We Need An Ecosystem in AI, And Every Company Can Win A Place In It
00:00:00The world is going to be very skepticalof tech and tech companies that saytrust us, we've got it, the future isgoing to be glorious. You kind of haveto deliver tangible benefits becauseit's too important this time around.It's too much of the economy for it notto be the case. True ambition is aboutmaking the impossible possible. I takegreat inspiration from sort of thepeople who are managing the Azurenetwork. We built in the last 15 monthsmore Azure capacity than we built in thefirst 15 years. I mean, it's crazy.wild. Our job is not to do Azurenetworking. Our job is to build theagentic system that does Azurenetworking. Right, the way to get toinformation, way to educate yourself,way to continuously keep yourselfupdated has changed so much. Maybe thenext big startup could be someone whobuilds a new university, a new pedagogyeven of how to get someone to go througha curriculum and find economicopportunity that's highly valuable.
00:01:11>> Please welcome Swyx, Saragosa, AlodGill, and Chairman and Chief ExecutiveOfficer of Microsoft, Satya Nadella.>> Hello.>> What's up?>> Uh I'm so excited to be here. Welcome toa crossover episode of No Priors in LaneSpace with Satya Nadella. Umcongratulations on an amazing build.>> No, thank you so much and it's great tobe with both of you. I listen to both ofyou or both the podcast all the time.It's great to be on it.>> Thank you so much.>> So, you were talking about um theseamazing uh announcements from across theMicrosoft estate all morning for I think3 hours. What is the uh what's the mostimportant reflection or takeaway youhave?>> I I'd say there areuh perhaps the the biggest one for me islet's sort of conceptualize this more asan ecosystem play as opposed to a singlemodel or even a single platform, right?I meanyeah, whatever I at least for me havinggrown up at Microsoft, having seenwhatever four major platform shifts uh Isort of fall into that uhuh camp where
00:02:26a platform is defined byfundamentally its ability to create morevalue about the platform versus what'scaptured in the platform. And so if youyou view what's happening right now, Ithink this morning's keynote washow can any companywhether it's an AI-native company or atraditional enterprise companyparticipate as a first-class participantwhere they can point to AI they create.Right? It's not that they don't useother people's AI. Of course, they will.But to me, what's the path? What's therecipe? How do I do it? What does thestack look like? What does the toolinglook like? What is valuable? How do youdo that? That's it. That's sort of ourjob to do.Yeah.>> Ecosystem strategy is uh verycomplicated, right? Because you end upbuilding certain components, partneringfor certain components, supporting them.You just announced this big suite ofmodels. Like tell us a little bit aboutthe uh training strategy for Microsoft.>> Yeah, soso the thing that we wanted to do withthe MAI models was to build and asMustafa talked about first of all, agreat lineage, right? Starting withpre-traininguh with very good data quality,
00:03:43uh doing all the ablations, making surebecause in in some sense it's becomeeven harder to build a clean lineagemodel because there's so much stuff outthereuh that you truly need to ablate out tobe able to have a fantastic pre-trainedmodel. In fact, that's one of thechallenges of a lot of the open weightmodels is they look great on onebenchmark or two, but they're not greaton practice. So, that's why in fact,even in the RFDs arepretty gone really excited about theseMAI models because how the heck can asmall 5B model hill climb and it goesback a little bit to what I think isultimately the key thing to do, which istry to pursue finding that cognitivecore. So, to me starting with a cleanlineage, then creating that ability forcompanies to be able to use this, right?Not just as a generalist, but to createtheir own specialist by building thishill climbing scaffold around it, right?So, it's not just the model, but youhave a hill climbing scaffold around it,thenyou will start building your RLE. Youwill start collecting the traces. Mostimportantly, you'll have private evalbecause we know all the evals out thereare good,
00:04:58interesting, but they're notreally that critical at this pointbecause they all can be maxed. And so,the point is each company will have itsown private eval. And so,that end-to-end platform storyaround our models is sort of what Ithink is interesting. And then the oneother thing, Sarah, since you broughtthat up isI do feel there's a new frontier. Likepeople talk about the frontier andyou're operating at the frontier.Interestingly enough, if you add alittle temporality to it,you can use, let's say, in in in fact,that the Lando Lakes demo we showed waspretty cool. We used whatever GPT-55,right? Then you collected a bunch oftraces, and then you took a 5B reasoningmodel and achieved higher. So, that isanother aspect of what it means toappear I'm know, operate at thefrontier.>> Yeah. I I think uh I first of all haveto congratulate you on basicallybuilding a frontier neural lab inside ofMicrosoft in two years. Um I'mwondering, you know, you have all thisAI strategy that you're rolling out. I'mrunning what do you know now that youwish you would tell yourself two yearsago with two three years ago. Threeyears for the Jensen partnership, twoyears for uh MAI.>> Yeah, I mean, I think the the thing whenthat I reflect quite a bit, right, whichis sort of obviously I got
00:06:13into all thiswhen I got excited by the the scalinglaws paper and you know, when you know,even the OpenAI partnership came aboutwhen those folks said, "Hey, we're goingto really throw a lot of computertransformers." Uh and they've helped,right? The thing that I always look backand say, "Wow, these thingsum do have capability that they'reclimbing up with I mean, this you know,this crude way of saying it isintelligence is log of compute." Kind ofworks.Now, what I think we underestimatedperhaps is the real-world complexityof deploying these so that they actuallydeliver the value in the real world,right? So, the outcomes as measured byany benchmarkis interesting, important, but the trueeval is when people out there are ableto do unique things that they only canvalue.And it's very measurable.Right? That I wish we had sort of evenlike had more in our consciousness,right? Which is as an industry becauseright now I thinkwhen people say, "Wow, I don't want atoken max." It's an artifact of us nothaving thought ourselves as an industrythat we are using tokens to create valueevery step of the way.
00:07:28So, I thinkthat's kind of what I wish we had gottenthere, but I'm glad we are here.>> What are some other use cases thatyou've seen that have created the mostvalue for your customers? Because I knowthat people talk a lot about code and Ithink it's pretty clear that that'ssomething that's having very large-scaleimpact. Are there other areas that youfind in common that your customers arereally benefiting>> Yeah, I think to your point, obviouslycoding is now God, but it's interestingby the way you love to even talk aboutthe coding, right? Which is coding iswork so well that we now have to rebuildthe IDE, right? I mean, it's kind ofnuts to see what we saw large is like,"Oh my God, I have these hundred agentsessions. I the cognitive load ittransfers back to me as a human is soexcessive that now I need a new UI."Oh, by the way, like the the chat as theonly artifact is also impossible. So,that's why we need a canvas. So, it'skind of interesting for all the thingsabout where is software needed or whereis UI needed? You kind of need that evenfor code, right? In a fully agenticworld. But that said, one of the thingsthat we are starting to see we startedseeing with co-work, but even some ofthe work we showed with autoautopilot,right? On what you see with claws, is agood one because if you sort of thinkabout a lot of human
00:08:44capital is doingthe glue work, right? If you now canaugment that with tokens/agentsthat are long-running, durable,right? Then your ability to scale evenwhat is still judgment and glue workgets amplified like coding does.So, you can like I'm positive that 6months from now we'll all be saying,"Oh, wow. Like all through night thenight there was a bunch of stuff thatall these autopilots that I have workingon my behalf with my delegated authorityso to speak, right? I can sort of giveneven my identity did a bunch of work.Then of course I'll need my new IDE tosay, "What did you do? Like I might didI do this work?" And so on. So, I thinkthat that's where compressing ofworkflows,completing of tasks, Uh where I think alot of the value gets created.>> you raise a really interesting point,which is there's the actual agent isdoing the code and then there's aharness around it.And that's the environment, that's thecontext, that's everything you'resetting up as a developer aroundactually a coding agent. What is theharness for the enterprise? Is there anequivalent concept for broaderproductivity work or how do you thinkabout that concept sort of generally?>> That's right. So so in some sense, youkind of want the
00:09:59harness to define themodels, the the data, uh and the tools.And so that you have a loop across thosethree. And so what we are trying tofirst of all make sure is each of ourproducts that we build, right? Whetherit's GitHub Copilot or the securitycopilot the stuff we showed with M dashor even the discovery for science, itdoesn't matter. All of them aremultimodal harnesses um with toolsaccess so that you can do thisprogressive uh disclosure of tools evenso that they're token efficient. Uh andthen you're feeding it with very richcontext because that's sort of the otherhard lesson we've learned in the last 2years is oh my god, the amount of workyou need to do to prep the context layeruh such that your plan can execute inthe most efficient way is where themagic is. So we have in our case, wehave the GitHub harness, whichessentially we're using across all ourproducts. It's available in foundry. Andwe're open like you can use your llamaharness, whatever, or you can use theumuh you know, any open harness or anyharness of yours and train with yourtools and multiple models and yourcontext. And so that's the pitch
00:11:14becauseright now a lot of dialogue is um hey,if I train the harness plus tools andthe model together, you get evals. Andwhat we are proving out is and the bestexample of that is what we did with Mdash, right? Because when it launched,uhit found bugs or vulnerabilities thatwere not found by Mythos.Uhand so there is existence proof, I wouldclaim, that you can have a multi-modalharnessuh that can in fact be moreuh performant in the real world.>> So the premise behind the uh training atthe independent frontier labs is really,you know, we're going to have thesemodels and we'll have an API businessand we'll support enterprises andstartups, buta first-party product, be itproductivity or code or search, drivesthe majority of revenue. That's adifferent value equation than you'redescribing. I think with the Microsoftecosystem, uh if if that's the case,tell me if it's the case, uh cuzobviously you have first-party productsand you have enablement products. Umwhat is the role of the develop Likewhat's going to be hard and the set ofskills and the value capture thedeveloper has in that world?>> Yeah, so I think that there's alwaysgoing to be the case thatsomeone who's super successful
00:12:29and as a platform builder can also havefirst-party products.It was true with Windows, it was true uhwith uh the the SaaS side and the cloudside as well with us and others and soon. But the thing that is is it shouldnot be a limiter to other peopleachieving that same success, right? ThatI think is the core difference, which isthe the network effects this time aroundaround intelligence are such becausethey learn from data and not really lotsof It's just a few samples that you haveto see to understand what's novel aboutsomething.So that's why the game becomes how toprotect. So that's why I would say everycompanyhaving private E valves may be thebiggest IP, right? I think about it.Like what's that private E valve thatyou can then use even a frontier modelto hill climb on and not leak the tracesmaybe one of the biggest driversuh of IP. Like so in other words,another acid test is you have an evalthat's private. You're usingmodel A. Can you switch it to model Band you know, climb up? If you can, thenyou're in control.
00:13:44If you can't, you'renot in control. And that's where eventhe harness decision becomes superimportant, right? So therefore, havingan open harness, letting all models comein, having your evals, your contacts,your tools help you hill climb, I thinkis the skills that an AI native startupneeds, a SaaS company needs, or everyenterprise needs.>> Yeah, I think in a very real way, yourMicrosoft historically as an operatingsystems company and then becoming acloud company, maybe like the third actis that you're a harness or evalscompany. Whatever Whatever the the sortof conglomerate of concepts that youwant to put together.I I I think like enabling every companyto have like frontier intelligence orwhat what I forget the the exact termthat you used.Is the is the mission, right? That isthat is the platform promise that youbuild with us, you will get yourintelligence for your data.>> That's it. That to me, that is the likeif there was one tagline for this entiredeveloper conference iscan everybody operate at the frontierwith their frontier intelligence, right?To me, that is so important becauseotherwise I I don't know how you achievestable equilibrium, right? Which is howdo I then go
00:15:00and say, "Wow, my companyis going to have a terminal valuebecause I now know how to continuouslycompound on top of what's a platformthat gets better, right?" So when likeWindows obviously came out, Adobe built,Autodesk built,or even like take what Jensen said, "Webuilt DX." And he built, you know, CUDAon top of it.Right? I mean, I always say to Jensen,"God, I got the short end of that."Right? I wishwe had recognized it. But nevertheless,but that idea that you can build aplatform layer that someone else canthen extend outand build their own intelligence layerin this case, I think is everything.Right? Without it, why have a developerconference? I can just come and have youall sort of just worship at the altar ofone model. But that's not a developerconference.>> Uh backstage, we had a discussion aboutwhat is IP or what is the value in acompany. It used to be the length ofhuman experience at a company. And nowit's this other thing, which is theevals, theexperience in sort of applying agents tothe company. Here, I just want you tolike flesh that out a bit more cuz>> Yeah, it's a great way to frame it,right? Because you have At the end ofthe day, every company is going to haveboth the human capital that is stillgoing to be super valuable
00:16:15becausehumansand their ability to find the gaps thatexist at all times is going to be theway we all will create value. Right? Imean, so I'm definitely in the camp thatthis is going to be about expressing newforms of human agency and ambition evenas token capital goes up. Right? So,let's say aany corporation has lots of tokens andlot of human capital. The question is,how do you compound the two? So, if youhave a like if you take in teams, I havea bunch of agents doing work and a bunchof humans doing work and the tracesbetween those, that is really importantcontext of how that enterprise iscreating value.Then, that goes back to train not ageneralist model, but to train the trainthe company veteran agent. Uh right?That is super valuable again, again,right? Which is when a company goes andsays it should in fact go on to thebalance sheet is how I think about it,right? That's In fact, there may be likehuman capital was never possible to goput on a balance sheetbecause you didn't know how to capturethe tacit knowledge. Whereas now I thinkyou can with the agents that havelearned through the through time throughall the traces.So that's what at least we think will
00:17:31happen.>> I think the SEC is going to have to haveaccounting standards for tokenexpertise.>> You're talking about the equilibriumstateand a stable equilibrium where companieshave this compounding value and can seeterminal value for themselves. Anotherchallenge to you know the consideredequilibrium of okay, there areapplications and workflows that are sortof common to a vertical or a horizontaland this was like the generation of SAScompanies and you know Microsoft haslots of SAS properties as well and thenthere are things that are very specificto every enterprise that they'redifferentiated against. I'm sure youhave heard much and participated much ofthe debate about the end of softwarebecause all these workflows are arecheap to generate now. Do you think theequilibrium looks different between whatagents get built in enterprises versusin their vendors in the future?>> Yeah, so I think what's happening thereis see we we had a particular way wecapturedI would say workflowin apps, right? Because we built up adata model, right? We schematized somepart of some business process.>> Mhm.>> We then built a bunch of business logic>> Yep.>> and then we put a bunch of UI on top ofit. Right? So that's kind
00:18:46of what everySAS company>> a little configuration.>> 20 20 years that was>> And that was it. So interestinglyenough, now you kind of get tore-litigate that vertical stacking,right? So I still think for example thatdata model that you build underneathevery SAS application is super good,right? It's like why reinvent it? Like Imy general ledger better be a generalledger. I don't need new schemacreation. In fact, that entityrelationshipis actually pretty good robust thingthat I want to feed.>> And you want to be stable.>> That's right. Then same thing withbusiness logic. Right, if you look atwe have this product called Power BI,right? It is like dashboards galorepeople created.The beauty underneath that dashboard isa very rich semantic model, right?Someone took the pain to create adashboard and do all the measures.And you want that that's business logic,right? I want that to be available tome. So, I think the challenge of theSaaS business model is we packaged oneway. We now have to learn how tounbundle these things and rebundle innew ways and discover new businessmodels, right? I mean, if you look at itwhat's happening
00:20:01today with Microsoft365 is a great example.Right, we have this thing called WorkIQ. In fact, what we are realizing is ohmy god, like if you look at it in fact,there's a historical parallel to right?We sold first exchange and SharePointandyou know, before Teams we had a thingcalled link server and what have you.And we thought oh, that's all going tomove to the cloud, but little did werealize that oh, the number of peoplewho will use servers in the cloud is 10x100x, right? Because people were notbuying servers, they were just buying asubscription.The same thing is now happening withM365 because with Work IQwe have exposed what was perhaps themost important database in a companythat never got used as a databasebecause it is only captive to our apps,right? It is all email operated on it,Teams operated on it, Word, Excel,PowerPoint, SharePoint. But now, likethis is one of the coolest things I getto do with Work IQ. I go to a GitHubrepo and I say, "Hey, I attended a bunchof design meetings last week related tothis repo. Can you capture all that andtell me what changes I should make. Imean, think about that. Right, itliterally can go look at all thosetranscripts, come back with a plan tochange a code base. Right, previouslyyou could never
00:21:16have thought of usingM365 for something like that. So, thevalue creation opportunity now in theagent world is in fact 10x more. But, itdoes require us to have, for example,there's going to be usage around M365,right? Which is going to be perhaps morethan even the end users, right? to evenre-architect. Like, in fact, like what Iused to serve an inbox or a mailboxcannot be used to serve an agent.Uh and so, that's sort of what we'redoing.>> I don't believe in like permanentbusiness models for any of thesedomains, but in the near term, do youhave a prediction betweenuh you know, outcomes-based pricing,token-based pricing,enterprise bundles?>> Yeah, the way I think about this isalways we have had like Let's even takethe per user pricing.>> Mhm.>> The per user pricing is really anartifact of someone creating a budgetneeding certainty.Right, because it's the most importantthing like somebody wants a budget,>> Mhm.>> they need a per user. And and per useris just a set of entitlements to usage.Right, that's kind of what it is. Andso, the way is if the first bundlingwill be take some usage, bundle it into
00:22:31per user stacks, and you know, then sellsubscriptions. So, subscriptions I thinkare going to be there, per user is goingto be there. Then, the next big thingwill be consumption. So, people will sayI want consumption. And it's alsopossible that people will say I don'teven want to pay for any of thesubscriptions or the consumptionsoutcome. But, remember,most people love outcomes until theyhave an outcome. Because once you havean outcome, it's like giving awayroyalty, right? I mean, like I I'vetalked to customers who love, you know,outcome-based pricing, and I say I'm allin until they oh my God, like what areyou talking about? You're sharing in myoutcome. No, no, no, I want you to goback to per user pricing and I want youto consumption price, right? So, I thinkthat debate will go on.But and all all the all of thesebusiness models have a particular timeand a place versus one to rule them all.And if anything, if you're a SaaS vendoror you're a platform vendor, having thatflexibility and quite frankly, we facethis with GitHub, right? We justrecently announced a per user pricing onGitHub.Because little you know,GitHub Copilot was constructed at a peruser level before we understood eventhe intensity of usage of agents, right?It was an interactive way for adeveloper to use code complete, maybe
00:23:46task. It is not like oh, I launched10,000 sort of agents that are going onall day, right? So, that is what theadjustment is about. So, now that wereally want there will always be a peruser. But they will have to be aconsumption meter.>> How do you think about the durability ofSaaS more generally? One thing I'veobserved is in a lot of enterprisesinternally, there will be teams thatalmost have agent euphoria. They're soexcited about the explosion of thingsthey can build that they're trying torebuild a lot of applications or goingto their SaaS vendors and saying we'renot going to work with you anymore orwe're considering an internal project.And it seems like in 6 to 9 months,maybe some of those people will comeback and say actually we we can'trebuild everything. How do you thinkabout what's durable in this world andwhat isn't?>> I think I think we have to go throughone full budget cycle on this to reallysee the uhuhthe sort of the emergence of theequilibrium. Because at the end of theday,there's marginal cost to even generatingthe app, right? So, in fact, it can beeven a a simple way to say it like ifyou should always acquire something ifthe marginal cost of building andmaintaininguh something on your own is higher. Uhright, that should be like it's aquantifiable right a quantifiable thing.
00:25:02And the maintenance part is important.Right, even like you got to rememberlike hey, you know, all the securitystuff that now AI will find you betterfix them too fast. Of course there's acoding agent to help you with but thenthat burns tokens. Right, so whoseresponsibility is it? It's kind of likea a cycle that you've got to thinkthrough. And I think we have gonethrough the excitement that I cangenerate a lot of software. I think thenext thing would be what software do Ireally want to generate? What softwaredo I want to use from others? How do Icompose these two into some agenticworkflow that I have agency over? Right,because I think there'll be very littletolerance for anybody who is inflexibleat the vendor level.But at the same time, I think thatanyone who has got that flexibilityshows up, delivers the value will beback at again. Right, we're sellingsoftware but we're just differentbusiness models in fact.>> Speaking about building software,one of my favorite moments from I thinka previous build maybe one or two yearsago was they had a big they they therewas a section of you building your ownsoftware. I'm curious if you're buildinganything now.>> Yeah, so I I think the you know, firstof all, let's face it. Right, buildingsoftwarehas made it possible for even theincompetence
00:26:17of a CEO of a company likeours,you can build. So thank God. But thatsaid, I I I I do feelthatyou know, something likeumGitHub Copilot to me and especially thenew sessions app or the new app has justmade it so much more possible for you tohave agency over artifacts that you feltyou couldn't touch before. Right, so tofor me as a CEO even to go to a codebase,to be able to learn about it. Like Iremember joining Microsoft long back,you know, first and then you say wheneverybody had to go in and look at, youknow, whatever Cutlers, Malak, or whathave you to learn how to do good C C++code.So now that ability to be more fullstack up and down is so good. But thatdoesn't mean every one of us should bedoing the same thing. The question ishow do you then have the ability toinspect things, learn things, seethings.I think it's just so much more. And soto me, what I'm building a lot of isthese long-running Foundry agents.Right, so there's autopilots. So theeasiest thing is to me, I think I justbuilt one even last week wherethe idea was,
00:27:32hey, can I have an agentthat is continuously monitoring,essentially my own chief of staffautopilot, right? We're going to havethat obviously in Scout. That's whatwe showed. But it is so easy and trivialto build. I took work IQ. I said, takework IQ, goand build a Foundry long-running agent,store all the memory inusing Raven, right? Basically as myback-end as a service. And lo andbehold, it built it. And not only builtit, I could say publish to Teams and itpublished the damn thing to Teams. Sothe abilityto haveyou know, some end-to-end project likethis complete is just pretty miraculous.>> you think that impacts the differenttypes of engineering roles that exist inthe future because right now I thinkthere's, you know, a dozen differenttypes of engineers that you can be fromQA, front end, etc. You know, there's abig swath. I've heard some people arguethat in four or five years we'llbasically end up with four engineeringroles. It'll be people who are managingagents. It'll be forward deployedengineers or FDEs. It'll be securityengineers. And then people working onlarge-scale infrastructure for a smallnumber of services. And then everythingelse just collapses into the agenticworld.>> Yeah,
00:28:47you think that's a correct view ofthe world?>> Yeah,>> mean I think I think we'll have toexperiment our way through it. But whatyou said is what there are some very atscale things. At LinkedIn, they didstructurally changeuh and you know, basically built up anew discipline called full stackbuilder, right? So they went and said,"Hey, let's bring uh people from designand product management, front-endengineering, all put them together. Uhbut also have an edge, right? It's notlike the design person still doesn'thave the design edge or the front-endperson doesn't have the front-end edge,but you can give yourself bigger scopein role so that you're not confined toone role. Um and thenequally, infrastructure has become verycritical, right? So in other words, likeI mean RLEs, I mean one thing we'verealized is even for the Excel team, forexample, building the RLE in which areward can be learned is actually one ofthe hardest sort of infrastructureproblems. Uh and so you kind of needeven new talent, right? Distributedsystems people even in what wasconsidered an end user app teamuh because it's a different skill set.So yes, infrastructure science is theother one, obviously. Um so I thinkwe'll see how these evolve, right?Where's the real
00:30:03I mean always the worldwill have a bunch of specialists. Umyou know, I think the generalist roleis going to be the most exciting, right?Because the leverage of a generalistum is where we're going to see themaximum returns, right? When when yousaid, "Hey, I coding." I'm now a generallike what I basically translatedknowledge work,right? Which I did where I created aWord document or a spreadsheet or evenuh and now I can build an app. Right?It's in the same sentenceuh right? That idea that, "Oh wow, mygeneralist skills have gotten higherleverage, I think is what we're going tosee across the board.>> Music to the ears of CEOs and VCs thatare like a little dangerous and a lot offun.>> Golden age for idea people.>> Idea people with a lot of agency. If youtake that idea of personal agency andyou just zoom it out to theorganizational context.UmUh my partner Mike Rynal who actuallystarted his career at Microsoft justwrote an essay where one of the bigtakeaways is it's an age where you canbe much more ambitious and you need tobe given the pace of the environment andhow quickly actually users and companiesare open to adopting new technologies.
00:31:18Um how do you think about I feel sillyasking this of somebody running a youknow trillion dollar plus companyalready, but how do you think about howMicrosoft can be more ambitious now?>> It's a great question. Um I think umI think the the thing in these type oftransitionsis to have a conceptual modelof how work can changeto go after outcomes that you couldhardly imagine previously, right? Infact, Kevin Scott has this nice line,right? Which is um when you can make theimpossible like when you're making hardthings easierthat's sort of one point of leverage buttrue ambition is about making theimpossible possible.Sonow the thing that is missing a littlebit in all of our organizations is whatis that new conceptual model of what canwe build? What was impossible and whatcan we build?And I'll give you one example of this,right? Which is I take great inspirationfrom sort of the people who are managingthe Azure net network.And they came to the this is from evenlast year.You know, we were scaling.
00:32:33You saw thatI I talked about sort of how we built inthe last 15 months more Azure capacitythan we built in the first 15 years. Imean, it's crazy. Right, it's prettywild. And it's the same team. So, theysaw that and they said, "Bob, this justain't going to work if we don'treconceptualize our work."So, they built Essentially, they said,"Our job is not to do Azure networking.Our job is to build the agentic systemdoes that does Azure networking." Right,these are the folks managing the 500plus fiber operators managing the van,right? All over. And fiber operationsultimately is a physical operation.Things get cut. Things get you know,have to be repaired. You know, we havefancy words called DevOps and so on.Basically, emails are coming in and yougot to go respond to them, take care ofit. So, they built this agentic system.They even have a character for it. It'scalled Miles and it sort of does allthis stuff. Right, they started sort ofscreaming for more tokens and so on. Andso, they were saying, "Look, I we don'tneed head count. We need tokens in orderto be able to manage our operation."That reconceptualizationof what their work is, right? They Theybasically took their work and made itmeta. Right, that meta work is now their
00:33:48new work.>> Mhm.>> Right, in the '80s if somebody had cometo us and said, "4 billion people aregoing to get up in the morning and starttyping." My model would have been, "Weneed 4 billion typists." But, we're notdoing typing. We're doing knowledgework. So, that to me I think is it,right? Which is whether it's Microsoftor whether it's any organization is togive ourselves permission to do newtypes of meta cognition, meta work usingthese new tools to change the outputsthat matter.And then really make the impossiblepossible. So, completing that dot orthat that connective tissue acrossthose, I think is where a lot of theenterprise value will get created.>> So, you talked about the data centers?>> Yeah, please ask.>> Oh, okay. Well,we this leads nicely into the datacenter build out. I always think justI'm just impressed at the sheer scale ofthe build out from Microsoft, but alsoeveryone else. That this is redefiningwhat it means to be a hyperscaler. And Ijust feel like that that that isunprecedented scale on finances,on the way you run the company, but alsothe communities that are that areimpacted.Then just talk a little bit more aboutwhat you're seeing on the ground. Likewhen you visit your>> Yeah, I think there are there are twoaspects of it. Obviously the the buildout is
00:35:03extraordinary.You know, nothing like this has happenedand it's great to be a one of theparticipants in it. But you brought upthe other part, right? I think at thispoint it's clear that unless we as anindustryare very principled about ensuring thatthe benefits of all the stuff we'retalking about are felt in real ways atthe community level, right? Because thisis not just a a campaign.Right, it has to be real where peopleare saying, "Look, this is not changingthe prices on energy for me. In fact, ifanything, it's bringing down pricesbecause long term there's going to be abetter grid. There is going to be moreenergy. Water consumption is in fact notsort of in fact water is beingreplenished, right? You got to really,you know, educate folks on truly what'shappening in the the closed loop systemswe're building. We have to invest in thetraining, the jobs, the tax base. Infact, the least talked about stuff isthe amount of jobs that get createdduring construction, after construction,what's the tax base that's there in thecommunity. And and all this has to bereal.And and if that is the case,
00:36:18then wewill have permission. If it is not, wewon't have permission. It's as simple asthat, right? Which is we we I think wehave have take it as an industry prettyseriously.Uh I think it's good for communities tobe skeptical, ask the hard question forus to do the hard work, earn that. Umbut at the end of the day, if this if wecan really be theI've always felt like in human history,if you use a lot of energy but alsocreate a lot of value for society,the story has been fantastic. If youdon't do that, it's not been that great.And this time around, I'm a firmbeliever that ultimately, if you do havea token economy that drivesproductivity, that drives economicgrowth, that drives broad-spread uh youknow, participation,better health outcomes,um then I think we will be in a greatplace. Uh and that's at least what weall have to be focused on.>> Yeah. It makes me think actually thatwith all these initiatives that you'redoing, might be easier to see ROI in thecommunities first before in enterprise.>> I I mean, I think both sides. In fact,it comes back together. It starts to bethe people in the communities are goingto be employed, are going to beparticipantsuh in the real economy, right? That's Ithink the question
00:37:34is. Like, if we ifthe broad economy is doing well and thecommunities are doing well, the dots getconnected. It's sort of the marketforces are such that we will connect thedots. And that I think is it. Like, yougot to be able to see the evidence. Youcan't be about any one company. Uh butit has to be broad economic growth andbroad you know, community permission.>> Yeah. What have you most updated yourthinking about currently or what haveyou most updated your personal models onregarding societal impact of AI?>> So, you're saying what's thethe>> What have you updated most on in termsof societal impact of AI?>> Yeah. I think the um thethe most umcritical thing is the first question weeven started with, which is we need totellthe story and make it realthat everybody has a real shot toparticipate as a first class participantin this new economy. And that's kind ofI think in the next 12 months, 18months, we need a way for people to say,"Oh, wow, I get it." Right? There'sgoing to be tremendous capability,tremendous amount of infrastructure,
00:38:50but I can see what is going to happenwhether it's the benefits like healthoutcomes or my ability to create astartup or my ability to run my localsort of store more efficiently. It'sjust happening and I see that benefitmyself, right? That to meyou know, earning that permission in apath-dependent way, we can't wait. See,the one thing you I've now learned is Ithink the world is going to be wayskeptical of tech and tech companiesthat say, "Trust us.We've got it. The future is going to beglorious."Uh you kind of have to deliver tangiblebenefits. Umand but frankly, politicians winningelectionsbecause they have advocated for that.That will be at least my adjustmentbecause without it umthinking that somehow because it's tooimportant to summarize. It's too much ofthe economy for it not to be the case.>> So, one very simple framework I havefor, you know, what are what is going tobe the broad benefit of AI umbeyond the communities just working intechnologyuh
00:40:05are sort of wealth creation. It'sgoing to happen in a ton of differentcompanies, startups and large companies.Then you have health care. You hadamazing demos today. There are companieslike Open Evidence. I think that ishappening. Um education seems likeanother one that's an obvious good wherewe haven't seen as much impact as I'dexpect. Do you have a hypothesis on whythat might be or if it'll come?>> Yeah, I mean I think this is whereagain, how we think about education, howyou know, recently I met with uh thefounders of Alpha School and learned alot about what they were going and goingabout. And it is fascinating to listenuh to how do you even rethink uh what iseducation really look like? Because Ithink it's actually very important. Uhand I'm not saying anythingtraditionally being done is lessimportant, right? I was even looking atthe uh it's fascinating to see I Iforget the which Stanford class it was.Uh the the Asian guidelines for CSsomething. Uhbecause you still need people to learn.Uh like it was an interesting AI classthat they were making sure people werelearning how to apply softmaxappropriately versus saying, "Hey, fixmy training run." Uh so I think learningconcepts is important. It's going to bea critical. But the way we create theincentives, what are the credentials,how we value those credentials,
00:41:21what isthe employment opportunity for thosecredentials. So I think that there is acomplete change that has to happenuh given the way to get to information,way to educate yourself, way tocontinuously keep yourself updated haschanged so much. So I think,interestingly enough, maybe the next bigstartup and success story could besomeone who builds a new university umor a new umpedagogy even of how to get someone togo through a curriculum and findeconomic opportunity uh that's highlyvaluable.>> Well, that has felt uhperhaps impossible for a long time, butit's a great note to end on andsomething that might be possible. Yeah.Thank you, Sayan.>> Thank you so much. Thank you. Iappreciate it. Thank you all.>> Find us on Twitter at No Priors Pod.Subscribe to our YouTube channel if youwant to see our faces. Follow the showon Apple Podcasts, Spotify, or whereveryou listen. That way you get a newepisode every week. And sign up foremails or find transcripts for everyepisode at no-priors.com.