0:01 People are worried about things like
0:04 privacy and and losing their jobs to AI.
0:05 How does a company like DeepMind build
0:08 the the trust of the general public?
0:10 What I want us to get to is a place
0:12 where the assistant feels like it's
0:15 working for you. It's your AI.
0:18 AI is scary. It's moving insanely fast.
0:20 And from an outsers's perspective, it
0:23 seems like there aren't nearly enough
0:24 guard rails. And some of these concerns
0:26 are actually legit. The ones that always
0:28 stand out to me are those surrounding
0:30 access. rich people with rich access to
0:33 the right tools or privacy. How do we
0:35 trust these big companies with all the
0:37 personal data? What does the world look
0:38 like when everyone is recording
0:41 everything and AI taking people's jobs?
0:43 One expert says a blood bath. Half of
0:45 entry-level white collar jobs
0:47 disappearing and 10 to 20% unemployment
0:49 could be 1 to 5 years. Some of these
0:51 concerns are less extreme than others.
0:54 Take the classic Skynet example. It's a
0:56 bit extreme, but they're all reasonable
0:57 because some of the things you've heard
0:59 me talk about before. First, the AI
1:02 race. These huge companies with vast
1:04 resources are trying to be the biggest
1:06 and best in the world of AI. Are these
1:07 companies prioritizing short-term
1:10 profits over long-term safety? Second,
1:11 even some of the scientists working on
1:14 it don't totally understand what's going
1:16 on under the hood. For example, some of
1:17 these models exhibit what's called
1:19 emergent behaviors. These are when they
1:21 produce outputs that even the engineers
1:23 who built these models had no idea what
1:24 they can do. In this video, I want to
1:26 look into whether or not AI is all doom
1:28 and gloom. Because if you listen to some
1:30 of the analysts, news outlets,
1:32 influencers, myself included from time
1:34 to time, it's the beginning of the end.
1:35 And it's impossible to put that genie
1:38 back in the bottle. But is it? Helping
1:40 me answer these questions is Deis
1:43 Habibus, a Nobel laureate, a knight, the
1:45 CEO of Google DeepMind, and one of the
1:47 most influential figures in AI.
1:49 Companies like DeepMind are the parents
1:51 of these AI children. And we're still in
1:53 the phase where the parent is
1:56 responsible when their kids mess up. So,
1:57 what steps are these companies taking to
1:59 ensure they raise responsible, well-
2:02 behaved young algorithms? It all starts
2:04 with trying to understand what's going
2:06 on inside the tech.
2:08 Can you sort of describe what's
2:10 happening under the hood with an LLM?
2:11 Like demystify it for people a little
2:14 bit. Sure, I can try. Um, at the basic
2:17 level, uh, what these LLM systems are
2:19 trying to do is is very simple in a way.
2:21 They're just trying to predict the next
2:23 word. And they do that by looking at a
2:26 vast training set of language. The trick
2:28 is not just to regurgitate what it's
2:31 already seen, but actually generalize to
2:33 something novel that you are now asking
2:35 it. LLMs predict the next word. For
2:37 example, if you go to a standard large
2:38 language model and give it the
2:41 statement, the quick brown fox, it will
2:42 likely complete the rest of that
2:45 sentence with the quick brown fox jumps
2:47 over the lazy dog. But the modern chat
2:48 bots that we use today are more like
2:50 question and response machines
2:52 fine-tuned to be more like assistants.
2:54 It's still doing the same thing, but
2:55 instead of trying to finish the
2:57 sentence, it's trying to answer your
2:59 question that you put into the chat. But
3:00 the trick here is that they don't want
3:02 that chatbot to just find a paragraph
3:04 from the original source material and
3:06 parrot it back to you. They want it to
3:08 come up with new information based on
3:10 all of the information it already knows
3:12 from within its training data. And if it
3:13 doesn't already know something, it will
3:15 either search the internet to try to
3:17 find it for you or in the case where it
3:19 doesn't have internet access, it'll just
3:21 make things up. And that is what we call
3:23 hallucinations. At IO, you announced the
3:26 new deep think, right, which is so much
3:28 more powerful and it's it's topping all
3:29 of the benchmarks for things like coding
3:31 and math and all that. What happened
3:32 under the hood that caused that new
3:34 leap? New techniques have been brought
3:37 into the foundational model space where
3:38 there's uh this called pre-training
3:40 where you sort of train the initial base
3:42 model based on you know all the training
3:44 corpus. Then you try and fine-tune it
3:46 with a bit of reinforcement learning
3:47 feedback. And now there's this third
3:49 part of the training which is we
3:50 sometimes call inference time training
3:53 or or thinking where you you've got the
3:56 model and you give it many uh cycles to
3:59 sort of go over itself and go over its
4:02 answer before it outputs the answer to
4:04 the user. What deep thinks about is
4:06 actually taking that to the maximum and
4:08 giving it loads more time to think and
4:10 actually even doing parallel thoughts
4:12 and then choosing the best one. And you
4:14 know, we've pioneered that kind of work
4:16 in the past, actually nearly a decade
4:19 ago now with Alph Go and our games
4:21 playing programs because in order to be
4:22 good at games, you need to do that kind
4:24 of planning and thinking. And now we're
4:25 trying to do it in a more general way
4:27 here. What's really cool here is how
4:28 Demis is highlighting how much of an
4:30 effort engineers and scientists are
4:32 putting into making AI more and more
4:34 accurate and removing the chance for
4:37 hallucinations. AI started with the next
4:39 word prediction, like the example of the
4:41 quick brown fox we gave earlier. Then it
4:43 evolved to test time compute where the
4:45 AI model would actually spend the time
4:47 thinking through its responses and you
4:49 were actually able to see this happen in
4:52 real time. And now the latest evolution
4:54 is what Demis just talked about which is
4:56 parallel thoughts. Now the LLMs are
4:58 thinking through a ton of different
5:00 potential responses all at once instead
5:02 of focusing on just one at a time. It
5:03 will then pick from all of those
5:06 responses or even combine responses in
5:08 order to give you the best possible
5:10 output. The ultimate goal here is to put
5:12 the most accurate and helpful responses
5:14 in front of you. You've mentioned that
5:16 the long-term goal is to sort of let
5:20 these AIs have like a world model. Can
5:22 you sort of explain what you mean by a
5:23 world model and what does that open up
5:25 to us? I think for a model, what we mean
5:28 by a world model is a model that can
5:30 understand not just language but also
5:34 audio, images, video, uh all sorts of
5:36 input, any input um and then potentially
5:38 also output. The reason that's important
5:40 is if you want a system to be a good
5:43 assistant, uh it needs to understand the
5:45 physical context around you or if you
5:47 want robotics to work in the real world,
5:49 uh the robot needs to understand the
5:50 physical environment. What sort of new
5:52 things do you think that'll open up to
5:54 people once they have that ability? Um I
5:56 think robotics is one of the major
5:58 areas. I think that's what's holding
5:59 back robotics today. It's not so much
6:01 the hardware, it's actually the software
6:02 intelligence. You know, the robots need
6:04 to understand the physical environment.
6:06 I think that that's also what will make
6:08 today's sort of naent assistant
6:10 technology and things like you saw with
6:11 project Astra that we show and Gemini
6:14 live for that to work really robustly.
6:16 You want as accurate as world model as
6:18 you can. So that's our glimpse under the
6:21 hood. LLMs are imperfect models that are
6:23 constantly being refined to become more
6:25 and more accurate with the eventual goal
6:27 of becoming complete world models that
6:30 help Ahi understand what's going on
6:32 around it in the real physical world.
6:34 But what's still unclear is how this
6:35 will translate into practical
6:37 applications that will significantly
6:39 improve society without a lot of the
6:41 downsides everyone is fearful of. So
6:43 you've mentioned things like AI will be
6:45 able to most likely in the future solve
6:47 things like room temperature
6:48 superconductors and more energy
6:51 efficiency and curing diseases. Out of
6:53 the the sort of things that are out
6:55 there that it could potentially solve,
6:57 what do you think the sort of closest on
6:58 the horizon is? Well, as you say, we're
7:00 very interested and we actually work on
7:02 on on many of those topics, right?
7:04 Whether they're mathematics or material
7:06 science like superconductors, you know,
7:08 we work on fusion, renewable energy,
7:09 climate modeling. But I think the
7:11 closest if you you think about and and
7:13 probably most near-term is building on
7:14 our alpha fold work. We spun out a
7:17 company called Isomorphic Labs to do
7:19 drug discovery, rethink the sort of the
7:22 whole drug discovery process um from
7:25 first principles with AI. And normally,
7:26 you know, it takes the rule of thumb is
7:30 around a decade for a drug go from sort
7:32 of identifying why a disease is being
7:34 caused to actually coming up with a cure
7:36 for it and then and then finally being
7:38 available to patients. It's a very
7:40 laborious, very hard, painstaking and
7:42 expensive process. I would love to be
7:44 able to speed that up to a matter of
7:47 months, maybe even weeks one day and uh
7:49 cure hundreds of diseases like that. uh
7:52 and I think that's potentially in reach
7:53 and sounds maybe a bit science
7:55 fiction-like today but that's what
7:58 protein structure prediction was like uh
8:00 you know five six years ago before we
8:02 came up with alphafold and used to take
8:04 years to find painstakingly with
8:06 experimental techniques the structure of
8:07 one protein and now we can do it in a
8:09 matter of seconds uh with these
8:11 computational methods so I think that
8:13 sort of potential is there and it's
8:15 really exciting to to try and make that
8:17 happen 10 years to a matter of weeks is
8:20 a pretty wide gap app. But to truly
8:21 understand this disparity, we need to
8:24 look at why it currently takes up to 10
8:26 years to bring a drug to market. It all
8:28 starts with the research phase. They
8:30 first have to identify a target such as
8:32 a protein or gene, which when altered
8:34 can treat specific conditions. The early
8:36 goal is to develop a compound that makes
8:38 that alteration. Once promising
8:41 compounds are found, we go through up to
8:43 7 years of testing in the lab and on
8:45 animals. And most compounds actually
8:47 fail at this stage for a variety of
8:49 reasons. is things like lack of efficacy
8:51 or toxicity. If the results are
8:53 promising, then the companies need to
8:55 get regulatory approval to get clinical
8:57 trials started on humans, which is a
9:00 process that has three phases of its own
9:02 and can each take several years. And
9:04 again, most drugs fail during this
9:08 phase. In fact, 90% never get past the
9:10 human trial phase. Once a drug does pass
9:12 all these phases, it then has to go
9:14 through another round of regulatory
9:16 approvals before finally being allowed
9:18 to go to the public. But here's where AI
9:20 comes in. That first seven-year long
9:22 discovery phase, it's going to be
9:24 crushed because AI can identify the
9:26 targets and compounds at an accelerated
9:28 rate. It can also detect toxicity and
9:30 side effects earlier, which helps to
9:32 weed out poor candidates before they go
9:34 to trials. The studies themselves,
9:36 they're also quicker because the rate at
9:38 which AI gathers and analyzes data is so
9:40 much quicker. The bottom line is we'll
9:42 get better drugs and treatments way
9:44 faster. But here's where it gets really
9:46 wild. In the beginning, AI was being
9:49 used to complete human tasks faster.
9:51 Now, we're starting to see AI training
9:54 AI, which when you boil it down is in a
9:57 way AI completing AI tasks faster. This
9:59 is where things really pick up. You guys
10:01 just announced Alpha Evolve recently,
10:03 which looks amazing, right? It's it's an
10:05 AI that essentially can help you come up
10:08 with new algorithms, right? How close
10:11 are we to AIS that are sort of designing
10:14 new AIs to improve the AIs? And then we
10:15 start entering this cycle. Yes, I think
10:17 it's really cool, a really cool
10:19 breakthrough piece of work where we're
10:20 combining kind of in this case
10:23 evolutionary methods with LLMs to try
10:26 and get them to get to to sort of invent
10:28 something new. Uh, and I think there's
10:30 going to be a lot of uh uh promising
10:32 work actually combining different
10:33 methods in computer science together
10:35 with these foundation models like Gemini
10:37 that we have today. So I think it's a
10:40 great uh uh very promising path to
10:42 explore. Just to just to reassure
10:44 everyone, it still has humans in the
10:45 loop, scientists in the loop to kind of
10:48 it's not directly improving Gemini. It's
10:50 using uh these techniques to improve the
10:52 AI ecosystem around it. Slightly better
10:54 algorithms, better chips that the
10:56 system's trained on versus it the
10:58 algorithm that it's using itself. This
11:00 is really important because it seems
11:02 like Demis is hinting at humans
11:03 eventually being removed from the
11:06 equation. AI gets better at training AI
11:08 and no longer needs humans to be
11:10 involved in its development. So where do
11:12 we fit in? The answer to that lies in
11:14 the end goal of all of these personal
11:16 assistants and agents. AI agents,
11:18 they've been sort of a a big talk in the
11:21 AI community recently. And how far off
11:23 do you think we are to being able to
11:24 give an agent like a week's worth of
11:26 work and it goes and executes that for
11:28 us? Yeah, I mean I think that's the
11:30 dream to kind of offload some of our
11:33 mundane admin work and and and and also
11:34 to to make things like much more
11:36 enjoyable for us. You know, you have
11:38 maybe have a trip to Europe or Italy or
11:40 something and you want the most amazing
11:42 itinerary sort of built up for you and
11:45 then booked. Um I I love our assistants
11:46 to be able to do that. You know, I hope
11:48 we're maybe a year away or something
11:50 from that. I think we still need a bit
11:53 more reliability in the tool use and and
11:55 again the the planning and the reasoning
11:57 of these systems, but they're rapidly
11:59 improving. So, as you saw with with the
12:00 latest project Mariner, what what do you
12:02 think the biggest bottleneck is right
12:04 now to to sort of getting that long-term
12:06 agent? I think it's just the reliability
12:09 of the reasoning processes and the and
12:11 the tool use, right? It's so and making
12:13 sure cuz each each one if it has a
12:15 slight chance of an error if you're
12:18 doing like a 100 steps even a 1% error
12:19 doesn't sound like very much but it can
12:22 compound to something pretty significant
12:25 over a you know 50 or 100 steps and a
12:26 lot of the really interesting tasks
12:27 you'd might want these systems to help
12:30 you with will probably need multi-step
12:31 uh planning and action. Removing the
12:33 mundane from our day-to-day sounds
12:35 wonderful but it also comes with the
12:37 inevitable questions about jobs being
12:39 replaced by AI. This is part of a
12:41 broader series of public concerns
12:43 surrounding things like privacy, data
12:46 security, and job loss that all big tech
12:48 companies are facing today. DeepMind's
12:50 association to Google comes with some of
12:52 the baggage. So that begs the question,
12:54 how does a company like DeepMind build
12:57 the the trust of the general public that
12:59 you can trust them with this kind of
13:01 technology? Well, look, I think we are
13:03 we've tried to be and I think we are
13:06 very responsible uh uh trying to be
13:07 responsible role models actually with
13:09 these frontier technologies. Partly
13:11 that's showing what AI can be used for
13:13 for good, you know, like medicine and
13:15 biology. I mean, what better use could
13:18 there be for AI than to cure, you know,
13:20 terrible diseases. Um, so that's always
13:22 been my number one thought there. But
13:23 there's other things, you know, where it
13:25 can help with climate, energy, and so on
13:27 that we've discussed. But I think we've
13:28 got to that you know companies is
13:30 incumbent on them to behave thoughtfully
13:32 and responsibly with this powerful
13:34 technology. We take privacy extremely
13:38 seriously uh at Google always have done
13:40 um and I think you know most of the
13:41 things we've been discussing with the
13:43 assistants they would be opted you know
13:45 you would you they'll make the person
13:47 the universal assistant much more useful
13:49 for you but you would be you know uh
13:51 intentionally opting into that very
13:52 clearly with all the transparency around
13:54 that. What I want us to get to is a
13:56 place where the assistant feels like
13:58 it's working for you. It's your AI,
14:01 right? Your personal AI. And and and
14:03 it's working on your behalf. And um I
14:04 think that's the mode, you know, that's
14:06 at least the vision that we have and
14:07 that we want to deliver and that we
14:10 think um users and consumers will want.
14:11 One of the things that you guys also
14:13 demoed at IO that I I got a chance to
14:14 actually test out a little bit earlier
14:16 was the Android XR glasses and those
14:18 were absolutely mind-blowing when I
14:21 tried them the first time. And uh so I
14:23 guess the flip side of the sort of
14:25 privacy thing is if everybody's sort of
14:27 walking around wearing glasses that have
14:29 microphones and cameras on them, how do
14:31 we ensure that the the sort of privacy
14:34 of the other people around us are is
14:35 secure? I think that's a great question.
14:37 I mean first thing is to make it very
14:39 obvious that you're it's on or off and
14:40 these types of things, you know, in
14:42 terms of the user interfaces and the
14:44 form factors. I think that's number one.
14:45 But I also think this is the sort of
14:48 thing where we'll need sort of uh
14:50 societal agreement and norms about how
14:52 do we do we all want if we have these
14:54 devices they're popular uh and they're
14:56 useful you know how do we want to what
14:59 are the kind of um the the guard rails
15:00 around that and I think that's still
15:02 that's why we're we're only in trusted
15:04 tester at the moment is partly the
15:06 technology still developing but also we
15:08 need to think about the societal impacts
15:10 like that ahead of time. So basically,
15:12 they don't know yet, which is
15:14 interesting and fair all at the same
15:16 time because ultimately when Demis
15:18 mentions the social agreements, he's
15:20 talking about government regulations and
15:23 legislation. AI is moving so fast and
15:25 we're all busy figuring out all the
15:26 other stuff going on in the world. We
15:29 haven't as a society stopped and really
15:31 thought about these implications. And we
15:34 need to because given the speed, we're
15:35 kind of running out of time. But it
15:37 makes sense that it's moving so fast. AI
15:40 is exciting. It's cool and the benefits
15:42 that it promises will change everyone's
15:44 life for the better. Just listen to Deis
15:45 talk about what he's excited for in the
15:47 near future. And remember, this is the
15:49 man who is on the absolute forefront of
15:51 this technology. So, I've got one last
15:52 question here. It's kind of a a
15:54 two-parter question. What excites you
15:56 most about what you can do with AI
15:58 today? And what excites you most about
15:59 what we'll be able to do in the very
16:02 near future? Well, today I think um it's
16:04 it's it's the AI for science work is my
16:06 you know always been my passion and I'm
16:08 really proud of what Alpha Fold and
16:09 things like that have empowered. They've
16:12 become a you know a standard tool now in
16:13 biology and medical research. You know
16:14 over 2 million researchers around the
16:17 world use it in their incredible work.
16:19 uh in the future, you know, I'd love a
16:21 system to basically enrich your life and
16:24 actually protect a little bit work for
16:25 you on your behalf to protect your mind
16:28 space and your your own thinking space
16:30 from all of the digital world that's
16:32 bombarding you the whole time. And I
16:34 think actually one of the answers to
16:36 that is that we're all feeling in the
16:37 modern world with social media and all
16:40 these things is is uh maybe a digital
16:42 assistant working on your behalf that
16:44 only at the times that you want surfaces
16:46 the information rather than interrupting
16:48 you at all times and and of of the day.
16:50 The thing about this technology is that
16:52 it's supposed to be the technology that
16:55 gets us away from the bombardment of
16:56 technology. We're sitting at our
16:58 computers and on our phones getting
17:00 flooded by negativity and toxicity on
17:02 social media every minute of the day.
17:04 It's refreshing to hear someone like
17:06 Demis, who's in one of the best
17:08 positions on Earth to build this future,
17:11 talk about how the importance of AI is
17:13 critical for our mental and physical
17:14 well-being. That we should be able to
17:16 cut out the mundane, remove the
17:18 toxicity, and focus on the things we
17:19 really want to do. Travel the world,
17:21 play guitar, pick up that hobby that we
17:23 never found time for, or most
17:25 importantly, spend time with friends and
17:26 family. In the end, after speaking to
17:28 Demis, I really felt like it wasn't all
17:30 doom and gloom, that super intelligent
17:32 and talented people are actually behind
17:34 the wheel and that they have a firmer
17:36 grasp than most people think they do. I
17:38 want to thank my guest Demisipus and the
17:40 whole team at Google DeepMind for the
17:42 incredible conversation. As always,
17:44 don't forget to like and subscribe, and
Chrome Extension