# What Google's Secret AI Lab Actually Looks Like (I Got Exclusive Access)

## Метаданные

- **Канал:** Varun Mayya
- **YouTube:** https://www.youtube.com/watch?v=EN9ZdSlvbsg
- **Дата:** 25.02.2025
- **Длительность:** 33:57
- **Просмотры:** 226,261
- **Источник:** https://ekstraktznaniy.ru/video/12040

## Описание

We got exclusive access into Google DeepMind and the prestigious AI for Science Forum, and we have some fascinating insights to share.
Watch this video to witness the brilliant minds shaping our future through groundbreaking projects like AlphaFold, AlphaProof, and SynthID. We’ll explore game-changing ideas, from cracking protein folding to designing gene-edited cows that slash methane emissions. Not just this, we even capture DeepMind's unique workspace culture with scientist-named meeting rooms and a library that would make any polymath envious. 
But beyond the tech, this video uncovers DeepMind’s real edge—its culture of curiosity. Whether you’re an AI enthusiast, a future innovator, or simply fascinated by how the world’s sharpest minds solve problems, this is a rare glimpse inside a world few ever see.

00:00 - Introduction to the AI Race and Google DeepMind
02:33 - AI for Science Forum
04:57 - The Polymath Advantage in Research and Success 
08:08 - DeepMind’s Founder: A Journey A

## Транскрипт

### Introduction to the AI Race and Google DeepMind []

so you all know that there's an open race to artificial general intelligence an AI model that can basically do all the tasks a human being can and when you think of companies that are in this race you think of open AI of anthropic of meta and then of course Google you know what there's a specific point here that you're missing Google is a commercial entity that does everything from search to email it's in the race to deliver those powerful AI models to you but the question is if Google is so busy doing all this other stuff who's making the models who's doing the research well that's where the plot gets interesting Google made an acquisition in 2014 of a company called Deep Mind founded by Demis hbus Shane leg and Mustafa suan yes the same Mustafa suan who is now the CEO of Microsoft AI Deep Mind is Google's AI arm but here's the thing most people don't know much about Google deep mind it's a mystery how it works what the office looks like what they're actually working on what the leadership is like it's a secretive company but I got access in fact Google Deep Mind along with the Royal Society hosted an AI for science forum and this was a science Forum that was polyic in nature you need to have a bit of understanding about AI little bit about medicine little bit about mathematics software in general people and even game development so there were four Nobel Letts and a bunch of really smart people there however there were only two creators there me and verasha so not only did I happen to see everything that happened in the AI for science Forum constantly trying to keep up with the conversation but we also got to see the insides of Deep Mind not just that I ended my London trip with a podcast with pushmi Ki who is the vice president of research to unravel the mystery of deep mind so I'm here to show you the Insider view of the company see as somebody who doesn't make the models but fine-tunes them and then uses them to make useful things that users can use or View or play being able to get this access for me selfishly means I can know what's coming from years away of course when I say some of these things a lot of people in the comments are going to shout at me like when I said AI would help solve design when I got preview access to Del a couple of years ago but I think for those that are curious and want to build on top of these models who want to work in applied AI this video will be very useful plus to see how the smartest people in the world think how they work what their Ambitions are what they read all of it is very useful they don't think about the world like normal

### AI for Science Forum [2:33]

people okay so let's start with the AI for science Forum the night before was a welcome dinner unfortunately photography was banned so there's a bunch of deans of colleges researchers and also the media there were reporters from the guardian the sun and a bunch of prestigious Publications like I said there were only two creators Dr Derek verit Asim and me now you might ask what are creators of the media doing there it's a science Forum the answer is very simple research and progress stays in very small rooms unless these researchers and media coordinate to get the word out there it turns out that even though awesome research is happening the world probably doesn't know about it for example did you know that Alpha fold 3 is now open source well now you know and that is the value of having the media and the creators there we can help this new technology and developments in science reach more people my audience is now 40% Global and reaches a 100 million views across platforms on sh form content and companies like Google and meta have me there because it's a very specific type of smart audience that we put together here the early adops that's right large companies think you the viewer are smart if you're watching this kind of content so if you're not already subscribed you should hit the Subscribe button anyway it was a great dinner and I learned some of the secrets of the Royal Society see the Royal Society is a society of distinguished scientists there are bunch of stuff like treasures and weird locks and lot of painting someone showed me a painting and said that's an actual painting of Isaac Newton there now I wasn't able to verify this but it seems like it's true the next day The Forum started and I got to hear all the panels so there's a lot of talk about Ai and there's also stuff I didn't really understand like Material Sciences also learned that they're making a new type of Gene edited cow where the microbiome genes are edited by targeting specific microbes responsible for methane production these scientists are aiming to create cattle that emit significantly less methane so they're actually addressing a major source of greenhouse gases in agriculture and that leads me to an interesting point one observation I made here how none of these people are stuck to one specific domain this problem statement of the cows is that the intersection of genetics microbiology and environmental science like my one big learning from this entire London trip and meeting all these awesome people is that most of this outlier extreme success comes from being

### The Polymath Advantage in Research and Success [4:57]

a polymath is somebody who has a wide range of knowledge across many different subjects in fact it turns out if you're really good at one thing you learn the skill necessary to learn things so you get very good at something let's say you've software engineering youve not only learned software engineering you've also learned the skill on how to learn something and then you can apply that learning skill to other domains take me for example I think I can do content I've done software for more than a decade I can do applied Ai and we built a lot of things in applied Ai and we can develop games as well we've made a bunch of Demos in the past both 3D and 2D actually in my personal opinion I think I can pick up anything in software in general but most people tell me most people in the world tell me you can't do all of these things you have to do just one thing and I tend to struggle with this line of thinking cuz the thing is I enjoy whatever problem statement I'm working on and if you follow me on Twitter you'll see that I spend months trying various things in that domain to understand it and get better at it these people at the event were so inspirational to me because they are like that but on steroids they're all in their 40s and 50s but there's a multi-dimensional universe in their brains I never once heard someone say oh that's not my domain I don't want to understand it instead they would ask questions well what does that mean is it similar to this uh this is how I think it works am I correct can you give me feedback on where I went wrong and most of these people have figured out that across domains there are ton of repetitive patterns especially when it comes to working with people or computers whether you're building a building or whether you're doing research search ultimately you are also working with people across both so a lot of repetitive patterns are there and most of these people today across all these domains are of course using computers like the geneticists are doing an insane amount of work on a computer so the better you are at adapting to in using new technology or software the better the geneticist you become like the only reason I know anything about genetics apart from what I studied in college is because of curiosity I don't know if you know this but I MA my whole genome in 2017 and then I use this online tool called Prometheus to give me more information on my genome like I Had No Agenda there's no I'm not trying to do this to pass some test I was just really curious about my own genome and the one thing about these people all these people there is that they're all very curious about life about the world about other Industries and domains I'm actually starting to lose faith in the idea of a person who's just very good at one thing and not curious about anything else you know all our lives we hear about the expert the person who's very good at one specific thing and knows nothing else about anything else or doesn't have any relevant information about anything else I think that is a TV level understanding of people like on TV we have experts we have astrophysics experts who come and talk about space and science or the economics expert who comes on TV and speaks about economics when I was young I thought that was the peak of success but now that I'm older and I get to hang out with people who are very successful not just 50 lakhs and 1 CR an like way bigger than that I noticed that they're all polymaths actually they're all really curious people that's why they're polym mats they are curious even if there's no commercial goal there almost everyone in the Forum was like that I'll give you an example sir Demis

### DeepMind’s Founder: A Journey Across Domains [8:08]

hbus the founder of deine started as a competitive chess player then he did game design he actually made a theme park game then he was executive designer of the game Evil Genius then stopped that went and studied Neuroscience wrote a paper on episodic memory and memory recall in scenes then start a deep mind and then made an AI to play Atari games then went and Sol protein folding in medicine and Life Sciences it's insane it's like domains don't matter to these people and it's so inspirational in fact just being there changed me in some way it kind of reminded me that life is short and we're all just here temporarily and you might as well do everything you're curious or interested in because not only is that where you'll be most excited but weirdly it's where you will also make money and attract other people to work with you because that passion is genuine and true and driven by your curiosity when I was 25 I would say things like passion is useless you need discipline etc I was wrong it's passion all the way through even if that passion doesn't make sense to other people one more thing I noticed is that Deep Mind puts in a lot of effort to explain things to people with basic science literacy like they could have explained everything in a very complex way but they made entire demos to try and explain their project for example on the human connecto or on Alpha fold they are very simple demos you can use so they're trying their best to communicate it to the masses it's the opposite of what some people on Tech Twitter do for example where they try to make things as cryptic as possible and do gatekeeping by not making things simple they enjoy the complexity because it gives them status but the folks in Deep Mind are very different they actually want the science to reach the masses so they try to keep things

### Tour of the DeepMind Office [9:49]

things simple okay anyway here's the good part the next day we got to Deep Mind the main office firstly to get in Deep Mind you have to sign a non-disclosure agreement so I can only reveal a small fraction of what I actually saw there and what I learned but I'll try my best sometimes senior employees at Google India will joke with me saying why don't you've seen more of Google than they have but deep mind is really secretive and that street that specific Street where Deep Mind is has a bunch of other tech companies there's meta right there too so anyway we first entered got our badges there's this amazing art installation in deep Minds entrance it's like looking through an infinite space very cool and all the rooms have names of famous scientists so we started on the eighth floor where we saw that General work setup we weren't allowed to record some of it but in general it's a bunch of very smart people working on variety of different things from Life Sciences to robotics in

### Four Exciting DeepMind Projects [10:48]

short I'll tell you four projects that I really liked and I was interested in the first one is Alpha proof which is solving International math olympiads pushmi told me and you'll see this a little later in the video that it does reasoning like opening eyes o1 okay there's a Chain of Thought going on but it uses a formal language called lean to reason so less hallucination it's almost like it allows it to think in cold the next one as you know is Alpha fold solves protein folding for decades scientists have been trying to figure out how proteins fold why because understanding this folding helps us do amazing things like designing better medicines or diagnosing mutations it looks at the basic pieces of a protein called amino acid and predicts how the whole thing will twist and turn into its final shape they're now solving problems in weeks that used to take years or couldn't even be solved at all for PhD students the next thing that deep mind is working on is Project Astra which is an always on AI assistant that you carry with you in your glasses it can help you see the world and with everything you see a tap in front of you don't know how to fix it you can say well AI help me fix this tap and the last thing that I was really excited about is of course Vo the video model I mean we work in AI plus content remember our product Alpha CTR that we built or the AI Avatar that we use that literally power the 100 million sh form views any and all advances in models for Content have an early user in me so one big Advantage for me to be there is well I'm able to use Early Access versions of these models before everybody else and therefore we're able to run these experiments faster than others okay two cool things about the

### DeepMind's Impressive Library and Cafeteria [12:24]

Deep Mind office first they have an amazing cafeteria with all sorts of food and drink but to be honest I don't care so much about food and drinks secondly they have an amazing Library like the library signifies how much they care about being polymaths and it has a lot of different books from Neuroscience to travel to physically based rendering in game development and all the way to Java and HTML I'm actually going to make a full video on all the top books in the library and one of my goals in life is to have all those same books across all those Topics in my house and actually read all of them I was actually surprised to see that I've read some of the books but one of my new goals in life is to make sure I have a perfect replica of that library in my house at some point not only that there are copies of the nature magazine all over the office the library is very inspiring and I think it deserves an entire video on it if you subscribe I'll make a full video talking about every single book or at least the top books there so make sure you do that and it also shows you what the Ambitions of the people in Deep Mind are it's not really money is a side effect right they've made a lot more money than everybody else because I think it's a side effect of just following their Curiosities and being really good at it most of them are already very rich for them it's actually their ability to satisfy their Curiosities and pursue this NeverEnding goal of learning more about the universe and our place in it but I didn't stop there you know I'm a person from the outside looking in and I wanted to talk to the vice president of research and ask him for everything how does Deep Mind work how do they hire people what kind of projects are they working on and whether llms have hit a wall and finally what the path to artificial general intelligence is an AGI system is supposed to solve all of these problem statements it's basically supposed to Encompass the entire library and instead of me breaking down the conversation for you I'm actually just going to place in the entire full conversation trust me it's worth a watch and it'll give you a lot of ideas around what the future is going to be like ladies and gentlemen this is pushmi Ki the VP of research of deep mind again a polymath and his work spans everything from using a camera to do motion capture all the way to Medicine here goes pushit

### Interview with Pushmeet Kohli: Inside the "Intelligence Factory" [14:37]

thank you so much for doing this um you know I had a tour of Deep Mind today like the office I don't think anyone like most people sitting in India know what's going on in Deep Mind and how influential you folks are in AI so what happens in these like mysterious Halls it's it's like any sort of institution any organization uh we call it the intelligence Factory what we do is we extract intuition from um raw material which is in our case data and this data can be data that has been carefully collected by scientists through experiments or it could be data that we have created uh through simulations uh but essentially there is this amazing amount of data that we are collecting about ourselves about the planet that we are living in and how do we make sense of it all it comes to a point where the data is so large that one single human mind is not able to make sense of everything that we have collected and so that is where we need uh techniques like machine learning and AI which are able to extract these hidden patterns within the data and allow us to make predictions that of quantities that uh are incredibly

### The Future of Programming with Alpha Code [15:58]

important I actually want to ask you a question on Alpha code because you've also worked on Alpha code I think the team has competitive programming for a long time has been this marker of intelligence you know for the human it's like well you know here's my lead code score uh but I think it's getting very good right the AI there is getting very good across you know all of these different platforms and a lot of software Engineers are worried right as in hey if this is going to be better than me at competitive coding at what point does it build software by itself do you have thoughts on this where this evolves over the next four five years so what is the act of programming problem solving I think the like one way to think about programming is basically problem solving but in terms of the task description what is the task of programming is basically taking a specification yeah of a problem and that specification sometimes is a informal specification sometimes somebody says oh I want you to write a program that does this that's an informal specification right of the problem or somebody might give you certain examples they might say well here are certain examples for this input the program should give this output user stories yeah exactly right so but these are incomplete and informal specifications of what you were supposed to do right and the act of programming is to take this informal incomplete specification and translate it into something that is formal rules very formal rules which is fully defined from the domain to the which is completely crisp yeah right and so the magic of a programmer is not in just doing that translation but also in filling the gaps so in unzipping the problem statement into its you know constituents Al also but because many times the problem itself is unspecified for most human problems right if I say U write a program to sort numbers there are so many different things that are implicit that when you write that program it should be it should take bounded memory it should not run for infinite amount of time correct it should run on a particular type of computer architecture and I did not specify that yet you got it right you understood what I meant right so part of sort of yes it's a very what Alpha code does is takes the specification and translates it especially the problems that you see in competition uh programming and does a great job in searching over this huge space of possible programs and tries to find the one which is consistent with the specification that is given at the uh in the contest right but if you think about the whole Act of programming that is a more General task yeah right where you have to somehow also think about what are the uh unsaid things in the program that the part of the specification that was not explicitly laid out but AI today is Al also very good at filling in that context right I mean it may not fill the context as well as a programmer who understands the problem space let's say I'm writing something for food delivery a food delivery app I'd know the local environment I'd say well you know if a driver stopped for 5 minutes then that's probably going to be a problem and the AI might not know enough of that to fill that in but I think it can fill in generic stuff because today if I go to Gemini and say you know give me a product spec for ABC it will do a pretty good job and it's the output is pretty long context and but I would say that the fact the act of prompting then becomes the act of programming in that case right where now you are sort of programming in a higher level language which is English right and the fact that it is an informal language you do not completely formalize and specify the exact semantics of what is it and yet the model is able to sort of work it out is something that you have to sort of learn to uh to use so you're saying the only thing that will go away is you don't need too much knowledge of the syntax anymore you don't need to be the best you know translator but you need to be able to formulate the problem well so this word the the prompt engineer sort of word then it becomes a real career then yeah it's like a prompt programmer like the prompt engineer is like prompt programming right but in some sense like we don't see it that way because we think programming is something very precise and prompting is something that is not precise like it's but in some sense like at the end of the day what you want is something precise yeah right and now regardless of whether you sort of use English to do it or whether you use uh a formal language like C++ to do it the end goal is the same right you want something which is precise which actually works for you I think that's this is the most nuanced answer I've got to this I've got for this so thank you so much for that you know I asked

### General World Models and Paths to AGI with Alpha Proof [20:57]

this question Demis on the round tables right on the media round tables and you've worked on many different types of problems the problem space of what deep mind works on is wide and you know there's a bunch of stuff that you've done in game worlds right game simulators stuff like that and there's also a lot of progress in robotics and there's been this crazy you know amount of chatter about General World models which is sort of like multimodal models but on steroids right and which is you know taking everything don't just take in language break it down into tokens and then try to predict the next word but Tak in everything Tak in sensory data if necessary Tak in everything do you feel like that might be I mean Dem is said on the round tables that you know that we probably need an offramp to llms to AGI right do you feel like that's the route to AI like what is your opinion on where that goes we need different we exhibit different types of like Demis has an advantage in being able to answer this question as because of his background as a neuroscientist I'm a simple computer scientist so I'll sort of give a simple answer I think uh my understanding of what how we do things we there are different types of um processes at play when we are um trying to solve intelligence tasks right we need sort of episodic memory we but we also need working memory yeah right and llms have been able to show that they are able to have amazing episodic memory right they can I mean they can learn yeah from a lot of data and then they can sort of generalize in sort of remarkable and very surprising ways but now there's an element of do they have a structured working memory do can they do can they accomplish long-term reasoning yeah right the both the slow and fast sort of thinking View and that is where some of the issues will be uh is there enough structure that is implicitly learned by these models that they can accomplish uh these kinds of task and can reason for a long time without making mistakes um so we we our belief is that there will be new advances that will be needed LMS will go along way but there will be sort of more advances needed to really accomplish everything that uh a human mind is today able to do any potential candidates anything you've seen that's excited you any New Paths so I think like just to give an example uh think of our model Alpha proof yeah so Alpha proof uh is um a system that we released we announced this year uh and which took on the maths problems from the international maths Olympiad and I don't know if you have seen these problems I have seen them and they have completely bowled me over like they are incredibly hard problems right um and they are challenging sort of problems and even uh some of the harder ones even field medalists would sort of find it difficult I mean they're not easy problems they will take some time to sort of solve them and what Alpha proof does it does tries it tries to find the proof of these problems by uh by doing search in the space in the very large space of proofs um by in a way that in the similar way to how alpha go did it right it sort of makes guesses about what could be possible proofs and then tries to explore them um and alpab proof sort of takes this even one step further so given a hard problem what it does it basically tries to solve variants of that problem this in a way where uh mathematicians even sort of work on it if you try if you are sort of asked to prove a statement about General numbers you will first try to sort of see what happens in the case of just one number or for the case of or for a restricted domain of even numbers or just integers before you try to sort of prove it for for everything right so it tries to solve these variants and in the process it learns something about the problem that allows us it to sort of really uh uncover some remarkable sort of proofs uh for these types of problems so you think something like an alpha proof is a much is a nice new candidate for path to intelligence yeah so I think it definitely has many advantages in terms of its ability to reason right uh because it uses it leverages uh sort of um a Model A large language model to sort of reason about what should be the right step but then also it has its formal elements the symbolic part which allows it to ground itself and explore things in a much more uh careful manner it has a piece of the Chain of Thought reasoning that we see with some other newer models today right where but some but here's the thing sometimes I open Chain of Thought like I open the reasoning list right because it's usually hidden on the other models and it's nonsense right sometimes it's like a breeze blue yeah then you know there's a table there's a plate on the table so it eventually gets to the right answer but you're like wait what is it doing to get here so uh but but very exciting that you think that that's a good path as well yeah so I think but in the case of alpha proof we are working with a formal language so the final sort of proofs that you get we know we can check them correct whether they are correct or not so unlike basically sort of uh Chain of Thought with informal language got it in the case of alpha proof you're using a formal language which is the lean programming language so you know whether it's right or wrong end of the day is doing Chain of Thought reasoning in code yeah that is insane that's a very smart idea yeah hey I want to talk

### SynthID: DeepMind's Watermarking Solution for AI Content [27:05]

about something else right and this is one of those parts of Deep Mind which uh at least one of the things you've produced which I think is it's a genius solution right uh which is syn ID let me see if I understood it correctly which is let's say somebody's generating text yeah you found out a way to Watermark that text right by saying look you know if you have a sentence and you have five words and then the sixth word is you know going to be generated there is a probability of the words that are going to be generated let's say the fifth word to be generated is dog MH uh not dog let's say if the fifth word to be generated is delve right uh just to give a better example instead of delve let's shift the you know the output a little bit let's use the word dig instead right and you do this you know in a bunch of places across the written text and therefore now you have a signature of slightly shifted probabilities of words and that's how you're watermarking it is my understanding of this correct yeah so I think syti is a watermarking solution what it allows us you to do is tag a generated content and then identify it and that's very important because if you see the the whole ecosystem of uh information some of it will be uh generated by AI you want to understand what was AI generated and what is not right and so synth leverages a number of different mechanisms to water Mark different modalities so for text it uses this sampling mechanism the one that you were sort of describing and that sampling mechanism essentially exploits the entropy in the llm because there is very little space so for the case of sort of images and video and audio these are high dimensional signals so you can throw pixel somewhere so you can sort of somehow inject certain sort of biases which are imperceptible to you and me got it right but yet would be persistent in the sense that if you take that signal you transform it some way we will still be able to detect that actually this was a generated signal even though you trans you changed it because we we are able to sort of extract that uh signature even after Transformations yeah because there's a lot of room I'm just thinking like a malicious actor yeah let's say I generate something with cnti yeah then I stop using Ai and I use something like a paraphrasing bot and there are a bunch of paraphrasing it Des so in the cas of text not because it has changed the word distribution uh so the text case is particularly hard yeah in the case of images and video and audio we are much more robust but the in text you have a relatively low dimensional signal and there is little room to actually inject a signature there right and so because too much and people would know exactly you will change the meaning or you will change it in a very sort of prominent way and therefore syn for text essentially exploits the entropy yeah of the distribution that the llm has so if you say the capital of France is now the entropy is zero because there's only one answer Paris so in fact you for this particular statement can't do anything yeah right it will be just one answer so we can't but if you said my favorite fruit is now it could be a mango papaya it could be sort of uh and it's much easier with words like Del right yeah but it's in the context right delve is basically it's much easier because you can replace it with many other words right and many other words are interchangeable so that's the entropy that we exploit to then incorporate a signature yeah very fascinating you know I don't know if

### AI-Generated Synthetic Content and Detection [30:53]

you've seen this right and I assume you haven't seen it so I'm going to show you once so I make content on Instagram that is short form content that is purely synthetic right I'm just going to show you an example for example this one right yeah this is a purely synthetic internet and someone just fixed app's award power button problem so the video I mean the lips are restitched right like it's it uses you know an avatar system for it the audio is completely synthetic it sounds like me it has intonation uh it sounds good and you know one question I have with syn ID would it be applicable to this because my biggest worry is look there are going to be a bunch of platforms where you know just upload a bunch of data by somebody or like you know an uninterrupted 3 minute clip and now the audio samples that you know some of these models required to just you know give you a nice output is like 3 seconds 10 seconds yeah so it's not very hard to make clones and if you see Twitter there'll be you know there'll be a bunch of you know uh people talking that sound like celebrities look like celebrities is there a mechanism put syn there yeah absolutely right so you can sort of so in syti we have one particular use case of syn is zero bit watermarking Yeah so in zero bit water marking we are essentially saying is this AI generated versus not yeah but there is another version of watermarking where you say I want to actually embed information in the signal yeah where the that where the person who is sort of consuming that signal then can understand oh this came from this was the origin this is the person who actually sort of developed it and this is the model that actually uh made it and so on right so give much more information to the consumer to be able to make an informed choice of what they are seeing that's very cool so I would say now when you click on images in Google maybe in the future we'd know very quickly what's Ai gened and what's not because slowly the internet is moving to much more AI like if you click on images on Google Now there just so much AI generated images right so it's crowding out the internet M very cool I think you know I've learned a lot I can ask you 500 other questions you know just it's super exciting having you here I have one last question right I went to the Deep Mind library today MH it was awesome MH what's your favorite book in the library so I mean there are many but across all disciplines it's like the polymaths candy store yeah it's an amazing place right I mean there's so many but uh um Charles darn is a personal sort of hero of mine yeah uh so the original species would be my favorite thank you so much pushit for doing this and yes you know I hope I get to see a little more of Deep Mind before I leave but yeah thank you yeah pleasure