Agents, Humanist Superintelligence, and Healthcare

Auto-generated from YouTube captions and lightly formatted for readability. It may contain transcription errors. There's a lot of talk about agents right now as well. In fact, you know, Microsoft Windows just now got Open Collage is sort of direct integration. What can they do with agents that we couldn't do with just a regular chatbot before? You can use those to organize your work on a day-to-day basis, track your health, run background tasks for you, find you the best price of things that you care about, things that you would previously write code for, obviously an agent is now going to help you do. Lots of my friends have been using it to sort of open their gates at home, control their HVAC systems. Someone was telling me the other day that um he asked his agent to help him, you know, lose weight. And so it's going to be easier than ever before to be a small developer with a few people and a bunch of agents building highquality stuff. The only test that we can put technology to is a very simple one. Does it make our collective lives healthier and happier? Does it enhance human progress? This is something that would taken two or three people to do pretty much full time. That is kind of a little superpower. People are going to be very excited that they're going to have a perfect health assistant in their pocket. I I really think that we're close to medical super intelligence. Mustafa Solomon is one of the most influential people shaping the future of AI. He co-founded DeepMind and now he leads Microsoft AI overseeing co-pilot and Microsoft's rapidly growing family of AI models. In this conversation, we talk about agents that can work for days. AI moving onto your devices, Microsoft's push for human- centered super intelligence, and why he believes worldclass AI healthcare can be available to everyone within just a few years. And we talk about so much more. So, please enjoy this wide-ranging and fascinating conversation with Mustafa Solleman. So, I'm excited to do this conversation again with you. This is uh the second time we've spoken. We spoke at the 50th anniversary and so it'll be kind of uh exciting to see how things have sort of evolved since we've had that conversation. Yeah, it's been a crazy year. It's wild wild times in the industry. Yeah. So, uh there was just a whole bunch of announcements. You know, the Microsoft Build event, the main keynote just happened and it was just announcement after announcement after announcement. Um I'm curious, what do you think is the what are the announcements you think have the most impact on the most amount of people? Yeah. I mean I think our main motivation was to try to land the idea that we as a platform company are building a hill climbing machine and that means at every layer of the stack from our own silicon to our orchestrators our harnesses and of course our own models we're trying to give developers all the flexibility they need to pick whatever they want so that they feel like they can build to the absolute frontier. Um, and obviously we have our own internal, you know, sort of Microsoft offering now between the Maya 200 chip and the MAI models that we're very proud of, but we're also partnering across the board with everyone you can imagine to make sure that developers feel like they've got the best ecosystem with us. Mhm. Can you explain the hill climbing machine a little bit more? What is what does that mean? So in in AI, you know, once you have a clear evaluation and you know exactly what your prediction target is, you want to make sure that the data that you curate is tightly coupled to the evaluation that drives it. And the harness that sits around an agent that tries to, you know, improve itself with respect to that data towards that evaluation objective is tightly coupled again to the full kind of cloud environment that you use. And that's what we mean by the hill climbing machine. you can hill climb once you have those four components of the stack tightly integrated into an RLE, a reinforcement learning environment that ultimately produces an agent that is customized to your objective that you fundamentally control. Got it. And it's so it's something that sort of gets smarter and smarter as it learns more about your business and what you're doing and the types of things you use it for. Yeah. And and that's kind of the evolution that we've been on in the industry. like the last three or four years have been about producing pre-trained models that have a kind of base level information about the world. They're you know generally understand all the web data and a good amount of the books and PDFs and general purpose knowledge. Right. The next phase has then been about like adapting those models to specific tasks using reinforcement learning and that's where they kind of hill climb towards a specific objective over time. I see. I see. So there's a lot of talk about agents right now as well. In fact, you know, Microsoft Windows just now got OpenC Claw sort of direct integration. What what now opens up for normal people now that we use agents? Like what what can they do with agents that we couldn't do with just a regular chatbot before? Well, increasingly you're going to be able to step out of the CLI and just use your agent as an app. I mean, you saw the companion app, right, for Open Claw, but for many other agents that'll be in there, too. And you can use those to organize your work on a day-to-day basis, you know, track your health, run background tasks for you, find you the best price of things that you care about, you know, I mean, anything that you could you would think to give to an assistant or things that you would previously write code for, obviously an agent is now going to help you do. Have you seen any like killer use cases in your own sort of experimentation? Lots of my friends have been using it to sort of open their gates at home, control their like uh HVAC systems, set up automatic timers for things. Someone was telling me the other day that um he asked his agent to help him um you know lose weight and so he has like cameras inside of his kitchen and so it was sort of letting him know that he was heading back to the kitchen for a snack a third time in you know 2 hours or something. So look, I mean there's something out there for everybody. I'm not sure that would be my choice, but like, you know, it's pretty cool to see the flexibility. Gotcha. Yeah. Yeah. So, he's going and opening his fridge and maybe he's getting a notification on his watch. Are you sure you want to be doing that right now? Right. Exactly. Very cool. Uh you guys also talked about uh web IQ. Can you quickly explain what WebQ is and what that opens up for people? I mean, obviously right now we have work IQ, which means that you know, for big enterprises, you know, your model should be grounded on the knowledge that you've already got in your own workflows. Web IQ kind of provides the flip side of that, which is that we want to have super fast, very efficient access to web content so that you could point your agent at any grounded external knowledge base and it should just immediately be able to provide good citations and accurate retrieved information from the external world. Gotcha. And does that when you say the external world, does that also include the sort of you know Microsoft suite of products? So it's like pulling from Excel and PowerPoint and Word and all those sort of things as well. Yeah, from your personal one drive of course. So be your corpus of documents, right? Um but also accessing the external web. So that'll just be regular real-time information from the web. Gotcha. Okay, cool. And then also um there was talk about project Solera. That was something that I hadn't previously heard about. I don't even think that was in part of the the brief that I was on earlier on. Um can you explain that a little bit what that sort of opens up cuz it kind of feels like it's putting agents sort of all over the place. So, you know, your like your kitchen example, this is sort of putting agents like maybe even in your kitchen and uh you had the the little uh badge thing as well. Can you explain what that is a little bit and what you see that opening up for people? I mean, you know, it's pretty incredible. The badge has been around for like 30 years or more and basically hasn't evolved at all and yet, you know, most of us who go to work every day in a big company have to carry around a badge to open the doors and so on. So, I think it's really exciting that there could be an everpresent, you know, aware kind of sensor system on your badge that basically makes it easier for you to ask any question of your everpresent AI um with obviously camera and vision. So, it's pretty cool and I like I like the kind of bold experimental nature of it. I'm very excited to to try and use it and it's a very flexible platform so anybody can like you know ship their own models, you know, ship their own harnesses, integrate it in their own products. So, it's really supposed to be a general purpose platform. Right. Now, the some of the examples you guys showed during the keynote where like in a doctor's office, which makes a lot of sense. That example makes a lot of sense. But do you see um examples of just like normal everyday like consumers using it around their house or is it mostly designed for business use cases? I think it's an open platform, so who knows what people are going to do with it. But the first I think hypothesis was you know there is this device that people carry around on their belt every day that is actually pretty important. I mean it's secure you know it it grants you access you know in a pretty sensitive way and so a lot of the basic infrastructure sort of already been built. Um and so the goal was basically to build on top of that and you know see what you know developers and different companies choose to do with it. Gotcha. Now I guess I guess my follow-up question around that is you know we all have phones in our pocket. But what what can we do with that that we can't do with our phone? Well, I I think it's ease of access, fast, stable, reliable. Obviously, it's secure. I mean, technically speaking, you could also use your phone to unlock a door anyway, right? So, I mean, I I I think that many of these deployments are going to be um somewhat overlapping because clearly all your devices, you know, do some portion of all the capabilities of other devices. I think that the the main motivation is it would just be great to add a few more features and and create an open platform um you know wi with this project that people can experiment with. Yeah. Yeah, that makes sense. Uh so you guys also mentioned seven new models, right? You've got uh you've got a new thinking model, you've got a new image model, you've got a new voice model. Um you know, seven different new models. What what can people do today now that these models are available that they couldn't do yesterday? Yeah. So I mean for us this is about taking first steps towards true self-sufficiency in AI and that means that we have to have the capacity to build our own models from scratch and show that we can achieve the absolute frontier. So for example in transcription it's the very best model in the world. It's also the fastest and the cheapest. Um our image to image and image editing models are now number two and number three in the world. Um you know beating out everybody other than OpenAI at the top for GT image 2. um and you know a couple of Google models for um editing but we're still better than Nano Banana Pro Nano Banana 2 right so it's a pretty big deal and I think that what we're trying to show is that taking steps towards building this hill climbing machine shows that we can actually create models that are of the absolute state-of-the-art so that as we models diverge in different directions you know we can actually build exactly what's needed for Microsoft and our developers so one example of that is um our code models are actually very very small and very very strong in highly inference efficient general purpose agentic use cases um and we've chosen that use case in our pre-train model in our RL climb and in our post-training um and that means that they're optimized for agentic coding and for the enterprise which is different to Google and say open AAI that also have to worry about kind of general purpose consumers right so I think over the next few years you're going to start to see more and more divergence in the lineages of these models. Um because clearly general purpose will deliver, which means that one model that can deliver across all the domains, but highly specialized models are always going to be able to push that a step further. And you saw that in the kind of work that we did on Excel and with Mckenzie and with a bunch of other companies showing that we can drive really outsized performance for 10x lower cost with very efficient in-house models. Right. Right. Now you I think you uh I think this was the wording you used but you mentioned it's all commercially licensed lineage. Yeah. Um what is what does that mean? And you know with it being sort of commercially licensed lineage does that have any sort of negative impact on the output of the model you know compared to to companies that might just use whatever they want in the training. I mean look I think part of the challenge is that we don't always know exactly what's in other people's pre-trained sets. Today we've also released 109page technical paper sharing in great detail exactly what has been used across the stack algorithms training infrastructure compute GPUs data content data composition. We've shared actually a lot on the kind of model design side of things as well. So we tried to be as transparent as possible knowing that you know our main job is to build trust so that people come build on our platform and that also extends to how we buy the data like we have paid a great deal for this data. We've licensed it very carefully. We've been very deliberate about not using open source data sets, right? You know, and we don't know what security bugs could be introduced through, you know, some of the more open source data sets that maybe haven't gone through the same kind of rigor. Whereas for our customers, you know, we we have some of the largest governments in the world on our platform. We have many of the largest companies in the world. So, we want to give them trust and confidence that the model that we give them is absolutely clean from top to bottom and is, you know, been, you know, acquired in exactly the right way. Yeah. Amazing. Um, so I'm curious about the the the sort of bigger road map for uh for your various platforms, right? You've got C-pilot, you've got the new um GitHub app that you guys showed off today, which is really really cool. I'm actually really excited to go play with that one. But, you know, you obviously have deals with OpenAI, you guys have deals with Anthropic, you guys are kind of working with with all these companies. So, I'm curious what the the sort of bigger picture is here, developing your own models. Is it sort of to um you know be the new state-of-the-art like to try to pass them like what what is the roadmap for all the models and you guys working with the other companies as well? I mean we are a global company that provides for or small developers, enterprise developers, governments and every kind of institution you could think of around the world. So as a platform company our job is to provide optionality. If you come to Foundry I think there's something like 11,000 models that are available. Mhm. Um you also get the absolute frontier so the best models in the world from anthropic and open AAI um and you also get the option to use our new in-house models which are you know growing and getting better and better. So the thesis is we shouldn't restrict anybody or try and you know lock anybody in or reduce their optionality. People want to make different choices for different types of experimentation in different settings. Um and that that's really the kind of core mission is optionality first. Right? So people might find that the open AI model it does really well for this specific use case for them but maybe the MAI the you know the state-of-the-art MAI model is better for this use case. So they could sort of switch between them based on the use case that these models are best at. Yeah. And each of the models have different trade-offs right and they have different inference profiles. Um they have different you know cost you know benefits or sometimes disadvantages. Some of them are very very expensive. And so I think as tokens become the sort of primary building block of new products, people are going to make very different choices about where to use frontier models, where to use smaller, more efficient models and making sure that they're kind of multimodel under the stack under the hood so that in the stack you can sort of swap between them. I think it's going to be a big part of it. Right. Right. But I also think like a really good use case as well is if like one model isn't working, you can have sort of fallback models where you know you tell it to do something with maybe an agent and if it's not getting it done with this model, the agent can say, "Okay, let's go try it with this model as well." That's right. And I think that you'll see constellations of models deployed into products rather than single models. I mean, this is definitely going to be a multimodel ecosystem over many years. Right. Right. There's also a lot of talk about um unmetered intelligence, which I I really love that that sort of metaphor analogy or whatever you want to call it. Um how close do you think we are to, you know, the current state-of-the-art models that we're seeing right now from, you know, Microsoft OpenAI, an Enthropic? How far off are we from those actually running like on device models of that level running on device? We're pretty close. um you know so we're going to be shipping our MAI models into the N uh 1x um you know later in the year um and I I think that the the general observation is that we have an enormous amount of kind of spare compute on people's local devices right um and those local GPUs or AI accelerators are getting more and more powerful so we should just try and use those as much as possible um so I think you'll see more and more workloads pushed to the edge over the next few years yeah do do you think the shift is these these devices is getting more and more compute on them so that they can run bigger and bigger models or do you think we're going to move in a direction where the we're going to get smaller and smaller models that run on less you know VRAM but are just as smart like which which direction do you think is the path it's going to end up taking? I think that, you know, as is often the case, you see both concurrently. So, you know, the largest companies are are going to continue to build absolutely enormous models on probably gigawatt scale, you know, data centers and and training runs that last for many, many months. And then you're also going to see hyperefficient small models that really do operate even on wearable devices on your, you know, local sort of PC at home, right? So, you'll see both concurrently, I think. Um, so I'm I'm curious how this affects the the various frontier labs, you know, Microsoft Anthropic Open companies like this that are going and doing these big uh, you know, data center buildouts. If if more and more of this is moving on device, isn't that sort of counterintuitive to why they would build the data centers? Like how or do you think these data centers are still going to be necessary as we grow? I mean, we've got two different types of workloads, right? So there's one workload that is training that requires these mega data centers that have like very very contiguous compute. So all of the nodes are like completely connected with the fastest possible connection that we can get. And then when we do inference, you know, which is obviously training your agent or running your agent so that it can run these asynchronous tasks in the background. That can happen on device. it can happen locally and maybe occasionally it calls into the cloud for a really like complicated or hard maybe PhD grade question that it doesn't have the local power to kind of resolve. Now at the moment the smaller models don't really know when they don't know but that's like a kind of big you know objective of training you you don't want like you know an absolutely frontier grade model telling you the answer to what the capital of France is you you want it doing a very expensive like background asynchronous complex task that takes multiple you know sort of rounds of of action right and we're also kind of seeing this um you know we're seeing this trend of the of of people wanting to let their models run longer as well, right? We're kind of seeing people set sort of set goals with their models and say run as long as it takes to meet this goal. And that might even be, you know, days sometimes, which you probably don't really want that running on your local computer if you actually want to be able to use your local computer, I'd imagine. Yeah, it's a good point. I mean, I think what we we're starting to see is people have separate machines for their local agent just to drive those asynchronous background tasks. And that's actually a micro version of what we see in the data center. So when we train for example MI thinking one I mean that RL run has been going for 13 or 14 weeks like maybe longer. Um and that's basically because intelligence is now a function of compute. So the longer you run this thing the more access to data you give it the more RL environments you provide for it the more it's going to explore this complicated path and find like novel information and knowledge in in in that world. Right. Right. Um, I want to I want to stick on the topic of uh data centers for a second here. Um, because obviously there's been, you know, in the media and in the the world there's been some push back around the data centers and and the the buildout of them, but it does sound like Microsoft specifically is is doing things to sort of mitigate a lot of the worries that people have around data centers. Can you talk into like some of the misconceptions people have about data centers and and what you guys are doing differently to sort of offset a lot of the fears that people have around them? Look, I I think the first thing is we've been very committed to our net zero target which we remain committed to which means that you know the vast majority of our data centers rely on renewable energy. Uh the second thing is that um obviously these data centers consume a lot of water but once the you know the core system is full up it's a closed loop liquid cooled system in our fair water data centers. So I think that needs to be changed like every five or six years and it's the equivalent of using the same amount of water that is used um you know by a regular restaurant you know in in a very short period of time. So it isn't as much as people were worried about and I I think it's good that people are concerned about these things because these are massive infrastructure investments and and really do change the landscape for people and so on. So I think the concern is justified, but I think, you know, it's also just important to recognize that we really are pushing as hard as we can to do the right thing. Now, as far as like the energy usage, right, that's that's the other sort of talking point, right? Is that um if the data centers are using all this energy, it's going to cause the local communities energy bills to go up, but I think you guys are sort of trying to offset that as well, right? Yeah. I mean, we we've made the commitment that we're not going to let the local energy bills go up and we'll offset. So people shouldn't see uh any any spike in their bills. Gotcha. Just sticking with what we've been sort of seeing in the media a little bit and talking into it a little bit. I'm sure you've seen the like commencement speeches where they're getting up on stage and every time they mention AI, the the audience isn't really having it. What do you think needs to happen for just the, you know, the general public and the the college students and the people coming up in the world today to sort of trust the AI and trust the companies that are putting the AI out there? Our hope and belief and core motivation is that we can create a humanist super intelligence. There are lots of people out there in the industry who uh are publicly saying that their motivation for building super intelligence is to create something that exceeds the intelligence and capability of all humans combined and then can keep improving you know for many many years to come way beyond our intelligence. And I think what scares people is when you say that it's hard to imagine how we would control something like that. So not only might it compete with us for resources, but if it were autonomous, it might be pretty hard for us to kind of keep it contained. Yeah. And this is something I have personally been thinking about for a very long time since I started Deep Mind and have worked for many years on AI safety and ethics. And I think that the only test that we can put technology to is a very simple one. like does it make our collective lives healthier and happier? Right? Um and that's why we invent things. That's the point of science and innovation. That's why we're excited to, you know, have new drugs that, you know, cure diseases that, you know, relieve our pain, that help us to live longer, right? And I think that that is just a very simple test of technology that we should put, you know, super intelligence to next. Does it enhance human progress? Does it make everybody healthier and happier? Mhm. And that I think is quite quite clear like we we've got to subject it to that test. Yeah. So so somebody that's um you know maybe about to go into college do do you have any advice of like what what would you tell them to study right now moving into the world that we're moving into? I think you still have to study software engineering just as you know people have been focused on for the last 20 years. Um I think that more than ever you have to study science and and maths. Um, but you know, we are also going to have to think hard about the philosophy and the ethics behind this because we're creating something that is incredibly powerful. It will feel very humanlike. It's going to be very dynamic. It's going to be integrated into our lives in ways that we've not really seen from any other technology before. And so thinking about the governance of these things and how they play into our politics and our civil service and our public duty to create the world that we want to live in, I think is going to be really really important. And maybe it's not so much the primary muscle that we've exercised in the last 30 or 40 years, but it's going to become really the main focus, I think, of people in the next 5 to 10 years. What What do you think um what do you think happens to this software and and SAS world um with AI cuz you know there's obviously all this talk about the SAS apocalypse and all that kind of stuff where people can just buy code their own projects into existence. So why would I go work with software companies? Um you know, what do you think the future of software looks like? Yeah, I think markets are very fickle and a bit hyperreactive. So, there's sort of these bouts of anxiety and then somehow things like rebound and so on. Um, it's definitely going to be easier to produce highquality stuff. And, you know, that has got to be good for everybody because if there are more people who can competitively produce more highquality solutions to real problems that humans have, then that is going to improve the quality of those solutions, right? It's going to be faster to take things to market. It's going to be easier to test hypotheses. It's going to be easier to explore a whole bunch of different, you know, product market fits. And I think that's going to be great for consumers. And it's actually going to be great for businesses, too, because we're just going to get better quality tools to run our organizations and our lives. And so, it's not clear to me that that results in the collapse of anything at all. If if anything, it's like going to be a flourishing of new invention, new discovery, you know, faster, more efficient things. And that's what we've seen for hundreds of years. You know, intelligence has produced all of these amazing breakthroughs and inventions. So, there's no reason to think that it doesn't continue to do that just because it's more of a synthetic intelligence rather than a biological one. Right. Right. I I sort of have this theory that instead of, you know, seeing more and more like of these mega companies, uh we'll see a whole bunch more sort of smaller pods of like little startups that are three or four people, but building stuff that's the equivalent of like much much bigger companies just with much smaller teams doing it. So there'll actually be a lot of job opportunities just in smaller pods as opposed to like the big corporations. I feel like we might kind of shift into that. Yeah, it's going to be easier than ever before to be a small developer with, you know, a few people and a bunch of agents building highquality stuff. And you know, maybe that's a good thing, right? That puts competitive pressure on some of the, you know, companies that maybe are fatter than they should be. And, you know, we'll have to readjust some of the roles and jobs and that'll look slightly different. Mhm. So, I just I just have a couple more questions. Um, I'm curious what you personally use AI in your own day-to-day life for. Like, what's what's your killer use case for AI right now? Um, I I use it a lot for tracking my health. So, um, I've connected Copilot Health up to bunch of my wearables, given it my EHR record. Um, you know, it keeps me on track with a few health conditions, long-term conditions that I've been managing, and I really like that. And I do use it for kind of coaching and guidance. Um, like I have pretty long conversations periodically. Um, sometimes they start from just a random question about something. Um, I like history a lot. I'm also actually a keen gardener, so I'm constantly researching plants and stuff and so yeah, they're probably my big use cases. Is there anything that uh that you know an AI model has done that's really surprised you? Last time we chatted, you told me a story about how uh you were headed to the airport and AI alerted you that your flight was late before the website even alerts you that the flight was late. that sort of uh sort of blew your mind. But I'm I'm curious if there's been anything recently where AI's just like really wowed you. Yeah, that's that's true. Actually, I forgot about that one. What actually happened there is that I I had asked the um check-in front desk if the flight was delayed. Um and the woman said no. And then literally a minute later, she came over to me and said, "Yes, it's delayed. How did you know that?" Which is really cool. Uh in Seattle, actually. Um yeah, I mean I think that I'm most excited about the healthcare applications. I I really think this is going to be transformative and and I'm I'm very confident. I mean we have seen this over and over again where you know if you if you just have enough really highquality data that hasn't been seen by the model before, right? Then you know you can produce magical leaps forward in diagnostics, in treatment, in preventative care. And um just earlier today we announced this partnership with the Mayo Clinic um and we had the president and CEO Jeano Fujer on stage with me and we're basically forming this multi-year partnership to train a health foundation model from scratch together um and deploy it in their hospitals and around the world. And so I I'm very excited about that. Very cool. Well, this is my last question and it was it's the same last question I asked you last time and I feel like you kind of already answered it, but let let's see if you have a different answer. Um, what's the thing that excites you most about what you can do with AI right now? And what excites you most about what we'll be able to do in the very near future? I mean, I think I think right now the thing that I'm excited about is that I can give my agents a bunch of very complicated long-running tasks and it can synthesize information for me in ways that I never believed were possible before. So, one of the things my agents do at the moment is every morning I get a pretty long and detailed briefing of all the activity that's been happening inside of Teams on email with all the docs that have been changed. And instead of me having to go and read all of those things, I just get this clean synthesized summary with links to everything I need to do, recommended actions, a prioritization for my day, briefings before each one of my meet. I mean, it's kind of remarkable, right? Right. I mean, this is something that would taken two or three people to do pretty much full time. Um, and that that there's kind of a little superpower. I I just feel like prepared for the day, clearheaded, and, you know, just armed with information, which I like. Yeah. Yeah. Yeah. I mean, I feel like the the next layer of that is almost like you put an AirPod in and it just sort of like an assistant just reads out what your day looks like for you. So, so we actually have made a a a voice model for for GitHub that basically, you know, just runs in the background, organizes all of your PRs and commits, reviews work that other people have done, proposes new PRs, and literally in voice, it's giving you a summary. So, you don't have to read the full PR, right? You can actually just get a synthesized statement, and then obviously it'll ask permissions and interact with you. So I I I think that the that the the the the headphone is going to become genuinely another surface in addition to the phone and obviously to your desktop. Right. Right. Um because you'll just be able to quickly check in which I think is going to be a very easy way to give instructions to your agent. Right. Right. And then the second part of that question, what's something in the near future you think uh we'll we'll be seeing that everybody's going to be excited by? I I think that people are going to be very excited that they're going to have a perfect health assistant in their pocket. Yeah. I I really think that we're close to medical super intelligence. Awesome. I think two to three years and it'll be possible to get access to the absolute best health care in the world which currently you can only get if you go to Mayo, right? And that's just going to be, you know, abundant and available to everybody. Amazing. Well, that's all the questions I have. I really really appreciate you taking the time again. It's it's always fascinating talking to you and I really appreciate you taking the time. Yeah. Thanks, man. It was really fun. Cheers.