I mean people often think that um this has sort of come out of nowhere but actually it's uh been a kind of steady March over 15 years of continuous Improvement year after year focusing on the same few objectives more data more compute larger training runs um and really the same core architecture this same idea that you can represent complex ideas with neural networks and they can be good for prediction has turned out to be true this has been proven out this is the kind of core thesis of Deep Mind and then um many other companies
now have been pursuing it for a long time I think the crazy thing is that it's really working now you know it it is genuinely a general purpose very effective way to predict pretty much anything in very complex environments and prediction is the primary skill of intelligence you know it's it's the main thing that makes us special as a species we're very good at guessing what might happen and then using that guess to creatively intervene whether we're manufacturing something physical or you know we're writing software or we're talking with one another or playing a sports
game it's a very fundamental skill so um yeah I mean it feels like a very creative and fun time to be using this new tool that everyone gets access to when you look back it's obvious what the the inflection points were the the first I mean you it depends how far back you want to go I mean alphao was clearly an inflection point I mean that that was although almost a decade ago now that was really the first large scale application of gpus so you know we hadn't really fully connected individual gpus in that way
before and it was one of the things that really gave Jensen an inspiration that he was on the right path I mean Jensen is incredibly Visionary and he has been betting on gpus since you know as early as 2012 um when he saw the cat neuron paper um in deep learning this was the first time that somebody at YouTube used a neural network to recognize cats in videos one of my colleagues actually Andrew ing um who's actually the founder of corsera um and an amazing teacher of machine learning you know that that was really an
indication that you need these very large parallel computers you know that can run you know thousands or in now trillions of operations per second um in parallel in order to search for the right combination of pixels that correspond to a cat versus a dog versus a zebra and then there's a series of breakthroughs like that along the way that gave everybody an Insight that actually this same core method is working and it's just scaling so it's often referred to as the scaling laws um how to kind of scale up more and more compute with more
data to improve our predictions and then I think most recently the things that blow me away are the quality of The Voice models you know can generate new audio in real time so it feels like you're having a conversation I mean it's just just a mind-blowing thing I mean it's sometimes you know you see a piece of technology sort of grow over a long period of time and even though you're very close to it you're still surprised when it works so seamlessly and it becomes second nature and then you take it for granted and then
everyone things is not interesting anymore you know like it's amazing that you have video streaming on your phones or you have audio recording you know that's just amazing to me you know we're in this amazing place in the history of science which is we have more breakthroughs than we know how to apply we we have a technological overhang it we've never had that in human history you know where where the same breakthrough is just producing more and more application and more products and more use cases um month after month week after week that's a great
example there's nothing novel about that it's it's just generate a script of two people talking to produce a podcast and then instead of training one conversational spoken model just add two and they take turns it's kind of amazing you know so it shows you that really we are just figuring out all the many ways that this technology can be profound and impactful uh it's very creative time definitely I mean there's going to be so many uh you know it's going to be an amazing teacher in you know it's going to give you a personalized curriculum
to help you learn whatever you know topic of interest you have whether you're young or old midlife or not it's going to be an amazing doctor in the future it's going to give you medical advice at near zero marginal cost right um it's going to give you medical coaching and wellness coaching help you adhere to you know stay uh you know to to stick to your regime if your doctor has given you a weight loss plan uh or you need mental health support and coaching um or you'd like to you know go to the gym
it's going to give you creative advice if you're trying to you know write an essay or produce a podcast or start a new business or get legal advice I mean the use cases are are endless and that's why it's a co-pilot because we see it as a kind of Aid as a consiliary as a you know as as something that is really empowering you to be your best self memory is the critical piece because today every time you go to your AI you know you have a new session and it has a little bit of
memory for what you talked about last time or maybe the time before but because it doesn't remember the session five times ago or 10 times ago it's quite a frustrating experience to people because you don't go and invest deeply and really share a lot and really uh you know look to build on what you've talked about previously because you know it's going to forget so you sort of tap out after a while and it turns you know into a shallower experience but we have prototypes that we've been working on that have near infinite memory and
so it just doesn't forget which is truly transformative I mean you talk about inflection points memory is clearly an inflection point because it means that it's worth you investing the time because everything that you say to it you're going to get back in a useful way in the future um you know you will be supported you will be advised it will take care of you know in time you know planning your day and organizing how you sort of live your life um so it's that capability alone which I expect to come online in 20125 is
is going to be truly transformative yeah it'll be expensive to begin with but as we have seen in almost every technology from the television to the camera you know once it's useful and you know some people can afford to you know give feedback to it and live with it and work with it then everybody wants it naturally and so when everybody wants it there's the you know economics of scale that really drive down the cost um you know today the very best models that we serve on Azure are now 99% cheaper than they were two
years ago really yeah and that's on Azure that's that's like that's like the best platform in the world for accessing these models right so I expect that cost curve to come down even more aggressively um in the next few years um and you know it it's always a kind of strange feedback loop because as the cost comes down of computation or of the algorithms then we can use more of it which means we can create even more complicated interesting powerful creative experience experiences so we we've been in this cycle for a century now almost of
computation where you know it gets cheaper but then we use more of it and then it gets cheaper and we get we use more of it and so we're in this kind of amazing feedback loop where so far more computation always delivers more value right it makes it more accurate or easier to control so it gets more precise you know you have Stu St istic influence and you can shape it to behave in a certain way you know and that I think is an amazing feedback loop at the moment um that we're kind of in
a sort of a beneficial cycle if you like not yet so I mean at the moment it's most reliable in sort of you know sort of English some of the European languages Spanish but in time there's no reason why the long tale of 120 languages in India or 5 and a half thousand dialects will not be uh um available like I think in the next 3 4 five years maximum because the same method has already proven itself to work in all those other languages so there's it's just a question of time and digitization and you
need less and less data once the main model is built you don't need the equivalent amount of data in the long taale of languages because it sort of bootstraps from the main the core language yeah so I don't think data is going to be as much of a limitation plus the more people talk to it and use it even though they don't explicitly give feedback they are implicitly giving feedback so if they end a session then that might be taken as a negative signal or it might be a positive signal or if they say no
that doesn't make sense or that's wrong or you know we learn from that feedback so that's why the more people use it the better it gets I mean co-pilot is growing like crazy in India I mean it's a huge market for us um which you know we're very excited about but partly it's because amazing knowledge workers lots of developers um you know obviously people who are by bilingual by default or maybe multilingual um and so you know they've been very eager to take up you know new technologies for a long time very open-minded very creative
so yeah I think that's why the um it's one of our fastest growing markets some of my strongest teams are in um this city same with hydrabad u we have very large ads engineering teams we have large research teams some of the biggest impacts that we've had in improving uh quality of search results for Bing come from here um so you know this is why I make the time to come visit here because it's critical Development Center for us so think of it as a conversational companion that you can ask any question anytime in your
own style at your own pace that has no judgment never gets frustrated never gets angry never gets bored never gets tired and is always available to reframe any question or problem you're thinking about to to your level whatever you feel like at that moment so it's completely responsive to what you need definitely I mean we have a huge co-pilot for work effort um M365 co-pilot and you know it helps you redraft emails helps organize your calendar helps you you know write creative documents in Microsoft Word or use Excel in time it's going to be a
troubleshooter so you're at your Windows machine and you get stuck with a setting maybe you're trying to change your Bluetooth or you're trying to fix a bug in your program or something it's going to be working with you at the same time saying um no like look over here open that close that window or you know it's going to see what you see in real time on your desktop on your mobile phone and that's a key transition moment because today you know you have to explicitly communicate to your computer in words or in buttons whereas
tomorrow you're just going to make a reference you're going to say this or that or look there or what color is it what do you think of that like it's going to be like having a companion by your side and that's a different plane of communication to what we're used to it's like an each new wave of Technology opens up a new plane of communication you know radio opened up one to many broadcast with audio TV opened up more than one but you know thousands or maybe hundreds of thousands to many with all the different
channels over time and then the internet opened up all to all communication and so each mode of Technology enables you to think and act and talk in a completely different way than what you've done previously and so that kind of draws out a different part of you you know you're you're going to be able to express ideas that you didn't even really know you had because there's a new modality at your fingertips and you know having an ever present companion that sees what you see and hears what you hear in real time is going to
feel profoundly different just as the touchcreen you know every time everyone used the touchscreen for the first time it's a profound difference the fact that you can swipe around and tap things and drag things you know some you know kids now you know you see even I do it every now and then you go to a magazine and I try to pinch and zoom the magazine I'm like you know scrub the map over and he's like it doesn't move so you you you learn a new subconscious way of trying to shape your digital world because
you have a new tool at your disposal and that's I think going to be a big big shift to always have that companion at your disposal to ask a question or to be creative or to be curious deeply empathetic yeah I think more so than humans I think so I think so yeah I mean it's conversationally fluent it's very relaxed it can handle interruptions it can be patient when it needs to let you speak it'll give you uh active listening so it'll be like uhhuh yeah uh-huh cool you know we do exactly yeah exactly so
it's it's uh it's a different way of talking to computers which is I think a huge shift I mean three years ago it had no reasoning whatsoever you just ask a question and it gives you the best answer it can today we now have models where before it gives you the answer it gives many many answers to itself in its own head in its own working memory and then it tries to rank those answers and say okay is this what the user is looking for in this moment will it make them bored is it answering
their question is it entertaining is it accurate okay yes all right produce it no okay okay let me produce another one that is more like this so it can do these things called Chain of Thought it can think in its own head um which is a form of reasoning and it can do that many many many times like up to 100 times um obviously super fast time as well so it's it's doing that in a second or you know a couple seconds before it produces an answer it allows the model to reflect on its own
answer and evaluate the quality of its answer and iterate that answer before producing the result you know so it it is like reasoning you think through a problem step by step um yeah it's quite different to how things have gone before I think that the important thing is to be open-minded and embrace the possibility um and be multidisciplinary because the the future uh high impact individuals are going to be able to speak a range of different Technical and social languages and operate at the intersection of disciplines specialization is still important but you have to have
specialization and breadth um because we now live in an all toall information exchange World um you know you see many young creators online today influencers and people in social media they're good at coding they're great at using new tools they're creative they can use all of the you know design tools they can make short films they're making music everyone is becoming multidisciplinary because they're learning from what everybody else is doing almost in real time so it's almost like the collective intelligence of our human species is is is taking off it's getting more and more capable
because we all learn from each other in this huge sort of morass of collective intelligence it's it's quite remarkable and that's just from all to all connectivity and reducing the barrier to entry to access tools that make us smarter
