Name: Brainstorm AI London 2024: How The Latest Breakthroughs In Generative AI Will Transform Business And Society
Uploaded: 2024-04-15
Duration: 19 min 2 s
Description: Zoubin Ghahramani, Vice President, DeepMind, Google; Professor, Information Engineering, University of Cambridge Dr. Anne Phelan, Chief Scientific Officer, BenevolentAI Dr. Marc Warner, Co-Founder and CEO, Faculty Moderator: Jeremy Kahn, AI Editor, FORTUNE; Co-chair, Fortune Brainstorm AI London

Zoubin Ghahramani, Vice President, DeepMind, Google; Professor, Information Engineering, University of Cambridge Dr. Anne Phelan, Chief Scientific Officer, BenevolentAI Dr. Marc Warner, Co-Founder and CEO, Faculty Moderator: Jeremy Kahn, AI Editor, FORTUNE; Co-chair, Fortune Brainstorm AI London 

Brainstorm AI London 2024: How The Latest Breakthroughs In Generative AI Will Transform Business And Society

Photography by Joe Maher Fortune Brainstorm AI

Conferences

Brainstorm AI: How The Latest Breakthroughs In Generative Ai Will Transform Business And Society

Zoubin Ghahramani, DeepMind, Google; Information Engineering, University of Cambridge, Dr. Anne Phelan, BenevolentAI, Dr. Marc Warner,  Faculty, Jeremy Kahn, FORTUNE

Thank you all for being with me and thank you all for joining us. So general A I is creating a lot of buzz. A lot of excitement about what is coming in the future from these models. I want to start by talking a little bit with been who's Amelia to write about Gemini. This is a very powerful model that Google has, has developed and has just come out with a couple of new versions of what is so special about Gemini Zubin. And how is this potentially transformative? Let's start there. Yeah, thanks Jeremy. It's a pleasure to be here. So, um first of all, Gemini is, is a whole family of models going from on device models to our very largest models. And uh what's really exciting about Gemini is that they're built multi modally from the ground up so they can process uh images, audio, video, et cetera alongside text. And our latest version of Gemini um actually has uh what's called a, a 1 million token context window, which means that it can process and keep in memory a tremendous amount of information much larger than any of the other models out there. It's like a whole book's worth of material. It's many books, a whole podcast, a video, an entire code base. And you know, if you think of that, not just if you think of that in terms of like the short term memory that the model has, that you can query and interact with the capabilities of that. We've only really scratched the surface of that. Interesting. Now, I know when Gemini came out at least one version of the model, there was an issue around the image generation capability of the model that in an effort to try to make sure that the images were you represent a diverse set of people and overcame some of the historical biases in the data that people complained about with these models. You guys created a model that it was actually very hard to get it to generate images of white people even in historically appropriate context for that. And it was quite embarrassing for you. How do you, how do you overcome issues like that? Because I think, I think that's one of, I think that's one of the problems now is that people want a model that you can just say, OK, well, you know, be diverse when that's appropriate and when it's not appropriate, you know, be correct to the historical context, but you can't just tell the model that apparently. So yeah, how do you overcome the first of all? Thanks Jeremy for asking, you know, it's, it's actually you know, it really highlights the fact that these challenges are difficult, right? First of all, you know, we want the tools, the generative A I tools to be relevant to people all around the world. So if you ask a model to create images of, let's say a nurse in a hospital or a school teacher or you know, a board member of a company, it has to be relevant to people all around the world. Google is a global company. We have, you know, people using our tools on every continent. Um but the guard rails that we put in uh in that case, we were clearly wrong. We came out, we actually said that we then um pulled the image generation for Gemini off so that we can improve it and put it back out in a way that's more acceptable to more people. Um You know, this is really part I think of the iteration process where you create tools, these are cutting edge tools, you put them out there, people find them really useful, sometimes, sometimes they find them offensive, right? And you know, I think this is a a healthy thing that the whole A I community is really grappling with right now. Mark I want to get you in here. You faculty help a lot of different businesses, try to figure out how to implement A I technology, help the government do this to some extent. I think, you know, this all comes down to reliability. You have these very powerful systems were very excited about the sort of headline capabilities, but then they start playing around with it and they're like, oh, but it's not reliable. How do I? And I think a lot of business leaders are struggling with the idea of how do I implement something that's not reliable? What advice are you giving people around that? Yeah. So we think that people need to think about these in a human centric way. So they are tools to be used in combination with people. And so if you're careful and you're thoughtful about how you design that in from the start, you give yourself this ability to sort of tune the workload between the person and the machine and that ultimately, well, one puts you in a much better place right now because the models aren't perfect and will have small failure modes that can mess up a particular workflow. But two, it does also future proof you to an extent because over time, you can tune that balance between the person and the machine so that as these developing capability and the truth is, nobody knows how fast they're going to develop and how far they're going to get. But as they do develop in capability, you will be able to tune that workflow as you go. And so you'll have this nice possibility of being relatively future proofed and benevolent. You're using these to help with drug discovery a lot of people are very excited about the potential transformational effect on science. Maybe you can talk a little bit about where you see this going and how big an impact you think these models will have on science. Yes, actually your comments resonated really clearly for me because we do think of it as human in the loop we have now the capacity to surface huge amounts of data by medical data. The field is expanding at a colossal rate, all mixed data, patient level data and to be able to carefully curate and amalgamate all that to be able to pose biological questions to our data foundations is something that we've given an awful lot of thought and some elements are moving really quickly in terms of automation and some are very heavily reliant on the human in the loop kind of process. But fundamentally drug discovery is very, very difficult, it's expensive, it's time consuming something like 95% of drugs don't make it through clinical development to launch and it costs something in the region of $2.5 billion per drug. So anything that we can do even incrementally to increase our probability of success in each incremental process step of the process and inevitably bring down costs is going to have a massive societal impact with the use of A I is carefully applied to drug discovery. And when you talk about carefully applying, I mean, can you talk a little bit about what's involved there and, and how do you prevent some of the potential failure modes? As Mark was saying of these models? Yeah. So I mean, I think incrementally, the first step of drug discovery is identifying the target that you want to modulate. So we spent a lot of time thinking about trying to integrate all biomedical data so that we can understand the signaling pathways and mechanisms of health and disease. So we can choose the right target because fundamentally, if you have the wrong target, doesn't matter how good you are, the rest of it, the whole thing's doomed really. So starting really carefully with the with the right target is a is a big first step. And actually, as we use some of these more generative models, really thinking carefully about the prompt and asking very careful questions of the technology so that you don't get these kind of hallucinatory answers that you can very carefully. It's not just one question, one answer. It's a series of very carefully crafted questions that get you a more and more refined answer back that then we can use, you know, actual data that we can use for the next step of the process. Interesting speaking of questions, we're going to go to questions from the audience in a minute. So please think of your questions. Um When you talk about using these tools for science, I know deep mind has done a lot of work around this alpha fold, you know, a huge success. Um recently, though there was a paper you guys had on material science that had been published in nature came up with, you know, thousands of new materials with unknown properties of science. It sounded very exciting. There's recently been some criticism though that a lot of those materials are not really synthes, sizeable, not easy to manufacture. Maybe we're getting ahead of ourselves. I mean, as someone who's been in this area for a long time, looking at the research potential of these models, how are you thinking about it? Are we, are we over hyping? What these can do for science? So I don't think we're over hyping because I think actually one of the most exciting things that A I can do for humanity is uh advancing our ability to create new technologies, new other technologies. So let's talk about material science. This is a very exciting area. Um uh So the paper in question um actually uh had uh 2.2 million new stable crystals that were automatically generated from the A I model. And if you think about what new materials can help you do, they can help you um design uh better batteries, they can help you um design better uh photovoltaic solar panels and all sorts of things that will help us address, you know, medical devices, all sorts of things that helps us address um you know, societal challenges. Now, um the uh the process that we stand by the results of the paper, it was peer reviewed in nature, which is really the top journal academically. Um, and, you know, it's part of a healthy scientific uh uh process where you put out some results and people come out and question them, you go back and forth and, you know, hopefully that the truth comes out. But the paper is really exciting along along with a lot of the other work that has been done in A I for the sciences in, you know, many areas like biology with Alpha fold or, you know, fusion energy and mark, do you think the hype sort of getting ahead of ourselves here in terms of what, what these models can accomplish? Uh Well, I mean, I, I like, I think that A I for science that deep mind does is unbelievable and totally brilliant. But I do think in other contexts it's easy to over claim for these models. So, like, you know, you look at when we were promised or fully autonomous cars and it's taken a bit longer than we expected. And I think if people are thinking in the context of like, these are just going to automate away jobs, one for one that will be like much further away than most people anticipate, hopefully much further away as well. I think it's both true and good that it would be further away. But none of that is to say anything other than like the science these guys are doing is totally amazing, right? Uh I wanna go to questions from the audience. So please, if you have, if you have a question, please raise your hand and please wait till mic gets to you and I will call you. There's a question here if we can get a mic to this uh lady here. Uh And if you could please state your name and, and your affiliation. Um, when you stand up. Hi, I'm Natalia Yashu. I lead product in a net tech called Learn Light. Um, and actually I just wanted to respond to the last comment you made around. Well, it takes us longer than we sometimes expect to progressive technologies partly is due to our acceptance or no acceptance of risk. Maybe we could have progressed far faster with autonomous cars if we're ready for more accidents, right? So it feels to me like it's the same with um A I, so I wonder how do you think we can actually find that balance between accepting some risk? Is there less a risk that we're ready to accept because it's technology and not a human error or actually do we need to? That's, that's a great question mark. I want you to answer that first. Like when you're advising people on that question, like how much risk should I accept? How do they, how do you reach that balance was obviously context dependent. So if somebody is saying we want an algorithm that will put a red jumper or a blue jumper top of our website. Then, you know, it's basically use whatever you want. It doesn't, it doesn't matter too much. Then in something like a medical context where the ultimate decision can in principle come down to life or death care much, much, much, much more greatly about um about that risk tolerance. Um At the moment, you know, I don't think there is anything better or more thorough than working with the organization themselves because it is important to remember that even though A I is a new technology, the thing you're using it for is still the thing you are always doing in almost all cases. So if you're a doctor, you are always making life and death decisions and now A I is coming into that. So whether that was like a bureaucratic decision or a technological decision or some simpler algorithm, in any case, you are used to wrestling with exactly the ethical dimensions of that problem. And so in the context of any given problem, there's almost always like a large amount of pre existing thought that you can tap into carefully and, and when you're dealing with a process that, you know, costs 2.5 billion and takes 10 years, how do you view that sort of risk assessment? And, and can you, you know, is it, oh, if we can save any time and money off? Of that. It's worth it or, you know, not always. Yeah, I think, I mean, there is a risk associated with it, but I think the risk can be quite a positive force for disruption. So when you think about drug discovery, as in designing the actual drug, we have very experienced and very seasoned medicinal chemists who can design, they're augmented now with all sorts of different predictive tools to enable them to make different molecules. And sometimes the tools will suggest something the chemist would never ever have thought of. And they inherently think well, it could be, that's not right, you know, that's too dangerous. But I should then, you know, you kind of think, well, actually maybe components of that could be useful. So I think it's the risk management if you, you've got to be open to the notion of risk and what it can do in terms of disruption, but there has to be like guard rails and controls around how you view that, that kind of prediction. So I think, yeah, the scope for this as long as it's carefully monitored, right? Other questions from folks in the room, please raise your hand. There's a question right here, please stand up and say, yeah, he l Southern Methodist University and the board of Directors of Hitachi. Um at lunch, we spoke about the importance of the triangle between on one hand, it backbone. On the other hand, data analytics, data science and domain knowledge you talked about the human in the loop to make sure we get the right quality. But often the person with the domain knowledge, it's the hardest to get into the loop. Can you, can you give some practical advice? How do we make sure that we actually get this human in the loop in the right way? That's a good question mark. Do you want to take that? Yeah. Well, we think of it as human in the loop and human over the loop. So initially, when you start building a system, you probably do want a human actually in the loop as in nothing goes out without human seeing it. But over time, as you build confidence, you can have a human in over the loop putting policies to guide it. Now, the problem is that you do actually have to build your systems in a slightly different way. Like if you want to make them properly human centric, it requires a different set of slightly different set of technologies, a slightly different set of design principles. So we've thought quite a lot about how you do that and have not not published it in an easily accessible form. So I, I promise I will write a blog and I will send it around. Uh but they're like in essence, the way you do it is you start by designing for human and machine to work together and then you structure the algorithms in a more modular fashion so that you basically inherently build in more explainability and more causality into the system. Fascinating other questions here. There's a gentleman here hand up uh marksman from the Times and how, how do you address the issue of concentration of power in A I given the resources needed to build models, etcetera and compute. How do we, how do you answer that one from Google Jeremy? Um So, yeah. No. Uh I, I think it's a, it's a very good question. Um When we look at the, uh what's happened in A I over the last few years, um it's actually really exciting to see uh how many start ups have come out with state of the art models. It seems like, you know, uh every week we have a new entrant into the field, the barriers to entry have been continuing to decrease with open models. We've also contributed to that. We, you know, we've published a lot of our work, but we also have produced open models and then if you look at the um cloud providers, um the whole model is to make these uh tools available very widely. Um So I think it's actually a pretty healthy ecosystem of competition right now. Um It's, you know, although it's expensive to train the very largest models, once you train it, once you can make it available and then people can build on it and use it. And we've seen an incredible thriving of the ecosystem So, um, you know, we'll have to see what happens over the next few years. But, um, I'm, I'm, you know, not concerned about that myself, but, you know, and that brings up an interesting question though, which we were talking about earlier about sort of, do you, do you build these models yourself? Do you buy them? Do you use open source models? How is benevolent sort of looking at that? And then also in the industry of pharmaceuticals, like there's this issue about power, all the big pharmaceutical companies have traditionally had a lot of data. How have you guys as a start up sort of thought, compete with that? Yes, in terms of the models, we have a hybrid model. Actually, we have some proprietary capabilities, we have open access capabilities, we have open access capabilities that we sort of customized and made them more bespoke to our needs. So, you know, in case of access to data, we've systematically brought in a whole range of different data types from the literature or makes compound structures, all this kind of totality of biomedical data so that we can reason over it. But we will always be have smaller data foundations and the really big pharma who have got decades of drug discovery data in house and that's for them to build their models and reason over. But for us, you know, there's a surprising amount of literature publicly available, well, literature and other data modalities that if you systematically build, you can get really decent picture of human biology and mark when you're talking about customers and clients, how do you advise them on this issue of bill versus buy? And do they use an open source model or do you use the proprietary model that seems to do better on a benchmark? What we, I mean, we think about A I is just better software, right? So it's the perennial build vs by question and the only two answers we know are wrong is all build or all buy. Anything in the middle is almost, is at least justifiable. And then our view is that it depends on how it compares to your capabilities and your competitive advantage. So if you have like if it's close to your competitive advantage, you want to be leaning to more towards, more towards build. If it's further from your competitive advantage, you want to be leaning more towards buy. Fantastic. That's all the time we have for today with this group of panelists. But thank you very much for being here and thank you all for listening and we'll get on to the next panel. Thank you very much.