AI Enterprise - Databricks & Glean | BG2 Guest Interview

00:00:00I think we have AGI. I think we haveartificial general intelligence. Wereally have. You you hear these 95% ofprojects fail, but like you know, likethat's that's that's actually what youwant. I I think the LLM is a commodity.People are not saying that, but it is acommodity. Like you can get gas fromthis gas station, you can get gas fromthat gas station, it doesn't matter.Just compare price. Is AI in a bubble?There is an AI bubble. Okay, so thenGlean is also in the bubble. Everybody'sin the bubble. No, I would I would saythere is a bubble.I I would say those three camps. Yeah,there is a super intelligence questcamp.I would be very worried there. There's asecond the researchers doing the, youknow, that's definitely not in a bubble.They're like the sober. Yeah, they'rethey're super sober and nobody caresabout them. And then there's,>> [laughter]>> all right, and they're probably the onesthat are right, unfortunately. And thenthere's the third camp which is ustrying to make this valuable. We're notin a bubble in a sense that we're notspending huge amounts of capital on whatwe are doing.Uh, we're just trying to get actualeconomic value instead of theseorganizations.

00:00:59Two legendary builders, Ali, Arvind.I'm so thrilled to get into this withyou because both of you have seen everysuper cycle I've lived through,internet,mobile,cloud,data and AI.Not just through the super cycles, butalso through the hype,the trough of disillusionment,and this time it's different.Today, we're going to chop it up on thestate of AI.You know, let's let's start with the20,000 ft view. Take stock of where weare.AI, we've seen consumer AI,billions of users. ChatGPT said theguns went off 3 years ago. Cloud,Perplexity, ChatGPT.People use it in the room.On the SMB and developer side, we've gothundreds of millions of users withCursor and Codex and Cloud Code and andand so on.Enterprise on the other hand,there's a lot of divide. It's hard tosee, a lot of fog of war.On one side, you've got models that areearningmath benchmarks and science benchmarksand engineering benchmarks.But on the other side, you've got theMIT report that's saying 95% of AIdeployments don't work.What's the reality?Bridge that gap for us. Lay it out asyou see it, a view from the top. So, Ithink first of all, I think

00:02:14we should wewe should know that people use AI intheir personal and work lives both. So,there's not so much of a divide. Likeyou know, everybody in your company isprobably using ChatGPTum and Claude and and other tools uh ona daily basis. Um the [clears throat]uhthe the thing that I I feel uh you know,is happening in enterprisesis you you hear these 95% of projectsfail, but like you know, like that'sthat's that's actually what you want.Like you you like when you are actuallyexperimenting with new technology, if ifif all of all of your projects arefailing, that means you didn't just nottrying enough, you know, at the moment.So, so I think when when I read thestudy like it was not a surprise for me.Um you know, we're going to actually seehopefully like you know, similar statsnext year, too uh because you wanteverybody in the industry to be to bereally eager and experiment and actuallyfigure out like you know, how to mix howhow how to actually make you know, getyou know, get benefits from thistechnology.This would make you guys by default the5% of AI that is working.>> [laughter]>> Which is uh one in 20.Maybe maybe you go to the 5%.What is what is a use case that isworking? And not just working like it'slike saving me time, but like it'sworking and it's transforming mycompany. Something that you can take tothe bank to to the CFO while while theCFO will not

00:03:30listen, but the legal won'tshut it down. All right, well, I mean,look, we're seeing a lot of use casesthat are working. Uhit's just that you know, you you justhave to it's not just you can justunleash the agent and it just works.Uh it's an engineering art. Like ifyou're going to have a company that'sgoing to be really differentiated likelike my company or your company oranyone's company and you want to beatthe competition,you can't just like you know, quicklyput something together and think thatyou know, the your competition is notgoing to do the same thing. So, that'sgoing to be you know, something thatneeds evaluations. It needs somethingthat you know, you're going toproductionize. It's going to takeeffort. You need a great team around it.But we're seeing a lot of them. LikeI'll give you some examples.Um Royal Bank of Canada uhbuilt agents with us that basically takeas soon as an earnings report comes out,so equity research analyst, their job isto put together these, you know, reportsthat say like you know, this is a buy,this is you know, hold and so on. Um theagent goes, gets the earnings report,gets all the previous earnings reports,gets all the competitors' earningsreports, gets everything that's going onin the market, does the full analysis,the news, everything, puts it alltogether and it can get the equityreport out in 15 minutes from theearnings call. Industry standard is 2hours.Of course, it's going to getcommoditized and others are going to dothat as well, but uh that's actuallyreally important use case

00:04:45that we'reseeing in finance.Um so, that's like finance example infinance, right? It's like and there'slots of examples like this. Siftingthrough hundreds of thousands ofdocuments, SEC reports, so on. That'sfinance.Um let's switch gears. Let's go to tohealth care. Health care completelydifferent. In health care, we have ayou know, customer Merckuh that in the life science spacecreated a model called Teddy. Teddystands for transformer enabled drugdiscovery.And um this is a transformer model kindof just like large language models thatcan predict the next word, but itinstead can figure out which genome ismissing if you remove a genome. So, itreally understands the gene regulatorynetwork and can really start telling youwhat's happening with gene expressionand so on. So, this is really importantfor drug discovery. It's the beginnings,but this is going to actually help us dothings that we couldn't do before. Let'spick retail also. So, I'm pickingdifferent. Health care is one. I gaveyou finance, right? The RBC one.Um let's go to retail, 7-Eleven.Agents that completely automate themarketing stack. I actually think themarketing stack is going to getdisrupted pretty heavily.So, umthese agents um can basically prepare,they can segment audience like thissegment wants to hear this and it canprepare all the

00:06:00marketing materialthat's like directly targeting you guysand it can put the campaigns togetherand do that.7-Eleven was doing this before as well,butyou know, this and we're seeing this atDatabricks as well. More and more isbeing done by agents and being automatedso you can just do it faster and you cansegment more fine-grained because beforeyou had to create the content for thegroups, that was a heavy contentcreation was something that was humanmanual labor. Now, you can actually dothat much much more. You can have allyour web materials completely customizedfor a target group. So, these areexamples where it is working. There arealso lots of examples where it's notworking. Even with Databricks, we're notjust the 5%, you know, we have some ofthat 95%, too, but some examples whereit's where it's where we're seeingsuccess. Ali, follow up on that offThese are great examples. Thank you.Maybe uh if you if you were to take it alayer up, what is common across theseuse cases or these organizations orthese CIOs that's making these use caseswork? Is there something that we canpattern match?Yeah, look, I I think the LLM is acommodity.People are not saying that, but it is acommodity. Like and you know, when Itook took econ classes, commodity waswhen it's interchangeable. Like you canget gas from this gas station, you canget gas from that gas station, itdoesn't matter. Just compare price. LLMshave become that way. Like it doesn'treally matter. This one is better rightnow, next week that one is better. Youcan't even keep up anymore, right?What's happening? So, they're acommodity.

00:07:15So, it's not about that. It'sreally comes down to your company,what data does your company have that'sspecial that your competitors don'thave?Can you leverage that and can you buildAI that really understands that data?Cuz that's not a commodity. There's notan AI out there that understands allyour business processes in your company,your secret sauce, and your data. That'snot a commodity. In fact, that's closerto the 95%. It really comes down tothat. Or if you have a complicatedprocess that just your company has, thisis how you deliver your product andservices in your company, and it's uhyou know, it'sthat portion can be disrupted with AIsomehow. If you can do that, now you canget ahead of your competition. But itcomes back to what makes your companyspecial. Unfortunately, a lot ofcompanies are just building uh commoditystuff. Like you should not be buildingthat cuz it's a thing that every companycan do. It's not special to your companyor to to your that's that's I think theproblem in a lot of the industry. Uhanother problem in the industry is thata lot of demo wear. It's really easy tomake cool demos with an AIand you know, therefore we're seeing alot of cool demos, but that's all theyare. Yeah. Well, you know, something wesay around at Altimeter quite a bit isyour AI strategy starts with your datastrategy. Yeah. So, you got to get thedata house in order first. And you know,there's a lot of reasons for for for use

00:08:31cases that are, you know, we weretrying, were not working. Maybe give usan example of the 95% of an AI bet thateither of you had at Databricks, atGlean, that did not work out and why itdidn't work out.It's actually an interesting thing youknow, with engineering today is youbuild systems and and never never beforehave you been in this mode where youstart with a great idea and doesn't seemlike good idea anymore like within 2weeks, you know, because we see a newdevelopment that happened. So, there arewe have like numerous failures in inengineeringuh on that front like you know, forexample, some of our fine-tuning work,building building models uh for specificuse case within our product like youknow, didn't didn't really pan out forus. And ultimately, the choice was thatyou know, we can go with uh alreadyuh built models whether they are smallopen source models uh hosted onDatabricks or or one of the large youknow, foundation models. The but ininternally like you know, from acorporate you know, use casesperspective, actually like we you know,we are also like in in many ways in thismode where a lot of our work actuallylike I would not say like fail, but itactually takes much longer than youknow, to actually generate success. Youknow, there are you know, we areactually trying to automate a lot of ourbusiness processes internally.Um

00:09:46and um like for example like youknow, one thing that I want is uh in ourcompany, I want everybody to actuallyknow exactly what their top priority forthe week is, what they want to work on.And and maybe, you know, we want an AIagent actually first tell them what thepriority should be and we want it all toall to be documented and we want asystem which actually then, you know,rolls it all rolls it all up and I get aview every week where I can actuallyquickly see, you know, what are all thedifferent people working on in thecompany and are they aligned is thataligned with, you know, what I want themto work on. And and this this is asimple thing like, you know,companies have alwaystried to actually have this, you know,as CEOs you always wanted and it'salways hard to make happen and wethought that AI would simply just, like,you know, magically do all of this workbecause, you know, like it has all thecontext and it has all the contextinside the company to make it happen butI still don't have it. So so so thingsdo take timeto actually, you know, to Ali's pointlike, you know, there isAI is just one more tool that you havein the toolkit. It does notsuddenly make building complexenterprise systems, you know,you know, like it doesn't make it likethat you can, you know, build it up likein in in one day. Doesn't it?>> Yeah.You know, the last time enterprises gotthis excited about a tool was calledRPA.

00:11:01Mhm. And we know how that ended. Itunfortunately fizzled out and, you know,somebody in the audience yesterday islike, "Hey, how is this timedifferent from RPA?" It seems likethe same movie, bigger budgets, betteractors.What's different this time? How how isthe nature of the architecture of thetechnology different from the previousautomation cycle? Either of you. Yeah,well, I mean I I first of all, RPA likeit didn't take, you know, it didn'tcapture my attention at all. So I haveno I actually can't, you know,>> [laughter]>> So so so I think II think like I I would not compare thesetwo technologies at all. Like, you know,you know, what we what we're seeing nowwith AI is so fundamental, you know,it's it's, you know, it's it's theuh you know, when when we saw it firstit was basically magic.And and we couldn't believe that this isa machinethat is doing this work. Machines justsimply cannot do these kind of things,you know, that we saw them do. Likewriting on their own, um having emotion,understanding emotion.Um so it's a umit's you know, it's it's fundamental,it's different and and theUm>> [clears throat]>> and and that's why like, you know, Idon't think, you know, we this thisthis technology is going to fizzle out.Um and it it's

00:12:16not like, you know, youdon't have to be like a financial expertor, you know, you know, like sort of adeep thinker on business. This is thisis obvious stuff. Like, you know, all ofus know, all of us feel it. All of allof us can see the capability of thistechnology and we know it's special andit's going to be it's going to bearound.Yeah. You want to hear my RPA? Please.You know, I mean it was rule-basedand the problem with it, especially ifyou're, you know, you're you wantsomething that automates what's going onon your desktop and automate the workthat's happening, it's just that there'stoo much unexpected things that happenand it's just hard and brittle to set itup. Mhm. It wasn't learning ever. Mhm.So there was like zero learning. It waslike you tell it exactly where the rulesand if you got something wrong, you youneed to go and go back and expand therules. Here, you have something that'slearning. Mhm. Right? So it can it canimprove and it can generalize and it canunderstand the patterns and do patternrecognition. Uh so that's thefundamental difference between thesetwo.>> 100%. Now, um there has been manystartups that have failed in thegenerative AI we're going to replace RPAwith generative AI models. There's manystartups that failed actually that Iknow of like pretty some high profileones.It's because um the paradigm we live intoday with AI is there's still problems.The biggest problem is that you bake amodeland that's where it's learned everythingit needs to learn and then you freezeit. Mhm. And then you

00:13:32launch it and thenmaybe you give it some context butthat's it, it's frozen. So therein liesthe problem that, you know, we need weneed an AI that really can sort ofcontinue learning while it's using thedesktop and clicking around. So I dothink this problem is hard to to to nailbut I think Arvin is right that it'slike there's no comparison at all. It'slike rule brittle rule-based stuffversus learning agentic system. Uh Ithink it's going to nail it perfectlybut we haven't really nailed computeruse yet. Yeah. Working on it. The numberone shift is this move fromif then else statements Mhm. to a moregenerative solution that figures out thesolution. Yeah. Um and so you're tradingbreath for for maybe determinism.Um that seems to be the difference andyou know, there's a lot of CIOs in theroom we've got here and they've gotbudgets coming up to plan.Uh if you were giving advice to them offlike, "Hey, based on everything I knowfrom my customer base,here's one thing or two things thatyou've got to figure out and and alignincentives on or it could be reliabilityproblem or org design. What advice wouldyou have for CIOs who are thinking abouttheir AI budgets right now?"Well, spend more.>> on Glean.Spend more, yeah, put it on Glean buttheI I think theumLike

00:14:47one thing, you know, whichwhich is important in AI market today isthat it's very new and there are manyplayers. In fact, um every everysoftware company is also an AI companynow. Um you can go and check theirwebsites.So so the I think it's it's just hard toactually figure out where to allocatethose budgets andwhat we tell people is that I think thewinners are yet to be identified and andsoexperiment with more vendors, do shortand shorter term contracts um and andyou know, while that's easy to say, uhit's hard to actually implement becauseevery product that you try has, youknow, it's a cost that you have to payto make it make it sort of even test it.So so you have to also pick productsthat are are easy to test. I mean, thoseare the the ones that don't require youto, you know, spend the next 6 monthstrying to implement something and youhave no idea what's going to come outafter that. Like, you know, the productsof today, the products that are builtwith the right AI, they should work, youknow,very very quickly for you.Crawl, walk, run.We're going to take a peek into thefuture. Shifting gears, you know, one ofthe things that keeps investors like meup in the nights isa quarter

00:16:02trillion being spent on Nvidiaon on the semi side of things. Assumingthat is just 50% of the capex, you'respending about half a trillion on capexand then you've got to earn about atrillion dollars of AI revenue for allof this capex to be worth it.Thisand just to put this in context, theentirety of the software industry earnsabout 400 billion dollars of revenue.This seems like a physicsproblem at this point. How do you how doyou think this this plays out? You know,you've got you've got to make about atrillion dollars of revenueto justify this present spend that'salready happening.How do you think this shakes out? Uhmaybe we start with you, Arvin.Mhm. Uh wrong person to start with but,you know, I'm an engineer and I don't Ishouldn't really, you know, think toomuch about who's spending what money.Like, you know, we're here to build ourproduct and add value. So that's so soso in some sense, you know, I'm I've notreally thought too much about thisproblem. But but if you think about AI,the you know, AI is not actually, youknow, extending software in a marginalway.It's a it's a it's a different productand in fact, you know, it's actuallygoing to grab a lot of revenue thatactually today is in services industrywhich is 25 times larger than softwareindustry. So there's there's a lot ofspend that is going to move. I mean,that the spend that

00:17:17you see happen on AIis actually sort of, you know, thoseservice dollars that are converting intoall AI or software dollars.Um and and I think the but with withthat said, you know,maybe maybe you have a more informedview on this.>> Arvin, do you think that's that's justto build on? You said, you know, I'm anengineer and I want to just buildsomething that's cool. I I do think it'snot binary, right? It's not like, "Okay,so the physics doesn't work out so thewhole thing will collapse." No, there'sgoing to be things that work and so itis a good idea to continue focusing onthe stuff that is obviously alreadyworking. Continue expanding on that. Uhbut I think if you zoom out, I thinkthere's like three paradigms or threekind of camps and I put Arvin in thethird camp. I actually put myself alsoin the third camp. But let's start withthe first camp. I think the first campis this quest for superintelligence campand it'syou know, I think all the frontier labuh labs are doing this. Like, you know,all three, four, or five of them,however you want to count them.And I think it's really still being alot of it comes from the scaling lawsmentality which is whoever has the mostGPUs and the most data is going to winthe quest for superintelligence which iskind of intelligence that's like onalmost like god-like. It leads torecursive self-improvement of the AI

00:18:32which then once you have that, it cancure cancer and solve all economicalproblems and we can probably 10x GDPover a few years period of time. So whatthe hell are you talking about thatthere's a physics problem? Like illanything any of your cost equations aregoing to pale in comparison to theeconomic value that this thing is goingto provide. So that's like one camp andthe way they're developing it is biggerand bigger clusters, more and moreenergy and that's how they're goingabout it. Um and that's where most ofthe capital is going, right?That's not the kind of capital you'respending or I'm spending but that's thatcamp. Um and then how do they know thatthey're succeeding? They're not justlike, "Oh, just trust us." They're, youknow, very smart people working on this.So the way they're approaching it,they're saying, "You know, we'll throwthe hardest questions we have at thewhatever AI we have now and if it nailsthem and we're making really rapidprogress. So what's your problem? Likewe're like look at math Olympiad. We'relike nailing these math Olympiadproblemsphysics Olympiad and programmingcontest. It's like better than any humanbeing. So like that's what they'rethey're throwing all the mostintellectually challengingThere's a second camp which are thepeople that created the originaltechnology, the the scientists whocreated the technology, got themthe computer science Nobel Prize for it,called Turing Award and that's, uh youknow, Rich Sutton who createdreinforcement learning which, you know,a lot of this stuff is built on.

00:19:47Youhave Yann LeCun who's one of the threefounding fathers and many others. Theyhave for many years Actually, I've beenyou know, I asked them for years.They've been saying that that first campis not going to That's like not even theright approach is their view.They're like, "No, that's just like autoregressive next token prediction. It'sjust probabilistically predicting thenext token. That's not how and usuallythey will say that's not how humanslearn. That's not how animals learn. Youknow,we operate in a different way. Yourbrain is not that way." And one exampleis that you know, even a child learnsvery quickly to walk and talk and dothings with very little data compared toyou know, like certainly no child isreading all of the internet's data fourtimes over before they learn to speak.So, I don't So, that's like camp numbertwo. Those guys by the way, they sayit's 20 years out. Mhm. So, they'resaying, "Hey, it's a physics problem andit's going to take 20 years to getthere."Which to me it's like, "I don't know.Like, leave me alone." Mhm. Uh let me doresearch. Third camp, which is I thinkwhat we are in, is I don't think we needsuper intelligence. Like, you know, Idon't think we need that superintelligence right now. Maybe they'llget there. That's awesome if they do.But, uh I think we have AGI. I think wehave artificial general intelligence. Wereally have it. We absolutely have it.It's like anyone who says we need to getto AGI,that's like it's it's a it's falsepremise to start with. We already

00:21:03haveAGI.Uh I came to the United States in 2009at UC Berkeley not far away from hereand I was in a AI lab. It was called AMPLab. The A was for algorithms and AI,machines and people. And these are allAI people. And back then, the definitionof AGI we had, we already have satisfiedthat. Like, Mhm.I know the discussions we had. And Iactually went back to some of thosefolks to see like, is it just me or whatwas the sentiment back in 2009? Andeverybody that I talked to said, "Yeah,that's by those standards we had AGI,but we've changed the definition now."We have those definitions, you know,ads. So, for 30 40 years we had adefinition of AGI. We've already hitthat.Now, we're changing it and moving thegoalpost. But, very obviously we alreadyhave AGI. Just use any of these LLMs andhave it do some reasoning.And certainly it's smarter than many alot of friends that you have that,right? Like, you know, let's let's notname our coworkers or whatever, right?Umso, you already have AGI. Now, now we'relike haggling over exactly how smart isit. You know, do you have a friendthat's smarter or not? Uh so, if wealready have AGI, we just need to makeit useful inside the enterprise. We needto just expand that 5% to be 10% 20%30%. So, that's why I think Arvind'sanswer is actually a good answer. Like,we have the AGI we need. Let us justfocus on solving the actual problemsinside the organizations. And I think wecan already

00:22:18that's that's enough toautomate a lot of the tasks and get hugeeconomic value out of it. We don'tactually need super intelligence forthat. That's good idea. If the superintelligence guys nail it, amazing. Thenwe've cured cancer. Um if they don't,hopefully the second camp comes up witha new thing in the next 20 years. That'salso awesome. We already have whateverwe need. So, Yeah. Yeah. Let us just doour engineering. Right. Yeah. Yeah.That's really good framing.And the way this manifests in the in theyou know, in the world is there's a datalayer. There's the intelligence layer,which is where camp one is presumablyproducing a lot of great models. Andthen there's the software layer wherethe users engage with.Where do you think value accrue if youwere to design 100 units of value acrossthese three layers? The data layer, theintelligence layer, and the software orthe application layer.Where do you think value accruesin the next 5 years?All right. This is a This is a toughquestion. I mean, I [clears throat]think the all all those three layersactually are very fundamental. Yeah. Ithought you were going to add a few morewhich are notYou didn't.Yeah, cuz I think I I feel like the likeas as Ali was saying that the models aregoing to be available to all of us. Youknow, they are going to

00:23:33be commodity.It's going to hard to sort ofsee that it the more spend goes to themversus you know, these layers on top. Umthe but how do you how do you say likeyou know, I it's hard to sort of come upwith you know, where the most value willbe.Umand And I also don't know if if actuallychanges from today's technologyarchitecture where again like you know,you think about in a pre-AI world,any any sort of like you know,enterpriseyou know, you know, application and datasystems, you know, you have you havedata systems. You you do have I guessyou don't have enough of thatintelligent layer today and then youhave the application layer. So, so I soI so I I guess you know, some somedollars will shift into it. And then wedo think that the intelligence layer isactually going to be pretty thick one.Maybe you know, it's it'll capture halfof the enterprise value.Anything to add, Ali? Uh yeah, no. I I Ithink that you know, uh yeah, there aremore layers in the stack depending onhow you want to do it. But, umI think as I said, the the LLMs is acommodity as Arvind said. You can getthem. Like, you know, but anyway, that'snot That doesn't mean those companiesare not going to be valuable. They canbe very I mean, TSMC is very valuable.But, um I'm saying they're going to bekind of like these fab-like companies.But, uh but they're interchangeable andwe've never seen something like thatever. I have not during all these Peoplejust switch LLMs

00:24:49like in one day.That's not the case with your you know,your iPhone versus Android or yourWindows versus your Mac or your anythingversus anything. Like, you know,uh you know, Google Sheets versus Excelis like huge religious battle inside ourcompany.Uh but but LLMs is like you know,because um it's a commodity as I said.It just speaks English or any languageyou like and you can and it gives youdifferent answers every time. Might aswell just try the cheaper one, thecheaper commodity or the slightlysmarter commodity. You can't even reallytell the difference, can you?Um so, then what is special is the datathat you have. Again, if your companyhas data that it has actually collectedthat your competitors do not have. Like,Glean is amazing. But, if you remove allthe data from Glean, it's there's no useto it, right? So, it's all about thedata that you have. Um and can yousecure the data also?Uh so, if we're going to have agentsrunning around accessing this data,like, "Oh, that's his HR data. Oh,here's the provost's salary information.Oops, I blurted out to all of you."Like, you know, like that's you know, sohow do you lock it down? Uh how do youmake sure that umthe [clears throat] there's governance?There's you know, there's also a lot ofworry around can What if it's using aChinese model? What if it's accessingthis information? What if it's sharingthis information with a competitor? Whatif it's interacting? So, the governancesecurity layer

00:26:04is going to be supersuper important. Uh but I do think mostof the value will accrue to the apps.Yeah. So, it's kind of And I thinkthat's common sense. It just I justdon't know which apps. Yeah. Uh I dothink Glean is amazing. I don't thinkit'sDo you think of it as an app orI don't know. But,Now, we see us as both app and aplatform. Yeah. Yes. So, I think it's alet's call it an app platform. I dothink it's amazing because it has thepotential to automateuh so much of the overhead inside of anorganization. Like, if you think aboutwhy do organizations have hundreds ofthousands of employees, you know, someorganizations or 50,000 20,000. A lot ofit is the coordination overhead of likeyou know, so many people have tocommunicate with each other. Hey, whathappened? What did you exactly mean bythis? Let's do a meeting where youexplain to me. I ask some questions.Let's Oh, let's invite these other guysalso and then write it down and thenjust The just the coordination overheadof organizations is massive. Right? It'slike this n squared problem that youknow, everybody needs to communicatewith everybody and they're communicatinginside their siloed org chart. But, howdo we get it across? So, this like youknow, through docs and Excel sheets andPowerPoints and meetings is how we likemove companies and organizationsforward. So much of that can beaugmented and be made more efficientwith Glean. So, that's why I think Gleanis amazing.Um but this is kind of like 2000. Yeah.And you ask what are the killer

00:27:19apps onthe internet. By the way, back then wethought it's like Cisco routers, youknow, Yeah. portals maybe with thousandsof links on them. Actually, I was likein I just started college andwe knew that the future of internetwould be portals, which are these webpages with a hundred links on it and youjust click click on the right link. Thisis before Google search. But, the futureof internet actually didn't look thatway. Ended up being you know, thingslike Facebook for friends and thingslike Airbnb for rentals and Uber foryour cab industryand you know,Twitter and so on. So, those becamegreat companies. So, I don't know whatthose are for the future. Uh they willpop up. Yeah. And they will be extremelyvaluable. But, okay. So, does that meanthat Databricks and Gleanthen basically will die and there'll bea new set of companies? No, back thenthere was actually an amazon.com alreadyin '98. There was already a Googleactually existed already in '98 and soknow, only a $300 company or somethinglike that, right? So, uh so, it's notbinary. We'll see what happens. I but Ido think there's going to be really alot of value will go to the future Yeah.you know, apps that will emerge.Speaking Let's double click into that.Yeah. The $300 companies of today atthat layer, software apps, Salesforce,ServiceNow.A lot of talk about software is beingdead. Um

00:28:34Satya calls them the crud apps.What is the future of this layer thattoday is called software that seems tobe heading towards becoming a database?Um and what do you see the the the thevalue accrue to to to those to thesethis part of the layer?Maybe start with you, Arvind.Yeah, I I I think that's aoversimplification.Um like for example, even to say thatSalesforce isyou know, it's just a database.Umyou know, it's a it's a full sort ofecosystem ofworkflowsand other applications, you know, thatare sort of built on top of thatinfrastructure. So, umSo, I sort of like you know, I haven'treally understood this concept of thatyou know, you have thisum like a you know, database you know,where all your enterprise data is andthen um and then people can just go andcreate dynamic UI experiencesuh on their own on top of that data. Onit like you know, every business can forexample, just create all the UI bythemselves on this. I don't I don'tthink you know, it's going to beuh happening like that because yes, youknow, AI makes it easyuh for you to build um you can have adatabase and you can build you can justtalk to AI and create a UI andexperience

00:29:50that thatthat is you know exactly what you wantit to be.But most times you actually won't knowwhat you want. Like you know, I think alot of like you know good thing aboutsoftware companies is that they actuallythink abouthow to actually take that data but thenpresent it in a way let you know makepeople interact with it or modify it ina way which sort of is natural and whichyou know drives you know like moreproductivity from a human. So so I thinkultimately like a software is anend-to-end stack in my opinion and allof these companies, you know, I don'tthink they're going away. I don't thinkyou know they're going to relegated tobecoming a database. Humans you knowover the last 20 years we got addictedto these screens. We scrunched over thescreens and we would input thisinformation with their with their keyswith the drop down and hey I met Arvintoday and this is what I learned. Itshould really be hey chat met with Arvinthis is what I learned remind me twodays to catch up with him. That thatwill happen. I mean that I think it'sgoing to happen. Yeah, that will happenin the next couple years and even Gleanyou'll you won't be able to want to typeyou want to talk to it. But I think thebig thing is data entry. Mhm. How doesthe data appear in that database?>> Mhm.And that's todayuh not completely automated. So you knowjust just for likeuh I I think a company that would bewell positioned to do that wouldactually kind of be Zoom.You know a lot of people don't thinkabout it that way. But Zoom is reallyshould be the the the perfect

00:31:05data entryuh application right? Cuz that's whereyou're having all the conversations andthat's where all the information'scoming out. And if they could you knowif it could work with Glean and getextract the most important informationYeah. store it all not in like astructured day table but like store thatinformation in system of record. Mhm. Ifyou had that that would be the fulldisruption of the SaaS We haveThat's that's actually is is is one ofthe most common agents these days youknow with Glean which is you take thesemeeting recordings you figure out likeyou know what you talk to the customeruh what were the action items and thenthe agent goes updates the notes inSalesforce with that like these thesekind of things are happening alreadyyeah. Meeting meetings is yeah is uhyou know like we in in Glean we havethis uhumpolicy [clears throat] where we recordevery single meeting internal meetingexternal meeting if our customers allowuh cuz there's so much so muchinformation you know in there. I uh Ijoined a meeting last week it was fourhumans and uh six AI note takers. Yeah.I heard about I think yesterday we weretalking about 17 note takers in one ofthe you know discussions. This this itfelt like the first you know it's likethe first scene of a movie where wherethe AI takes over. Clearly there's a lotof sprawl. There's there's like almosttoo many tools and and consolidationcoming>> [snorts]>> at

00:32:21some point. But uh but maybe yourguys's personal workflow you know youguys are CEOs in the age of AI a lot ofCIOs in the room they've got more jobsthan time on their hands. How are youusing AI for both your personal selfand how are you driving yourorganizations they're both largeorganizations to to adopt AI and andbenefit from it?Um maybe give us a glimpse of yourleadership in in the age of AI. MaybeAli we start with you this time. Yeah Imean we have agents for all kind ofstuff that we use you know everythingfrom you know we have agents that arereally good at understanding ourcustomers. We have an agent Raffi's thatRaffi's is the name which Raffi. Yeah ifI want to understand anything about anylike you know tell me best customerstory on this. Like you know I told youabout RBC Royal Bank Canada but I canjust ask it I need a use case I'm goingto get on stage I'm going to talk aboutfinance sector. Give me a use case thathas these it'll like just find you allthe information collected. So it'sreally really helpful for me for thesekind of things like when I get on stagelike this.Uh but also customer if you go into acustomer meetinguh you know I want to tell customer Xabout their biggest competitor Y howthey're using Databricks. Now maybe Y isnot a competitor is not using Databricksso then I shouldn't use I should use Zwhich actually is using Databricks.Maybe that's like the number twocompetitor. How do I get thisinformation super quickly? All of thoseare prepared at Databricks. So on thego-to-market

00:33:36side a lot of this is beingcompletely automated and we're usingthis.The marketing stack I already mentionedis heavily automated already like the alot of the tasks that happening inmarketingso we're seeing that stackthat happening. Um then there'sengineering that's like a whole bigthing like you know that's how we sortofand I think there's a whole changemanagement and how to do it right.You know initial attempts at automate alot of the software engineering atDatabricks kind of failed even there'snothing wrong with the AI. The problemis the humans and how we were organizedbut that's you know so those are likethe two big orgs Databricks is a big youknow 6,000 person go-to-market org and3-4,000 person R&D org and then there'ssome back office stuff. Those twoalready we're seeing heavy automationusing agents for all kind of the task.Then there's back office so that'sfinance and these functions. Finance isall on Databricks and it's all all theforecasting all the sort of it's allmoved to machine learning based. But ittook them a long time cuz they had theirExcel models and they're very proud ofthem and didn't want to you know but uhum again there's a change managementthere. We actually had external datascience team build the AI modelsand then eventually they became goodenough and nowfinance has taken those over

00:34:51and likeyou knowuh finance has kind of moved from Excelto Pythonlargely at Databricks. But it was ajourney cuz you know most of us speakspeak Excel. Similar thing is nowhappening to HRand other departments as well but Ithink they're like you know I think ingeneral HR departments are likeyou know even like they're not theclosest to doing this kind of analyticalwork with you know Excel and so on. Uhso maybe that's not quite as far alongbut yes it's we're seeing it everywhere.Yeah.Anything to add Arvin? Same for usanything I can share some of my ownpersonal usewith with it like so I one of our agentsis daily prep agent which I really lovebecause you know every morning um ittells me like you know what my day isgoing to be what I need to read what Ineed to prepare. Like most of themeetings you know I will have not havecontext it actually brings you know likethe plan for those meetings for me. Sothat's that's one of my favoriteagents you know that helps me feel moreconfident like you know for like how I'mgoing to do my meetings in the day.Um the other oneuh which I which I shared yesterday alsoumthe like you know I I've changed myinstinct and I think you know changingchanging instincts you know take taketake take a long time um and you knowwhen when you're the CEO like you're the

00:36:06boss and everybody listens to you andyou can just say [snorts] like you knowwhenever you have you know a smallquestion curiously just go and asksomebody and they're going to like youknow uh put 30 people on the taskactually get that get that answer for meand this is You're going to have a prepmeeting before the prep meeting beforethe meeting.>> so so all of that so so that so that'ssort of like you know and and so butit's sort of like for me it was easy. Ijust get to ask somebody and and thatyou know I changed that because I knew Iwas actually causing like you know a lotofthe the that was very expensive. So theso today like you know my instinct is tolike whenever I have curiosity wheneverI have questions when I need to dodata analysis when I need to writesomething you know my my letter to thecompany every month all of those thingsyou know like fundamentally I use youknow AI of course you know Glean in thiscase but to actually help medo my do my tasks. Yeah.>> More more I I think theyou have toyou you have to sort of have thatum belief. A lot of people won't do it.You have you have to have that beliefthat AI is a good collaborator. It's notgoing to do the work for you but if youuse it you're going to actually producebetter output eventually. Even if youdon't save time you know for the firstyou know first few months but you'reactually going to improve the quality ofyour output.

00:37:21Fascinating. Wellthis brings me to my favorite part ofthis conversation which is rapid fire.Short answers are fine long answers arewelcome.UmStart with the 12 months from now.Are the big AI companies that we know oftodayup or down? We'll start with OpenAI 12months from now stocks up or down?Ali and then Arvin.Up.And I'll say revenue will be up. I don'treally understand how stocks work.>> [laughter]>> Anthropic.Ali or Arvin.Up same. Okay.Arvin Let me let me talk can I get Yes[clears throat] of course. BecauseChatGPT is going to continue growing andit's on fire and it's what everybodyuses.Uh so is Gemini by the way. And thenAnthropic because more and more you knowcoding we've only like eaten into asmall portion of that market it's juststarted so.Is AI in a bubble yes or no?There is an AI bubble. Uh like sayinglike okay so then Glean is also in thebubble everybody's in the bubble. No Iwould I would say there is a bubble. I Iwould say those three camps. Yeah. Thereis a super intelligence quest camp.>> Mhm. I would be very worried there.There's a second the researchers doingthe you know that's definitely not in abubble they're like

00:38:37They're sober. Yeahthey're they're super sober nobody caresabout them.And then there's [laughter]right? And they're probably the onesthat are right unfortunately. And thenthere's the third camp which is ustrying to make this valuable. We're notin a bubble in a sense that we're notspending huge amounts of capital on whatwe are doing.We're just trying to get actual economicvalue inside of this organization. So II don't think it's binary but there is abubble. I mean there are startups withzero revenue worth you know 10 20 30billion.That's a bubble.Yeah.Same I mean I think theuh there are quite a few companies wherethere's a lot of optimism and valuationswhich are well ahead of the businessthat those companies have. In like Iguess you can say like you know comparedto non-AI companies like of course AIcompanies do havehigher higher higher multiples and umbut I think it you know that sort ofcomes from that you know thatthere's a good reason for it you knowcuz you know these are AI companies aregoing to growmore than non-AI companies for sure.Yeah.My favorite game at Ultimately we askour CEOs is a long short game is if youwere to pick a company a product, anidea that you're long, that you think isgoing to be a bigger deal than it istoday, what is that? And then short,which is, you know, there's more sizzlethan there's steak, more more hype thanreality.

00:39:52Pick a long, something that you'rereally optimistic on. Same order. Aliand then Arvind.Mhm. [snorts]I am very long on agents.You know, I think I'm very long onspeech. Speech as an interaction. Like Ithink keyboards are kind of basicallygoing to disappear completely. Wehaven't actually nailed speech. I know Iknow it feels like we have, but wehaven't cuz you're still using yourkeyboard. So as long as you're using akeyboard, we haven't nailed speech. ButI think we're this close to completelyeliminatingkeyboards. So I think that's that's abig one.What's What would I say? It's like, youknow, I do think coding is a little bitover hyped. I don't know if I wouldshort it. It's I mean, I think it'sstill the future. So I think that'sthat's one of them. I think automatinglike customer service and support is alittle bit over hyped. So, you know, Ibasically I think the things that theindustry thinks are like amazing andwe've made great progress. We probablyhaven't done as progress on it. And thena lot of the other things that are beingignored, you know, we're going to havebreakthroughs in those. So.Fascinating. Yeah.Yeah, and for for me, I thinkthe products that aregoing to change the paradigmwhere instead of you building a productand you know, expecting people to cometo you,if you understandyour user, your customer umvery

00:41:07deeply and actually bring the AI tothem. That's the category that I'mexcited about. I want I want I want tosee more proactive proactive AI productscoming coming to the market next year.Yeah. That that's That's That is what isgoing to actually take it from a5% of the users being power users to100%. Yeah. Yeah. Yeah.Your favorite AI tool that you use inyour lives.I think Glean is awesome. I mean, ifthat was not clear.Let's go.So he uses it all the time. I actually alot of the questions I would ask fromthe team. The The thing you said youchanged, I I I first ask Glean and thensee, you know, if it nails it or not.Then if it doesn't, then I'llspin up a 30-person team to go spend[laughter] a week and have threemeetings and all that to get, you know,the explanation of some simple conceptfor me. But usually Glean nails it.Yeah.Well, for for me, I'm excited about notetakers.I've I've used Grain and Otter.ai myselfand Fathom and a few others. But notetaking is actually fascinating. I mean,I think the I I feel like, you know, ifyou if you take those notes and then ifyou utilize it the right way, like forexample, what Ali was saying, like, youknow, that becomes the source of whatthen actually creates knowledge, savesdata in your systems. That's going tochange,

00:42:23you know, how companies work.Yeah. You know, in closing, I'd love toget your vision for your companies.We'll start with Ali's favorite tool,Glean.Congrats, you just announced crossing abig milestone, $200 million in revenuerun rate.You've You're signing big deals, $10million deals. You've got super users,I'm seeing you're seeing casual users.Paint us the vision for for Glean fromhere to a billion in revenue. I thinkwe're still doing annual planning, whichalso some, you know, AI companies aretelling me that's that's old school. Butbut but we're doing it regardless. We'redoing it. That's just because they'reearly startups. Like Did you Did you doannual planning when you started Glean?No. No. So. So, the But but I think forusumthe the thing that I'm most excitedabout again isSo we think a lot about AI literacy andhow do you get everybodyalong on this journey? And we're notseeing it right now. Like Glean is aheavily used product, but but stillthere's there's a big variance betweenthe top users and and and and the ones,you know, at the bottom.And and that's what we want to change.So for the future for us is we want tobe We want Glean to be this um verypersonal companion for every person inevery company

00:43:38in the world. um This Thiscompanion with which, you know, is is isyou know, you have a very confidentialrelationship with this companion in thesense that whatever you ask thiscompanion, you know, whatevercommunication you have with them, um youknow, it's it's fully privileged. Nobodyelse gets to see it. But this companionknows everything about you and your worklife. It knows your day, it knows yourweek. It knows who are you going tomeet.Umyou know, in the day-to-day, it knowsyour weekly goals. It knows what youknow, what things you're not good at orwhat your career ambitions are. And withall of that, you know, uh this thispersonal companion is umis sort of helping you now with yourwork. Um it you know, hopefully takesmajority of your tasks automatically. Umwhat you know, works on them before youask it to work on them. And and that'sSo that's sort of the vision that we'reyou know, taking our product to. We havemost of the you know, foundation forthis in place already. Um Today you haveto come to Glean to get most of thatwork done. In the future, we want Gleanto actually come to you and do thatwork.Fascinating. Well, we can keep going fora bit, but I'm being called on time.Thank you so much for chopping it upwith us. You got a lot of alpha, a lotof insights here. Really appreciate it.Thank you. Thank you. Thank you.Thank you.

00:44:56All right, gentlemen. Thank you so much.