the future and say chat GPT version number fifteen comes out, and it is extremely good. It has no bias. We all trust it. We trust it so much that whatever we put into, you know, like, the example that you had, it's just going to be really, really good.
They tend to be on your side. If such a system is trusted intimately by us to a to such a degree that when that particular system, this system that we trust tells us, if you do not take care of climate change, of this and that, then this is going to happen, then we probably solved a big portion of the problems that we have.
Because a lot of the problems that we have is is that we don't trust certain voices that we maybe should trust. But then if you look at what's happening with things like YouTube, Facebook, all these algorithms exist, they could give more honest information.
They don't because they're driving attention and polarization. So I'm not sure why you believe that this system existing was gonna be honest and truthful. Fox News exists. That's yeah. It's like Yeah. Not gonna go there. Exactly. Yeah. Well, we have, we're out of time.
Do we does anyone have any last thoughts? Any last things you wanna leave the audience thinking about? I feel like we left on a bummer. We're all gonna die. No. We're not. I honestly I I am very optimistic. I'm also a huge Battlestar Galactica fan, the new the new one, not the old one.
And, so, you know, they have a plan on the plan. The silans don't have a plan. That's the sad thing. They don't have a plan. It's up to us. Right? So, I do think that there's a tremendous amount of opportunity in the field.
I think there's a tremendous amount of things for people. Even you were mentioning about school. I think that every child on the planet can now have an on demand tutor. The way that's gonna level up educational opportunities, I think is amazing. I think that's fantastic in and of itself.
If for no other reason, ChatGPT exists, think about the educational opportunities for people in underserved communities. Amazing. But I do think there's there's a lot of risk, and I think we need to be thoughtful about what we're building, how we're building it, the kind of feedback loops that you called out.
I I just think a lot a lot more mindfulness needs to it needs to be approached with mindfulness more so than, like, this is amazing. Let's do all the things. Do you wanna give closing remarks? Because we're already overtime. Yeah. I I I mean I'm just leaving.
You guys hang out for a while. Right, Ryan? It's a big issue. It's a complex issue. Yeah. I mean, there's there's loads of calls for optimism in the kinda medium term. We'll be able to build a lot of really cool things to make people's lives better with this sort of tech.
And I I think there is sort of a a medium to long term discussion around AI safety. I think it's worth taking seriously, but it's a very complex discussion and it's evolving at the moment. I'm probably I see. I'm I didn't realize I was so cynical.
I think I'm pretty much cynical person here in the end. But, we started as me being really optimistic, funnily enough. But I think I just look at history and think history is gonna repeat itself, and we just have entrenched power using this to further entrench power.
That's what happens. That's what systems do. So until we actually look at the system that we're creating and think carefully about it, I don't have faith that AI is gonna make this rosy future as much as it helps people access education. I started my career in education.
I think that's important. It helps level up lots of people, but that thing, I don't currently have faith that's gonna help the people it most needs to help because I think it will be made for profit, not made for help. That's okay. You don't need faith.
You need a language model that is really good. Okay. So here's what's going to happen. Here's what's going to happen. All of us in this room, our capabilities that we have nowadays, some of you are better, some of you are a little bit worse in certain areas, All of this will be leveled out.
It will level the playing field completely again. We all have to learn how to use language models. We are the last stupid generation, if you wanna call it, because our kids are going to be so much smarter than us. And I think this is a great chance to be really exceptional with a new set of tools that we just did not learn yet how to use.
Awesome. Well, join me in thanking the panelists. Fergal, Sandeep, and Lea. Thank you so much.