Lila Ibrahim, COO, DeepMind 
Richard Socher, CEO, you.com 
Moderator: Verne Kopytoff, FORTUNE


Brainstorm Tech 2022: Delivering On A.I’s Promise

Conferences

it's good to see everybody. Thank you Layla and Richard for coming, appreciate uh one housekeeping note will be asking the audience to want some questions, questions at us towards the end of the session? And so please have some good ones ready? Uh for Layla and Richard uh to start off with um you know, since we're talking about a eyes, you know, fantastic promise, uh what is the milestone sort of the big sci fi achievement that would make you say eyes promises is finally realized that, wow, it's here, Richard, um I think AI and the definition of AI is actually continuously changing to what we are just not quite able to do yet. So as soon as we actually solve something, it's just a large language model that can write text for you and an entire essay, we don't barely call it AI anymore. So I think we we keep shifting the definition of AI and we're gonna keep doing that and we see a I right now already writing proteins generating photo realistic images based on a simple description uh and so many more things that would have been impossible 5 10 years ago. And I think as we as we keep doing that, well, we'll keep pushing that definition of what A G. I. Or artificial general intelligence might be. Thanks for having me here. It's good to be back in the United States. Deepmind is headquartered in London and actually our goal is solving intelligence to advance science and benefit humanity. I kind of audacious goal? Um but quite exciting. So I think that's actually my answer to your question is can we use artificial intelligence, artificial general intelligence to understand the way the universe works. Can we understand better understand chemistry, biology, physics, and can we use that to solve world problems and uh issues like life sciences, better understand disease, climate sustainability. And I think that ability to use um a more generalized intelligence uh to augment human knowledge to have some of these breakthroughs um is really going to be quite spectacular. And and how many years do you think it will be to get to general Intelligence? Uh It's anyone's guess right now. Um we were quite excited. I think the pace of advancements, I think as Richard mentioned, um it's speeding up. And so I think there's a lot of new approaches that are happening. The field is um getting more active. So it's quite an exciting time right now and um finding and I'm sure we'll get into this, but it just means that we have to continue investing more in some of the responsibility and ensuring that we are actually solving problems and not creating more. And and Richard, you've you've had some sort of spicy things to say about what some companies are doing in ai what what what's the beef? Yeah, I think ai right now is this kind of dual state, on the one hand, you have uh research towards a gi and continuously harder goals. The other hand is here, right? It's changing all our lives as we speak every time we make a search there a lot of Ai models. And unfortunately right now AI is often uh trying to just suck you into engagement loops online and a lot of very smart people at various companies and other search engines and social networks mostly try to spend time on getting you to spend more time looking at content, eventually click on an advertisement. And I think that's not ideal. I think it's important that as the I influences us, we need to have some explicit control over it. So at dot com and you search engine, we actually give you explicit control and you say, oh you want to see more Reddit results or less twitter or vice versa. You can actually up and down vote the rancor and change the neural network that you know, basically influences your information diet. And I think that's an important aspect of Ai that people that are influenced by it have some control over it and Atlanta, what are your thoughts on that, you know, Ai for good versus ai forgetting to click on more ads. Um, well and and that's really when I, so I have 30 years of experience in technology prior to deep mind and it was one of the reasons I chose to actually join this organization was because I felt like the mission was really what I was looking for, of how to bring my experience in tech and really focus on the impact that Ai can have. Um we're unique in the way that Deepmind works in which we are very much a research focused organization of kind of a bottoms up and top down and when we start new areas of research, we really think through what are the potential impacts and from the beginning design in a lot of the parameters that we want to consider around what might the risks be. Um where do we not have expertise that we want to bring in and who are the partners that we might want to work with um to help take some of the technology forward so that it does have the type of positive impact into society. And what what what do you think could be done to encourage ai for good more of it. I mean obviously like I said a lot of, a lot of it now is being done to get you to click more on facebook or on on ads. Is there anything specific that can be done? I think a lot of this right now is about um collaboration and um awareness. So when um the founders of the mind, I was we were comparing notes from 10, 10 plus years ago when the company was founded and at the time they had a hard time getting investments. In fact some of they were somewhat nervous to tell even their professors at the time what they were going to uh go do and start this this company at this at that time I was actually in venture capital in Silicon Valley and a I wasn't necessarily on our radar as something that an area we wanted to invest in. And I mean you just heard Richard like rattle off a bunch of examples of how ai has The terminology has advanced in people's understanding even my 12 year old daughters, my twin daughters like they know you know this is just all around them and how they interact with devices and what they're learning in schools etcetera. So a lot has changed in the past decade. So I would say that awareness is increasing increasing which is an important part of it. But I think collaborations are our key. Um Can I may I give an example? Okay, so um one of the areas that we've been working on in Deepmind is we initially started off with an environment of simulated environments and gaming just to try to develop new methodology around artificial intelligence systems And instead of so trying to innovate in that space, we then wanted to apply that more generally to real world problems. So we took on an area of protein folding um and protein you say, what is what are proteins? They're basically the building block of life. And if you can predict a three D structure of a protein, you can better understand its function, you can better understand disease, which comes with therapies, you can understand um how to deal with industrial waste. There's all sorts of interesting opportunities now Scientist biologist have been working for over 50 years to try to solve the protein folding prediction problem with our advanced Ai systems. We were able to work with biologists to make great advancements. We about a year ago we gifted to the world a database of millions of proteins and have been working with organizations were providing the um the model predictions but we're not the ones to be able to take this. This really is a collaboration and it's an opportunity to be equitable and access to the protein. So for example we've been working with a neglected disease folks to help democratize some of the understanding of diseases and drug discovery as an example. Um or we're actually working with an enzyme innovation center that is working on plastic eating enzymes. So we were able within days to provide 100 proteins where that they can start doing their research. So I think this type of collaboration public private collaboration, we're accelerating. We're leapfrogging areas of work that have people have had bottlenecks and trying to make some of these advancements. Really good organizations, smart people And this is a great example of where technology can help augment human understanding to benefit the world. And uh the protein space is really interesting at Salesforce research. Actually one of my last project when I was chief scientist there was to build a large language model similar to the large language models that can now right text If you actually are not familiar with this, you can go to you dot com slash right and have a whole blog post written for you. Uh but that same technology you can use to write english sentences. You can also use to write proteins. And so we actually then synthesize these proteins and generated completely new kinds of proteins in the wet lab that are inter bacterial and have, I think a lot of potential to change all of medicine. So it's one of those really interesting ones to get to your earlier question. I think ethics is becoming more and more important, especially in that part of the eye that really does influence our lives and it is essentially still just a tool and it's only as good as the systems, the people and the data that feed it and work with it. And so I think it's it's extremely important that we make it more transparent uh and also collaborate and and how do you, how does deep mind handle the problem or the challenge of of ethics when it comes to um yeah, you know, again, I'm fairly new to this field. So my experience in tech and when I was looking at where to it did, did I really want to spend this part of my career, like the prime of all this experience and apply it into artificial intelligence and you know, there were a lot of organizations at the time and I honestly, the more I learned about artificial intelligence, the more uncertain I got Like um and I kept thinking about this is the world my daughters are going to live in and can they go to put them to bed at night and they know what mommy was working on. And so I actually, when I was interviewing at deepmind, I spent about 50 hours interviewing just to make sure that was the field I wanted to go into. Was this the organization? And did I think I could have an impact and I often describe it as almost like a moral calling of yes, this um, I felt like there was something to contribute here from my experience in bringing new technologies um, into, into market and what do you mind? What I've been really excited about is from the very founding of the company, uh ethics and what we actually call it, pioneering responsibly. There are two of our values. Um but it's one of our first principles um and we think about things in terms of responsible governance. Like I mentioned earlier having these multidisciplinary teams of people that talk about our work very early on, not bolted on. We bring in outside experts as needed um responsible research because we are a research led organization. This is, you know, we've partnered with organizations like women in Machine learning. Um Deep Learning INDABA, which is an african effort to try to help the artificial intelligence that the computer scientists kind of come up to speed on the researchers on the latest trends co founded by some of our employees and then we've also do a lot of work around socio technical research, so areas of like queer fairness and D colonial ai and then the final area responsible governance, responsible research and then responsible impact. In the example of Alpha fold, our alpha fold system in collaboration with the outside world. To take try to understand what are the root node problems that we can deal and open up new fields of science again, to try to benefit a broader, broader societal impact. Let's go to audience questions. If anyone has a question, please look out for one of our assistants who has a microphone and paddle anyone. Thanks so much love this panel. Mat McElwain with madrona lila, super interested in how what you've been doing with alpha fold has been pushed by the competition from Rosetta fold at the Institute of protein design in Seattle. And not only from a research perspective, but from an open source perspective in terms of you all being more proactive about open sourcing after Rosetta fold took that path. Um, so if I can, I wasn't sure about your question, but maybe what I can do is one of the things that I mentioned is we really think about the responsibility all the way through and we wanted to make sure that whatever we were doing could go out into the world and uh how could we ensure that it would be used for good and mitigate risk because we had spent the majority of deepmind history focused in simulated environments and this was really one of the first broad scaled impacts. We wanted to take a very thoughtful approach to how we did this. Um we're not a traditional organization in that sense and I would actually say even being based in europe versus being based in the U. S. Has some difference in how we approach this rolling out of the technology and the collaborations that we have. We're approaching one year since our alpha fold system launched. And actually it's been quite exciting to see. Um I am generally, you know, with my background in tech, I love seeing so much goodness happen. Um and I think that when more organizations participate that it actually helps make us all better. Um the fact that we're having conversations about ethics and we're having comp conversations about ownership of your privacy etcetera. I think these are really good things because that's going to help hold us all accountable. I do love the push to open source more. It's kind of fascinating. But whenever someone has An incredible new capability, it usually only takes 3-12 months for someone else and another organization to replicate that capability and then usually open sources. So just today, actually hugging face. Open source, one of the largest language models. Full disclosure is one of my favorite investments. Um but you know, they are now at the very forefront and all the models, many from different organizations are put out there on the hugging face uh hub to be able to be explored by everyone else. And now that people used to think you couldn't train these very large foundational models in an open way but they showed that it is possible. So it's a very exciting time for ai to move forward quickly. Last question. Hi Jim Swanson from johnson and johnson. So very familiar with the protein folding. We're doing a lot of good work. I think party with you guys. Um one of the things that we continue to watch for is bias nous in the data either unintended or intended. Can you talk a bit more about what you're doing to look at uh parody of data to make sure bias is not built in and how you're assessing that over time and what should everybody should be looking for in that and that's important space. Go ahead. Richard. Yeah. Boy Quick. Please we only have 45 seconds. Yeah boy. Um we can talk for an hour about um bias. Yeah. As you know, uh many of you here are familiar with kind of the problem of thinking of your training. The model to predict the beauty of a person for instance and silly classifies that were trained. But the training data was just uh data samples from past beauty pageants that didn't have any woman of color in the dataset. And so they classify would not be able to possibly say a woman of color can be beautiful, which is terrible. And so this is just one of the many examples of bias is the same thing happens in giving out loans uh and you know restricting people in different areas of the world from getting loans or just in the wrong zip code because the classifier picked that kind of bias up from past uh biased samples and and data in the world. And so it's a problem that has no silver bullet. You have to really look into each industry and very carefully understand who is infected by that classifier or that machine learning algorithm and then do your best to try to mitigate and try to sample data from other areas or just get rid of certain features entirely. That might be relevant but would then result in a racist or sexist. I want to thank you we are out of time. I want to I want to thank both of you for talking about ai with us and thank you very much