Deepmind

Fortune Editors

Fortune's annual list dedicated to the idea that business is uniquely suited to address society’s biggest challenges.

CTW.1122.Deepmind-Alphabet

DeepMind, a unit of Alphabet, is using artificial intelligence to solve daunting scientific challenges. Its AlphaFold software recently cracked a problem that had stumped biologists for decades: how to predict the structure of a protein based on its genetic sequence. Its predictions are becoming a go-to tool for scientists researching everything from breakthrough drugs to enzymes that can digest plastic.

Change the World

Change The World

Impact Segment

Sector

Get insights into Deepmind including business profile, stock information, news, and rankings from Fortune.

Deepmind Company Profile, Stock Price, News, Rankings

legacy

deepmind

Companies

Deepmind | Fortune

Deepmind Company Profile, Stock Price, News, Rankings | Fortune

Beatrice Nolan

Is AI-generated ‘workslop’ to blame for a lack of productivity gains?

/2025/09/23/ai-workplace-productivity-workslop-nvidia-openai-deepmind/

Maya Ackerman

I’ve been researching generative AI for years, and I’m tired of the consciousness debate. Humans barely understand our own

/2025/09/10/what-is-consciousnous-human-artificial-intelligence-tech-addiction/

Sage Lazzaro

Agentic AI is the technology’s new frontier, and CAIOs are toiling to get agents implemented correctly

/2025/09/03/agentic-ai-is-the-technologys-new-frontier-and-caios-are-toiling-to-get-agents-implemented-correctly/

Sharon Goldman

AGI talk is out in Silicon Valley’s latest vibe shift, but worries remain about superpowered AI

/2025/08/25/tech-agi-hype-vibe-shift-superpowered-ai/

Microsoft AI CEO Suleyman is worried about ‘AI psychosis’ and AI that seems ‘conscious’

/2025/08/22/microsoft-ai-ceo-suleyman-is-worried-about-ai-psychosis-and-seemingly-conscious-ai/

https://fortune.com/2025/09/10/what-is-consciousnous-human-artificial-intelligence-tech-addiction/?itm_source=parsely-api

https://fortune.com/2025/09/23/this-is-a-test-of-eye-on-ai-for-september-22/?itm_source=parsely-api

https://fortune.com/2025/09/23/ai-workplace-productivity-workslop-nvidia-openai-deepmind/?itm_source=parsely-api

Agentic AI is the technology's new frontier, and CAIOs are toiling to get agents implemented correctly

https://fortune.com/2025/09/03/agentic-ai-is-the-technologys-new-frontier-and-caios-are-toiling-to-get-agents-implemented-correctly/?itm_source=parsely-api

Alphabet’s Isomorphic Labs has grand ambitions to 'solve all diseases' with AI. Now, it's gearing up for its first human trials

https://fortune.com/2025/07/06/deepmind-isomorphic-labs-cure-all-diseases-ai-now-first-human-trials/?itm_source=parsely-api

https://fortune.com/2025/08/22/microsoft-ai-ceo-suleyman-is-worried-about-ai-psychosis-and-seemingly-conscious-ai/?itm_source=parsely-api

Researchers from top AI labs including Google, OpenAI, and Anthropic warn they may be losing the ability to understand advanced AI models

https://fortune.com/2025/07/22/researchers-ai-labs-google-openai-anthropic-warn-losing-ability-understand-advanced-models/?itm_source=parsely-api

Google DeepMind 145-page paper predicts AGI will match human skills by 2030 — and warns of existential threats that could 'permanently destroy humanity'

https://fortune.com/2025/04/04/google-deeepmind-agi-ai-2030-risk-destroy-humanity/?itm_source=parsely-api

https://fortune.com/2025/08/25/tech-agi-hype-vibe-shift-superpowered-ai/?itm_source=parsely-api

OpenAI and DeepMind are losing engineers to Anthropic in a one-sided talent war

https://fortune.com/2025/06/03/openai-deepmind-anthropic-loosing-engineers-ai-talent-war/?itm_source=parsely-api

When culture meets code, the future gets interesting. Grammy-winning artist and producer Wyclef Jean has long blurred the lines between creativity and technology: from mentoring Haiti’s next generation of roboticists to consulting for Google DeepMind. Jaeson Ma, CEO of OpenWav and a media power player behind ventures like 88rising and TikTok’s early success, is betting on the next big wave of digital storytelling. Together they explore the convergence of AI, entertainment, and the global creative economy, and what it means when artists become architects of the future.

Wyclef Jean
Grammy Award-Winning Artist, Producer, Composer; Chief Creative Officer, OpenWav 

Jaeson Ma
Cofounder and CEO, OpenWav 

Andrew Nusca
Editorial Director, Fortune Brainstorm

Soundwaves and Signals

Conferences

What must an AI company do to justify an eye-watering $300 billion valuation? Answer: Make money—and lots of it. As OpenAI reconfigures itself for revenue, the U.S. company is looking further afield than its home turf for growth. We talk to its international chief about scaling responsibly, partnering globally, and the unending pursuit of AGI. 

OpenAI's Global Ambitions

Sebastian Guth, Chief Operating Officer Global Pharmaceuticals and President, Bayer U.S.
Terra Terwilliger, Director of Strategic Initiatives, Google DeepMind
Moderator: Kristin Stoller, Editorial Director, Fortune

COO Summit 2024: AI And The COO - Leading The Transformation To A Futuristic Workplace

Sebastian Guth, Chief Operating Officer Global Pharmaceuticals and President, Bayer U.S., Terra Terwilliger, Director of Strategic Initiatives, Google DeepMind, Kristin Stoller, Editorial Director, Fortune

COO Summit 2024: AI And The Coo - Leading The Transformation To A Futuristic Workplace

Sebastian Guth, Chief Operating Officer Global Pharmaceuticals and President, Bayer U.S., Terra Terwilliger, Director of Strategic Initiatives, Google DeepMind

I've already spoken with a lot of you over breakfast. And I think the number one topic that came up was A I, which is why I'm very excited to be here with both of our panelists. But the number one thing I heard was people asking me how, what are the specific use cases? I need a case study for how other people use it so I can bring it back to my company. So Tara Sebastian, I want start by asking how are you each using A I both internally and externally and special? Let's start with you. Sure, you know, look for us A I would deploy very widely across the organization and we really deploy it in three different buckets. One is driving workflow efficiencies as probably many others in this room. And there's a couple of examples that come to my mind. Um um One I would highlight uh is uh medical coding. Uh So in our industry, when we run very large state scale clinical trials, uh we have to translate medical events that are often described in plain language into code. Um uh so that we can ultimately analyze the data. Uh and that's something we do quite extensively through the use of A I we've estimated in a, in a recent uh clinical study that, that has saved us about 100 and 70,000 hours of work. Um So it's very significant uh and, and drives efficiencies. Another example in that bucket of workflow efficiencies is regulatory submissions uh were mandated by the FDA and many other regulatory bodies around the world to submit very extensive documentation. Uh as we seek the approval of clinical um of products of medicines. Uh And we use generative A I to populate about 70 to 80% of those dossiers today. And if you translate that into efficiencies and savings, uh and ultimately acceleration is significant. The second bucket in which we use A I um is to advance patient care. Um So, um an example in that space is uh is mammograms. Uh Some of you may know that it's breast cancer Awareness Month. Uh And some of the audience may have gotten or may regularly get mammograms. Uh But what many individual people don't realize is is that it's in fact, not one image but hundreds of images that are being taken. Uh And we um provide software to radiologists that helps select the best image uh so that they can uh appropriately diagnose uh the patient and, and, and uh mammograms and breast cancer is one area we do this in other spaces as well. What's most exciting to me personally. As I think about A I though is the use of A I and technology in drug discovery. Uh because at the end, you know, we're in the business of bringing medicines to patients as quickly as we can. And I'd say that's a pretty challenging task. Um We spend about a billion and a half to 2 billion on each medicine that we bring to market. It takes us 12 to 15 years and we fail most of the time. Uh uh And, and, and technology can help us significantly accelerate uh that work. Um An example in that space uh for us is our partnership with a company called Recursion, uh A company in which we are an investor uh and a company that has arguably and likely the most advanced A I guided drug discovery platform in our industry. Um uh And we partner with them in them, bringing technology so deep expertise in technology, data sciences uh and, and use our uh library uh of uh of uh compounds uh to essentially identify and develop novel cancer therapeutics. And, and that just gives me chills because, hey, that's, that's the use of A I to drive real impact and ultimately develop medicines that otherwise would not see the light of day. Absolutely. And it touches everyone here too. Just like Google deep mind. Tell us what you're doing over there. I think probably everyone in this room is, is using the tech that you developed. So tell us about what you're doing. Thank you. So, Google deepmind is Google's Artificial Intelligence Unit. We create Gemini the foundational launch language model for Google as well as a host of other technologies uh which are while not in the lang large language model category are also being used for, for example, fundamental scientific discovery just as you described Sebastian. So we two are using internal and external uh the model in different ways internally. Uh legal is an example of where we have found great use for Gemini uh in our legal system, you can think of legal problems as having three different categories. There's retrieval of information, there's summarization and then there is reasoning and we are finding Gemini is valuable and across all of those categories, retrieval of information, of course, is finding the information that you know, you have somewhere, but you can't put your hand on it right there right at that moment. Uh summarization. And that's, that's really the basic I would say use case of simply retreating and finding that information. But it saves a remarkable amount of time. We're estimating that 40 to 50% of the request to our legal team could be handled simply by that kind of automation on summarization. Again, you're retrieving the information, but you're also pulling out what is most relevant. This is a more difficult problem particularly in the case of a legal context because you don't want to miss an item that might be important. So our team is increasingly using generative A I to to review for example, scientific papers and say what does need legal review in this case and what does not. Uh and then third reasoning as probably the most exciting category, legal problems as you know, are a good test case for gen A I because they are often unstructured, ambiguous and can have many different answers depending on jurisdiction, depending on business choices. So reasoning we think is the the next frontier here and the initial results are promising would not be complete without also adding to your point, Sebastian about scientific discovery. And this is an area. Our our mission at Google Deep Mind is to build A I responsibly responsibly to benefit humanity. And we are particularly proud of our work in scientific fields including alpha fold, which is the transformational protein folding technology. We created it. We won the international challenge of protein folding with the with this technology and in fact, released all of the 220 million plus proteins in the database for free for researchers to use in their own applications including drug discovery. Excellent. And I'm gonna come to the audience for questions in just a second. But I I have one for both of you. There's a lot of fear mongering when it comes to A I uh A I is gonna steal my job um both internally and externally. I feel like people are very, very scared to, to use it whether it's, it's bias. And again, people being scared about how it's gonna affect them. Um What would each of you say are the, is the biggest pitfall that you would advise C OS against when implementing A I into their business strategies to start with you? Sure. So first in our industry, I actually don't think that A I is going to steal jobs uh because hey, unpacking 3 billion years of evolution compressed in a cell is pretty damn difficult and that will continue to require also uh human capacity. Uh And, and the art of science uh which we're marrying up with technology. Uh big pitfalls, I mean, in our industry, we, we cannot and will not compromise patient safety. Uh So we're building a lot of redundancy into uh the way we work. I gave you the example of medical coding earlier. Uh And in that example, we use a four eye principle. So we in fact use technology, but then we still have humans double check that whatever data entry um we put into the system is accurate. Uh And that's only appropriate given that, that uh patient safety cannot be compromised. Um One piece of advice I would have from my own personal experience is that in large organizations and maybe that also resonates with some of the audience, there's at times a risk to use technology for technology's sake, you know, and in my mind, uh at the end of the day. Um It's a means to an end and the end in our case is to bring new medicines to the world at a much faster pace than otherwise than we would have otherwise been able to. And it's not about chasing shiny toys for the, for just for the sake of chasing them. What advice do you have? Yes, I would echo that in that A I is an unparalleled tool for fundamental scientific discovery. So how can we look beyond the immediate applications and think about how much there is yet to be discovered to solve fundamental human problems. Now that said, certainly we want to be mindful of how we're asking our people to use artificial intelligence internally because of those fears you just discussed to me the most important principle. There is co creation. So how can you with your team core? How the A I will be truly helpful to them in doing the job that they are already doing? They are the experts in doing that role. And I believe there's plenty of low hanging fruit. There is plenty of paperwork to be automated so that high skilled people can go do more high value tasks. Uh But that co creation piece is key. I'd say also making sure that, that you're valuing the time that people are putting in to learn how to use these systems and rewarding them for doing so in tech, we have a wonderful uh culture of dog fooding and using our own products internally. Uh And, and I would simply say that it's important to value the time people invest in that. And then third think forward about what this means for people's careers, both in reality and in perception, what new training might people need? How might this contribute to their training and learning in a way that's intrinsically valuable? I want to come back to training in a second. But does anyone in the audience have any questions here and, and raise your hand and we can run, run a mic around see one over there? Yeah, just in general. Um you talk about the growth of A I and the data, the need for data centers all over the place. What do you guys think about like power for the future and your needs power for the data centers. Uh Power for the data centers. Uh It's an important question. It's a question on many people's minds these days. Uh Again, we have to think about the fundamental scientific advances that are going to be possible with this technology. Uh We have, we actually at Google Deep mind work on novel power applications as well. So we have uh a piece on optimal power flow on how to increase efficiency in data centers to reduce the amount of power that is being consumed. We also uh we do work on fusion, nuclear fusion and how to optimize reactors. Uh And the, the fusion reaction inside of reactors. Again to get us closer to that place of renewable energy. Excellent. Thank you. Chair. Any other questions? He went over here and over here. Hi. Hi. So I work at an autonomous vehicle company, GEN A I based autonomous vehicle company and I found it really interesting with the release of chat GP T and the popularization of, of gen A I, there's been kind of an increased fear of, of A I and the potential impact it could have in society. I was wondering if you guys saw something similar and what that shift looked like between the A I conversation and gen A I conversation. Maybe I get us started, you know, in our industry. Um There's very few individuals that are ready to put their health into the, into the hands of machines. Uh and rightfully so, uh in my mind, uh so we've had a lot of conversations uh over the years more so in recent times to your point because it's suddenly in the public domain on, on, on how we use uh A I responsibly and how we ultimately a commend the work physicians and clinicians do, but don't necessarily aim to replace it. And that's a philosophy that keeps driving us because we continue to believe that uh that humans will for, for time to come play a very important role uh when it comes to advancing the health of those that we serve. And I would doubt. I think all of you play a role in that too. And how you deploy these technologies in the world in your workplace? Makes sense. I saw one over here. Thank you. Hi, thank you for the reminder on purpose driven A I, I think that was beautiful. So, thank you so much. So, my question to both of you is that what is the importance in your mind of learning about A I from other industries? And if that is important, how do you set your organizations up for being open to receiving ideas from other industries around us? Thank you. I believe this comes back to co creation again. So we want to build artificial intelligence with people, not just for people. So that intelligence comes into us from a variety of different routes and customers are a wonderful source of information. Uh We also work with outside organizations across a variety of uh represent constituencies, industries to try and understand what, what will create value for people. And I'd say at GDM, we are especially focused on the scientific community because again, we want to understand what will be of most value to scientists. How can we help them create the next fundamental breakthrough that will transform our futures and maybe just as a quick build, you know, in our industry, we learn quite extensively from others. Uh The promise of A I in the pharmaceuticals industry is massive, but we are for very good reasons. As an industry, probably progressing somewhat slower than others. Um You know, I, I sometimes described as jokingly as the race of the turtles. Um but that's a very conscious choice. Uh because I said before, you know, we're not going to compromise patient safety and the work we do has such a big impact on those that we serve, that we cannot take shortcuts and we probably, and possibly can't sort of experiment uh to the same extent other industries can. But that also gives us an opportunity which is to learn. Well, thank you for the incredible questions. Thank you, Sebastian and Tara for being here. I appreciate it. Thank you. Thank you.

Jeff Dean, Chief Scientist, Google DeepMind and Google Research, Google Interviewer: Jeremy Kahn, Fortune

Brainstorm Tech 2024: Google’s Gemini Man

Everybody. Thanks for joining us, Jeff, thanks for being here. Um I'm gonna go to the audience for Q and A pretty early on in the session. So please have your questions ready. Wait for Pat Mike handler to come to you and then stand up and please say your name and state your question. Um So Jeff in the introduction, Andrew talked a little bit about the merger between Google Brain and Google, Deepmind, Google had these two separate A I fundamental research units. What was the reason for combining them and what benefit have you seen from that, that merger? Yeah, I mean, I think both organizations were doing fantastic work and you know, a lot of different A I innovations have come out of both groups prior to the combination, you know, things like Transformers and Burt and alphago and various other Alpha related projects on the Deep Mind side, Alpha fold. But I think as we realized that we were both striving, you know, both organizations were striving to build, you know, large scale multimodal models that can do lots of interesting things. It made sense to kind of combine the people working on those ideas so that we have, you know, a better set of ideas to build on to pool the compute. So we sort of focus on training one, you know, large scale effort like Gemini rather than multiple fragmented efforts. And so that's been really good. I think it's been really great to see people come together and really just kind of work together on, you know, building awesome stuff. Right. And has there been a change in sort of the tempo with which you're able to sort of move things from research to product as a result of that merger or a change in the culture at all? From the from that merger? I mean, I think obviously we want to build, you know, things like Gemini that can sit underneath a lot of different products. And so we have a pretty strong focus on, you know, how do we make great models that are great for, you know, cloud applications, but also, you know, sitting underneath various search features or Gmail or other kinds of, you know, existing Google products, but also can create new kinds of products. So that sort of large escape of product focus is really important, but we're really focused on the fundamental research and making the best things possible and then rely on product teams to kind of, you know, take those building blocks and figure out how best to use them in their setting. Great when I talk to companies about using generative A I one of the things that always comes up is the rate of hallucinations from these models. And it seems to be one of the things that's created a bit of a bottleneck to actually deploying them in any kind of real use cases. I know you guys are working very hard on this issue. There was some hope that when multimodal models came around like like Gemini that that would reduce hallucination rates because the model would understand both the language and also the visual imagery that went with that language, it's made some difference, but that hasn't completely gone away. How are we doing on solving the hallucination problem? And is there kind of a, you know, do you think it's a solvable problem in the current architectures? Yeah, I mean, I do think we're making progress like it's a difficult problem for sure because the models are trained to generate, you know, probabilistically plausible sentences and those are not always true. So, but I do think there are some glimmers of hope. So, one of the things in the latest Gemini models is the fact that it has this very long context capability. So you can put lots of information that you're trying to sort of use and in answering questions that you that you're doing, and that information to the model is much clearer than sort of the model, the training data that the model was exposed to in the pre training phase, which is sort of like a big jumble of tokens that are used to generate these probabilistic models but is not as clear as you know, information in the context, right? So that information I think is much easier and the model tends to hallucinate less about that information. I think that's, you know, one promising direction and we're obviously pushing on lots of different other directions, right? At Google IO. This year, I know you guys made announced that you have this thing called project Astra, which is an attempt to kind of build an a a model that have agency. What what kinds of actions will astra be able to perform and how soon is this going to be released? Yeah, I mean, we're hoping to have something out into the hands of test users by the end of the year. So it's pretty exciting. I think the ability to combine Gemini models with models that actually have agency and can perceive the world around you in a multi modal way is going to be, you know, quite quite powerful. I think we're obviously approaching this responsibly. So we want to make of the technology is ready and that it doesn't have unforeseen consequences, which is why we'll roll it out first to a smaller set of initial test users. But I think it's going to be pretty interesting because all of a sudden, you know, the model can be perceiving the world around it around you and doing, doing stuff on your behalf. Right. Interesting Bill Gates recently said he thought, you know, large language models, there were sort of two more turns of the crank in terms of scaling up these models further. But he also said that he didn't think that those two turns of the crank would kind of get us A G. I'm curious what you think, how many turns of the crank are left on sort of transformer based LM models and where will those turns? Get us? Yeah, I mean, I think as you see the models grow in capability, some of that is due to scaling. So every time you increase the scale of the model and the amount of training data that these models are trained on with the model capacity as well. You do see new capabilities kind of emerge, you see the models, you know, hallucinate less, you see them sort of able to do things that they couldn't do before at smaller scale. So that that is a promising sign and you know, a couple more generations of scaling will get us considerably farther. I do think we're going to need some additional algorithmic breakthroughs, which is something that we've always focused on is how do we sort of take scaling but combine it with algorithmic approaches that combine well with scaling to sort of improve the capabilities of the model. I think things like factuality and reasoning capability where the model can sort of uh sort of imagine plausible outputs and you know, reason its way through which one makes the most sense or the kinds of things that are gonna be uh important to really make these models robust and, and more reliable than they already are. Right? I want to go to questions from the audience. We have a question, please raise your hand and we'll get a, get a mic handler to you. There's one right down here. Um If we can please get a mic to this gentleman and if you could please stand up and say who you are? Alright. Thanks, Mark Paper Master. I'm CTO at A MD. And uh Jeff, we're all um excited with the advancements like you, you all are doing with Gemini, but we also all are talking about the the demands of power. How do you think about it holistically from your role in terms of everything from the chip engineering all the way through the uh models in the application stack? Yeah, it's a good question. I think one of the things that, that I really like about Google is we have a fully integrated system. So we have, you know, we design our own chips for A I, although we also use other people's chips and then we have our own sort of system software on top of that, we can make efficient use of those chips. When we first introduced the tensor processing unit in 2016, you know, that was quite a big advance in terms of energy efficiency. It was about 30 to 80 times more efficient than CPUS and GPUS of the day. And we've now been through multiple successive generations of CPUS that have also made pretty significant improvements in energy efficiency. And we've also, you know, for many, many years focused on the overall energy efficiency of our data centers. We have very low P we have, you know, a lot of data centers of ours are powered by percent or more clean energy, which is super important as despite this, Jeff, you just reported that your co two emissions have gone up over the last several years, despite these sort of improvements and despite these optimizations, so can you talk a little about what's going on there and sort of how, what's the path out of this? I mean, if, if we have, you know, carbon footprint scaling at a rate that's, you know, going up at the same kind of slope as capability, maybe that's not where we really want to be. Yeah, I mean, we have a public commitment to by the end of 2030 you have essentially 100% of our data center operations powered by clean energy. And that's not necessarily a linear thing because often we make, you know, we uh work with clean energy providers, solar, solar, you know, capacity providers or wind providers and some of those things come online in 2027 2028 29. And so those things will provide sort of significant, you know, jumps in the percentage of our energy usage that is from energy. But then also we want to focus on making our systems as efficient as possible, right? And, and is, is, is the access renewable power become a constraint on sort of how quickly you can scale the technology. Um You know, I think there's been a lot of focus on the increasing energy usage of A I and you know, from a very small base that usage is definitely increasing. But I think people often conflate that with overall data center usage of which A I is a very small portion right now, but growing fast and then sort of attribute the growth rate of A I based computing to the overall data center usage. So I think it's really important to look carefully at all data and really understand the true trends that underlie this cool other questions from the audience. Please raise your hand. If you have a question, if not, I've got plenty more for Jeff. But if someone else in the audience has a question over here in the back, the gentleman in the colorful shirt first, I think, and then if you can please stand up and state your name, hey, Will Wilson from antithesis, Jeff, what is your current take on the debate over A I risk and alignment and all that stuff? Yeah, I mean, I think it's a, it's a good and healthy debate for sure. I mean, I think there's, you know, many different viewpoints along a spectrum of, you know, optimism and pessimism and, you know, how much concern should we have for, for, you know, disastrous consequences? I guess I'm somewhat in the middle. I think there's a huge amount of benefit from A I technologies to the world. So we should make sure that we sort of do, do our best to make sure that the world gets access to those things for things like education and health care. But we also need to make sure that some of the, the risks of A I are sort of mitigated or diminished in, in rolling these things out for the positive benefits. So it's, you know, obviously hard to predict the future in a very fast moving field. Uh I think probably the most extreme viewpoints on either side are probably not, not the most uh likely outcome. Do you have a take on, on the California Senate bill? 1047? Do you have a view on whether that's a good approach to regulation or not? I don't, I mean, I would say in general, it is really good for governments at all different levels to be thinking about how they want to regulate A I, I think there's certainly, you know, I sort of think of it as a very general technology and often the regulatory process probably should focus on how is that technology being used in a particular setting, right. So like if you have autonomous vehicles that makes sense to have like a certain set of regulations there in health care settings, there's already sort of a relatively comprehensive framework around roll out of new technologies and health. And so I think, you know, adapting those for new A I based things makes a lot of sense. Cool. I know we have at least one other question because I saw a hand up before. So why don't we go to the uh someone else? Um Just here. Yeah, please say it stand up and say your name. Yeah, hi. This is Sidar Taral with s software company work. So Jeff, I'm a zoo. So I have a soft spot in my heart for Google. But having said that one of the challenges we have is as we look at models, we don't want to be locked into one particular model because we don't know which model or which company is going to drive innovation the fastest. So what's your advice for companies like us in terms of how we should think about not locking ourselves out of future innovation by picking Gemini. Yeah, I mean, I think I hear those concerns. I think, you know, we think we're going to innovate quite quickly and then give people what they want in terms of models. But we also recognize there's a whole bunch of choices. This is partly why we, for example, have open source models called Gemma that we've recently introduced. We just provided Gemma two updates a few weeks ago of 9 billion and 27 billion parameter models and the 27 billion parameter model on the LM SIS um sort of elo rankings actually is the top ranked open model at the moment or it was at least a week ago. I haven't looked this week. This is a fast moving field but, you know, open source models can provide people some comfort that they have sort of a little bit more control. But I also think a lot of people want sort of a PS that they don't have to necessarily spin up a whole bunch of GPUS themselves. They can just sort of use it as a building block for, for getting stuff done. Cool. Uh Other questions down here in front. There's a question I see. Uh if we can get a mic to this gentleman just here right here in the front row. Oh Can we get a mic to this to him right here? Sorry. There we go. Thank you. Please say your name. Hey, my name is Ken. I'm uh CU and co-founder at union. Firstly, I'm a huge fan of your work. 20 years. Every paper, everything. This is what we have built on. Secondly, I think this is mostly looking into the future, purely technical question, not talking about the implications of business, but what's the one thing uh in a that will surprise you in a good way, in a bad way in those cranks that we're spinning. Like let's say we have this two, the next generation. I'm not even talking about the second one. What is the one thing that will surprise you in a positive way? And what will disappoint you if it doesn't happen? Like, yeah, I, I mean, I think um one of the things that is probably going to be a surprise to people is the improvement in multimodal capabilities. You know, I think the first couple of iterations of Gemini really sort of introduced multimodality, but I think it's going to be a much more naturally integrated thing for many different kinds of modalities, audio video imagery, both on the input and generation side. So I think that's going to be an exciting new, new direction. Uh you know, I'm the, the what will surprise me on the downside that's always hard to predict, right? But it's always gonna be disappointing when it happens because you're like, oh, I could have sworn this thing was gonna work out but it didn't. And you know, I think this is part of the nature of kind of a a research driven but still fairly close to product kind of development framework is you're constantly trying to bring new research innovations into the thing you're building at the largest scale and make sure that those things have been de risked at a smaller scale. And sometimes those interactions, you can really predict really well. And we, we do our best to sort of scientifically rigorously evaluate how they're gonna interact. But when they scale up, sometimes they interact in ways that you can't predict. Ok, let's, let's get another question while we still have time. Let's go right here. You got, hey, Healy Ier boom pop. Thanks for being here. Um One thing I think about a bunch is the fact that education in the States is like we're 60 or 70th out of 194 countries. And I think with access to A I, it's wonderful because as internet and mobile phones are around, everyone can ask questions and get answers the trick, I guess. And this is kind of a very high level question which I think is the stuff you think about probably is if you're getting the answer all the time, how do you also balance the responsibility of giving that person who might not have education to think critically? Do you think about the answer? And how do you caveat? How do you say look a I may be right? May be wrong. And how do you make sure that I guess critical thought remains center in humanity as A I becomes more prevalent. A good question. Yeah, it's a very good question. First. I think there's huge, huge potential for A I to help in education, right? We know educational outcomes when you have a personalized tutor are two standard deviations away from those when you have a much more sort of group focused setting where you're not being challenged at the level of an individual and, you know, someone who knows what, you know, and knows what you don't know. So I think that's the potential. I think, you know, part of deploying these systems in education if you want to have the system really challenge people in how they learn best and make sure that they're actually mastering the material before they sort of go off, but it does enable them to kind of take the right educational path for them. And I think that's, that's hugely huge potential. Yeah, I just want to ask a quick, there's been a lot of concern that the use of generative A I is churning out a lot of misinformation. A lot of poor quality content. Google obviously is a company that its entire business is based on finding good quality information on the internet. And there's a lot of concern about the poisoning of the information ecosystem. How are you looking at that at Google and how you're concerned that this technology you're helping create is also kind of rooting the very thing you built the business on? Yeah, I mean, I think this gets to our responsible A I principles like we've, we've always focused on how do we roll out A I technologies in ways that sort of benefit people and that sort of minimize things like harmful bias or misinformation or other things like that. And so it is true. These models can create misinformation. Misinformation though is not a new problem. It's something that's existed for years and years and, but A I also gives us tools to battle misinformation so you can do both. I'm afraid we're out of time, but I want to thank you all for listening. It's been great, Jeff. Thank you so much for joining us. Thank you for having me.

Deepmind

Videos