Dust’s Gabriel Hubert and Stanislas Polu: Getting the Most From AI With Multiple Custom Agents
01:03:07 Share

We've asked the entire world to move from calculator technology, punch the same keys, you'll get the same result. The caster technology ask the same question, you'll get a slightly different result. This is not happen. This is the biggest shift, the use of the tools that we have since the advent of the computer. We're asking an entire cohort of the workforce to move to a drastic mindset.

And the only way you get that is by having a risk reward ratio that you're comfortable enough is like, you know what I am, i'm not asking you to be right hundred percent of the time. I'm asking you give me a draft that save me time many, many, many times over. And the distribution of r wise, something that i'm comfortable expLoring within its rating on. And I think that, that is really one of the predictions that we see in people who've tried to ChatGPT or in people who just curious with new technologies. They expect that some of it's gonna a bit broken, but the upside scenario to them is so clear and so telex that they are willing to make that trade off for that local risk to to get things started.

Welcomed the training data this week. We welcome Gabriel hubert and standard last polo, the confounders of dust, a unified products to build, share and deploy personalized A I assistance at work founded in early twenty twenty three after spending years at stripe and open a eye, second time founders gave in stand started dust with a view that one model will not rule them all and that multi model integration will be key to getting the most value out of A I assistant.

There were early to be convinced. Ed, the access to the propriety data you have in data silos will be key to unlocking the full power of ai, and they know that you want to keep that data private. We've work together for eighteen months in their predictions have been consistently pressured.

So today we decided to ask them about these predictions. We will get perspective on how they see the model landscape evolving, on the importance of product focus, over building propriety models, and on how A I can augment rather than replace human capabilities. And Gabriel, welcome to training data.

Thank you. Glad to be here. Yeah, thanks. good. And if you have IT to be here.

guys, first thing that I want to ask is you started this company in early twenty twenty three at the time. IT seems like one model might rule them all. And that model of the time was there is A G P G.

Three point five. I don't know if four had yet come out, but that was way ahead of the curve, and people were super blown away. Ugus came out with a pretty contrarian view that there actually would be many models and that the abilities stood those together and do advanced flows on top of that would be important? So far you been completely right. How did you get the confidence to make that decision a year and half ago?

Yeah think on on the on the model. But IT was clear hear that um many labs were already emerging. IT was not clear from the kind of general audience, but for the people that knew, the dynamics of the market was called many up to emerging. And I think um IT was uh kind of natural to us that there would be competition in that space and as a result, uh, there would be value in enabling people to quickly switch from one model to another to get the best value depending on value cases.

Yeah I think from from use that point, the point of being able to quickly evaluate and compare IT is obviously important. Looking ahead or already at some the conversations we're having a IT seems that the levels of scrutiny, security sensitivity of the data that being processed may also influence some different cases. So we're exciting these seeing people thinking about running smaller models on device for some use cases. And you can imagine a world where you want to be able to switch between and A P I call to a frontier model for something that's, uh, less sensitive, absolutely crucial to get like uh, cutting edge reasoning capabilities for and some smaller classification or some ziemer ts that could be done locally while the interface that you use a for your agent or your assistance that remains the same. And that switching requires the ability to have a layer on top of the models.

You have have been read about this every time ah as you've called this out. And so many other predictions over the past couple years of partnership have been uh non obvious and then correct uh I think this is still not obvious as in there will be many models. You'll have some local models, you'll some A P I call and then you actually, as a customer want to choose between them or want to have some control.

First of all, why do you think that will be as and why would there be multiple models? Secondly, wasn't that could have tracked away by some sort of router mechanisms, some hiper visor layer? And and does that happen? And would you be that hypervisor later? Um yeah help me understand that I think .

this media to two modes of Operation and we think about the futures。 So it's it's a by model distribution basically of the future. There's one where the technology as IT tends today keeps uh progressing rapidly. In which case is still going to be competition from rather big lad because we incredible need for G, P, S.

To build those latter and latter models because the only way we know to get those more Better, mostly by scale in that world, this kind of dynamic of being able to switch to the best model at time. Tea will remain true for a long time, I guess, until we reach whatever IT is that uh is at the end of that uh dynamic and then this D I potio asm, we can told but that uh later in more details that the accident, maybe the technology plato ing. And in which case it's not going to be one model.

It's going to be a gassin model and eventually everybody will get this the model. And eventually on your macbook m six, you'll able to to to train GPT six uh in a few hours, in a couple years and then the kind of rotor, uh the rotor need those kind of disappears because the technology is really communities. Ze, uh, in terms of trust producer tokens and every company, we will have that on model. We we would have our model in that world.

Oh, we get to push you on that. So there's you're building a business where you you sort of win regardless of which one of those world to be go into. Which one of those world do you think we're going into.

This one is definitely tRicky. And it's it's interesting because so in terms of capability of the models we've seen or had, the perception that the system was moving very quickly for the best years, we seen larger context support for audio, support for uh image and stuff. But the same time, the one call seeing that matters for a changing the world is the reasoning capabilities of those model rights.

And the current reasoning capabilities of those models has been actually pretty flat over the past two years. Uh, there are a lot of g as you is training, which is roughly slightly over two years ago, if I remember correctly, for the end of the internal train as any. And so that means that I would have passed to use terms of reasoning capabilities. It's been somewhat flat also think so.

There's the Kevin sky point of view, which is there's actually exponential progress, but you only get the sample that progress every so often. And so in the absence of a recent sample, people interpret IT as having been flat when in fact, there's just sample bias you can see IT.

So do you do you do you think he's right? Do you think there's actually expansion progress we just having going to see IT yet? Or do you think it's actually like asm coding and reasoning breakthrough s have not progressed at .

the right one might hope. Ah as far as i'm concerned, I have a strong feeling that it's a IT IT isn't been moving as fast as I would have expected in my most uh, optimistic views of the technology. And so that's why i'm i'm i'm allowing myself to to ask the question and or simply consider the difference .

in our what do I think only predictions for twenty twenty four also is that we would .

have a major reasoning breakthrough. Do you think it's coming? Yeah good, tough one because this one doesn't come for sure and even u pu doesn't matter that first uh hasn't come yet. Uh and there's there's many reasons to to believe IT might not be a uh uh a core technological limitation.

You you can make many opposition as to White maybe the case that IT takes time um do the scale of the clusters require to train the next duration of model is humongous and IT involved a lot of complexity from an if infrastructure and reprogramme ten points because G P U fails when you scale to that many G P U Better tester, they failed pretty all the time. All the training is very synchro across the cluster. And so I might just be the case that scaling up to the next order of magnetizes needed is just very, very, very hot and there wouldn't be kind of a uh, inner and limitation is just a phase where we learn how to go from red one to red five. But for G, P.

S, basically, and you are a, you are at OpenAI of pretty critical one time, so people know you from the dust experience. But one thing that I have to be remembred about stand, is you are a critical researcher at OpenAI from twenty nine through late twenty twenty two, got a one, a bunch of onder ful publication. Some of them relate to mathematics and A I. You work on this with a it's giver and the crew OpenAI, do you think that mathematics will be essential to this type of reasoning breakthrough? Or is IT orthodontal is something that we're actually to learn on textual language data?

I I remain quite convinced that is a great environment to to study. And IT was the tested that we added time with a game of uh then funding. He was working at fair and executes in subjects and IT was exactly IT was really shared as time we read where friend is competing the in in the workspace a three friend by the ideas um I think the the idea there was that mathematics, and in particular its form of formal mathematics that gives you perfect fiction, is a very unique environment.

Study reasoning abilities and to push reasoning abilities because you have a verified so you know it's constrained by being able to verify delicate of the model that in an informal set up would require humans checking them some extent. And so that very bit is probably something that has to and look something at some point he doesn't yet for any reason, but that something shouldn't look so y so and remain extremely relation on the kind of mass and formal mass and L M. S. studies. I remember one of the .

ways you were presenting as me when I was still very much shrimping up as a IT matches the door to software, software, the doors of the rest and and started with some of the critical systems that were the very only ones to have been hand proven and hand verified as an example of how much more cost IT was to do IT by hand and do IT by machine and an indication of the future gains we could expect from being able to extend that and democratize that.

Um you guys see a lot of action through the dust A P I calls. When you build the dust system, you're able to choose what type of lying model to use. Uh, you're able to call many different models like me as a user.

I often call not just a call three, but I call GPT four and I call the dust a system and I call I in my customer instance, I select one of many options. What have you seen in terms of trends, what's performing really well? I've personally been super impressed by the anthropic models as of late, but you guys have a much closer review of that.

I I think the um I mean so would have Carried out on on on trains. You know you going to have the usual coding of vices, the grass Green people going to want to switch just to see what he looks like on the other side. And so when you're observing the switch, is not nessa observing a conviction that the body on the other side is Better, observing the convention people want to try, but IT is true.

We've got a great feedback on on clothes, latest senate release and and empirical. We seeing some stickiness on on on on that model um in in our use base. I think that uh in a word on the street is uh for some coding application code stratas actually performing very, very well.

We haven't yet made IT available through dust, but it's yesterday. Oh, that will get sorry. See, this is the thing you'd get for reni amp esco and waking up. So you're reporting at seven o clock in the hot day.

So yeah code trial apparently is a is really interesting on some on some current capabilities and then you have to mix IT in with the actual experience that people are getting. So reasoning uh cannot be fully uh made independent from latency uh latency at some points last year could be basically a way to tell the time. In serenity co, you could see latency literally in the API as people waking up on the west coast.

So people have use cases that may be more or less torrent of those. The we cover the german I models and robes models opening as a mastro right now. And and we have seen some interest in moving away from the defauts, which when when we first launch with of an a ice models, not say that I isn't isn't forming grabble .

over the past year, there's been a lot of enthusiasm about open source models, and it's actually one of your predictions down. You have these great predictions every year about A I. I was really enjoy reading them.

One of them was that at some point this year, an open source model take the brief lee ah for L M quality that doesn't seem to have happened yet and IT also seems like the enthusiasm uh around not the enthusiasm around but rather the lead slashed acceleration of uh the open source models and comparison to the close source models has maybe slow down a little bit. Maybe back to that. Kevin's Scott about we're sampling at the create time is the continous times we haven't seen IT. But where do you think the open source ecosystems gonna will IT actually at some point uh surpassed the closed sce eco system.

I mean, everyone said that that that that echoes with what we said there is ready. And that by model distribution does one dirigo tion where open, this goes nowhere. And this one distributed open, this wins the whole thing right? Because, uh, if the technology plato, uh, open with obviously catches up and eventually every can train there, the high quality model themselves and at that point a there is uh no value in going for a for preparatory model.

Um so I think that a there's a cn I O where open was really eased the winner at the end uh which would be a first turn of event obviously uh and then uh in the current dynamic is true that the open service has been lagging so far, obviously there um um I think that the one that has to be called out is really a facebook or meta efforts because they have what he takes to try next the model and so far has been releasing every model very openly. And so that's exciting to see what will come out of them uh, in those next four months uh to maybe make the reaction true. The caveats to that is that assuming the best model are the largest, which is a somewhat safe uh, assumption yet IT can be discussed. Uh IT means that that mode will be mangos to some extent and that means that even if it's open source, nobody will be able .

to make IT run right.

Uh it'll just got to much money. Uh, you'll need A G P, S just to do in france. Um and so that will really trump the kind of usage of those even if they are Better in the current set of offers in terms of costs of running them. It's a point for .

the consumption that interesting because that means that you might still have a world where there's a lot of A P I based inference, demand for po, best inference regardless of whether the model on the other enders controlled, hosted open weights, whatever, because the techno abilities to for that one.

one of your founding assumptions are related to molecular and model performance. And this goes back almost two years now, was that even as of two years ago, the models were powerful enough and potentially economically viable enough that you could unlock a huge range of unique, compelling applications on top.

And at the bottle neck, even at that point, was not necessarily model quality so much as product and engineering that can happen on top of the model. I don't know if that's a consensus point of view today. You know we still hear a lot of people who are sort of waiting for the models to get Better. For what it's worth, we happen to agree with you. But the question is, what did you see in twenty, twenty two that gave you that point of view? And if we fast for to today, what is your lived experience spin, deploying the stuff in the enterprise in terms of where are the product and engineering unlocks that need to happen to bring the softer for vision?

My trigger potful for living up on Y I was seeing in playing with you before. And IT IT is IT was coming from two very control ducati motivations. The first was I city city for IT is crazy useful.

Nobody knows about. Nobody can use IT yet. And still IT exists and little almost already in the A P. I. I mean, at the time IT was two p three point five in the API which was kind of slightly small oration of G D but on the same train day so IT was a crazy good mother um which was and IT was basic cortex based model and IT was much Better than tragedy ity and I was available .

in the appeal.

And yet the A R of was ridiculously smaller at the time, like in existence, bio standards of what we see today. And so that was kind of the motivation and and that was mixed with the fact that I I, I, I was starting to feel the, I mean, had the intuition that he would be hard to invent a not facial mathematician with the current technology.

And so I was kind of seeing IT a not a date, but a very long bus slow fast food on where was working on at the same time. I was the the util of those models already when you use them for your day to day tasks. So that was first motivation.

And the very contradiction motivation I shored was Gabriel at the time, was if that technology goes all the way to H. I, it's the us. Train to build a company. So Better we Better do this right now because, oh, wise next time is good sheet. And I absolutely didn't answer your question, but like geron vacation, I I think I .

got me excited. And when we did stop brainstorming on on the ways to to deploy this just raw capability in the words what where IT made sense to dig was one inside on some of the limitations of the hybrid fine tuning. At the time that people were talking a lot about fine tuning, a lot of consultancy firms were selling a lot of slides that were essentially telling big companies to spend a lot of money for tuning.

And and the two things that cut up for me was then saying, you know, one is expensive and you do IT regularly and nobody knows that i'll have to do IT regularly. H, and too, it's really not the right idea for most of the things people are excited to find tune on and particularly like fine tuning on your company's data as a bad idea as opposed to maybe sometimes find tuning on some specific tasks where you can see gains. But um the idea that bringing the context of the company, which is obviously every real company's obsession, like how does this work for me? How do I get this to work? The way I like IT to work was going to happen with technologies that weren't just changing the model itself, but rather controlling the data.

IT has access to controlling the data any of its users have access to. And those are somewhat hybrid models between new world and old world. The very old world version of IT is the keyhole is is still the same, the sea, says the one, deciding how new technologies exposed to members of the company, the god will they are in place, the observably that's available to the team to measure its impact in and any data kes.

Those are old software problems that they still need to be rolled out on very new interfaces because the interface is now of these new assistance, these agents um and then some of the new problems are around access controls. Does access controls look and feel the same in a world where you have half of uh, the actions done by non humans? And I might want to have access to a file that's like two thousand and twenty.

Like do I have access to the file? Yes, on in two thousand and twenty four is like, well, maybe an assistant might have access to the file and can give me a summary of IT that leaves out some of the critic information I should not have access to, but still gives me access to some of the decision points, are important for me to move on with my job. And that set of primitives, that set of new answers, just doesn't really exist.

And how documents are stored today. So if you think about deploying the capability in a real world environment where people are still going to have to face those controls and and those god rails, the product layer is actually very thick. The application layer to build the logic and the use ability to ensure performance, but also adoption is is, is quite thin. And that was the I think that was the go to say, right? There's a lot to do here when I get started.

Maybe you can dig into that because when we interacted in you two, twenty, twenty three, you want give you toy twenty three, a lot of people were still starting these unna model companies. And you guys had a very specific opinion, which is the future is application later. And there's going to be a lot going on under the hood, and we're just going to be an abstraction layer on top of that and let things happen.

As I see, as IT happens, we're going to succeed, in any case, by building something that people actually use and love. First, how do you have the conviction for that? Secondly, how is that have been playing out? What has been the hard part about IT event in the sea OS and the enterprise enterprise ploy ments ah you guys have been wei ahead of the curve on rag. I mean, ever I was talking about fine tuning but you guys have done so much in terms of uh, retrieving IT was is before I was even called that really retrieving and actually making smart decisions around information. Walk to the step by step of from the idea of application layer to where you are today.

You can imagine the application like a conviction existing in a world where you still decide to be a frontier model. The reason we split the two is one IT seemed like a lot of money for a lot of risk, and I mean, a lot of money for a lot of risk to try and develop a frontier model or an equivalent to a front model, and also make a bet on the way IT was going to be distributed. And um it's our internal slogan was note no G P S before P M F.

We don't see the value in training our own model until we actually know which use cases it's gna get deployed on uh and there are much cheaper ways to explore and confirm which use cases are actually going to make most of values and generate most of the engagement. Um the second a reason was really about this this this data contradiction like the fact that the cut off dates for training on internet data a are hard to set continuously. The fact that you can actually get an internal understanding of what happened last week in a front model means that for tuning is a hard problem, that IT is not a solved problem scale.

And so if you walk from that conviction backwards, that means like it's there are many cases where it's not self. So another technology has to be the one to deliver most of the most of the games and and extracting a small piece of context uh, from documents where IT lives, feeding IT into the scenario, the workers that you need helpful. The one trend that seemed interesting was that actually many decisions require limited amounts of context and information to be greatly improved.

So the context windows at the time that were a small we're already compatible with some scenarios saying let's just bring the the information in. And what we've seen over the last year, of course, is the the increase in size of those context windows, which just makes IT easier to expose all the right data no more than the right data hopefully to the reasoning capabilities of the of the frontier model. Um and what we experience is first, all IT takes time for people to understand these distinctions.

It's hard and you have to get yourself out of your own bubble regularly to realize that is true. The world um the few ism quite evenly distributed yet and people have varying assumptions on what IT means to roll out A I intently or roll out the capable of of these fronty models on on their workloads and you have to walk them back on your what they really care about, which is always very simple things. You know I want to work faster.

I want to know the stuff that I missing out on. I want to be more productive or more efficient in some tasks that I find repetitive, and then only bring the explanation of what technologies is gonna lay that when it's absolutely necessary, because people will worry about their experience and how they feel about more than how it's working under the hood. Ninety nine percent.

The time at the big insight that happened and then I think willingly into we have been for a while, and it's great to see some of the market also doing that is people are actually really good at recognizing which tool they need in the tour box are getting. We ve we've we've not respected users enough and saying you need a single user that does absolutely everything. The rounding problem should be completely obstructed for you.

You should ask this question to the one oracle, and the oracle will reply. People are pretty comfortable telling the screw driver from a hammer. And you know, when they want to get to work and they need a screwdriver, they're very, very disappointed with one they get is a hammer.

And IT sounds like a hammer response. And so specializing agent, specializing assistance and making that easy to do, design the poy monitor, iterate on, improve all those words that require this. A IT was IT was quickly apparent to us. The people were very comfortable with that.

And h and so the the number one question that made us feel like we had an insight to hang onto an Ellen on words everybody asking us about us was obsess with the top use case and say, what are people using IT most for? What is the top use case across companies? what? And I know I can almost see the the amazon eyes trying to decide which deposit com they're going to vertically and integrate that which vertically use cases we know is build as a specialized version of this.

But I think the full story is fragmentation. I think the story is like giving the tools to a to a team or to a company to see opportunities for workers to be a improved on a augment and understanding the legal bricks that are going to help them do that. So rather than encapsulate the technological breaks that are useful and of take them away from users, exposing them at the right level gives people a unmoral autonomy and ready to the ability to design things that we had never fold up. Some of the they will come up. We we little imagine ourselves that I didn't makes .

sense like the fragmentation in providing people with the lego blocks to see what sort of use cases emerge. Just to make IT a little real low, can you share a couple of use cases that you've seen in your customer base that have been unique or surprising or particularly valuable? Do something .

to a little more, Angel. This obviously, that people are thinking that very of obvious use cases that have been interestingly and quickly deployed our enable ment of sales teams, support teams, marketing teams. And that is essentially context retrieval and content generation.

H, I need to answer a ticket, you know, I need to understand what the answers for the ticket is and generate a draft to the ticket. Uh, I need to talk to a customer. I need to understand which vertical therein and how our product solves their problems and draft and melt follow up on their objections.

Um I need to uh prepare a blog post to show how we differentiated from the market again, like i'm going to go plan to what makes special and generate without tone voice. Those were were pretty obvious and and quite expected. What i've been excited by is to see two types of things.

One, very individual assistance, personal personal aches. Uh, people generally, actually quite Young people in the first years of career asking for advice on a weekly, on a daily basis. Like how did I do today versus my goals? Where do you think I should focus my attention in the coming days? Can you actually break down my interactions on slack and and in in notion over the past couple of days and say where I could have been more.

I'm getting the feedback and i'm sort of sometimes talking too theoretical. Ally, can you point out the ways in which I can improve on that in the in these two notes that about to send? And so that's exciting because all bet was no, we want to make everybody a builder.

We want to make everybody able to see that. It's not that hard to get started. And by reducing the activation energy there to see small gains immediately rather than wait for the next model or the next version, that's going to really solve everything for them.

Personal use cases in grave, the second family of use cases that, that i'm excited by, uh, essentially cross functional. So where the data silos exists because the functions don't speak the same, they speak the same language, but they don't speak the same language. And so understanding what's happened in the code base when you don't know how a code is powerful, having an assistant translate into plain english, what the last poor request has been merged does is powerful.

It's powerful to people that were blocked in their work um didn't know who they should bug to actually get an update. So you know marketing to engineering, sales to engineering, the other scenario are in your extracting technical information from a long sales call is powerful because IT means that the engineer doesn't need the abstraction of A P M M or A P M to get nugget gets from the last call with a key account. They can just actually focus the attention of an assistance on that type of content on their own project and get those updates.

So I I say that the family of assistance that, that we're excited by because they really represent, I think, the future of how we'd love, uh, fast moving, well performing companies to work where the data is useful to you and the decisions you should make is always accessible. You don't need to worry about which function decided on IT or created IT. You can access IT, and that fluidity of information flowing through the company helps you make Better and faster decisions. Stay in day out. Um yeah any other examples are missing down that you think you're so good.

But no, yes, I think I think what I wanted to add is, is the fact that as he said, the usage is extremely fragmented. We we see over and over the same seno.

And so we have data to back that kind of proposition is is as we we built just as a sandbox ks, which is maxi extremely powerful uh and exam flexible, but also as the uh complexity of making activation of uses, uh not trivial because when you have horizontal sandbox like products you like, yes, but for what? And so july, the pilot face that goes with all users starts by clearly identified two cases. So they really kind of try to consider the question, what are the use guys I should care about for my company? And tried to agency a couple of them.

And we always see the same pattern. We see first use case, get deployed usage thoughts. We try to remove literally to another use case. Second use case gets deployed, usage picks up a bit more.

And then we generally goes for a face where the usage is kind of flats increasing slowly and eventually reaches gAmber critical massive usage and all of the scrog to something like seventy person of the company and that kind of the pattern of of kind of of our users and the scarring to seventy percent. Uh, the usage picks up the time the original loose case that were identified by the stakeout ders become trust anecdotical competitive. The rest of the usage.

And that's where we we feel like this provides od value. And it's very hard to know for us what are all those these case because for a we have examples of company with a few hundred people and a few hundred assistance. And so it's just it's just hard to answer the question. What are the best case is .

like these are great examples in that, that calls to mind an analogy that I would like to try out on you guys and you may pick on this and logy, but this this is what just showed up in my brain, which was um a lot of those cases as you describe. You could imagine some sort of vertical application being built around those use cases. And the analogy that comes to mind is there are gazillion an vertical applications.

And yet where does a lot of work happen? Spread shit. Why does this happen in spread sheet? Every knows that uses spread sheet there.

There they're flexible. You can you can customize into your hearts content. And so the analogies that wonder about is this almost like the spread sheet of the future.

You know, some of these applications may get peeled off of vertical specific applications, but even then, people are still gonna come back to the to the personal agent because it's just it's there, it's available and has as your data familiar, you know how to use IT. You can build what you want quickly and simply and effectively. Like, is that a reasonable and analogy for what this kind of is?

I think it's an amazing energy for another thing that i'm thinking about, which is IT took me the art, took the longest time to get to start work together way back this. This is like, I don't know, I was twenty years ago, fifteen years ago. And then one point stand uses IT for something is like, oh, well, this kind of like a cool rapper interface, where you can just get the result of your functions in real time.

And yeah, this is now the work. This thing is like it's a cool plant to facebook and engineers. I get IT now.

And yeah, I think it's also interesting for that experimental cost is very, very low if you think about um the way in which like some of our customers try try and describe that the games that they're experiencing that they are seeing their excitement for the future is some functions. We've had eighty percent uh, productivity gains. Some functions were in five percent productivity gains.

And we're not even sure that we're measuring them, right. But we're seeing gains when the specialization of the assistance is close enough to the actual worker that is able to augment the distribution problem of that with a vertical zed A, A vertical ze of assistance is almost impossible to solve. How are you gonna a get that deep into that function at a time where budgets a tight decision making on which technologies is going to be a fit is sometimes complicated when sometimes that's where the performance gains the most obvious.

One of our users has seen, like eight thousand hours a year, shaved off two workflows for an expansion into a country where they decided not have a full time team. And so basically sparing you some of the boring details, but like the ability to review websites, compare them to incorporation documents in a foreign language, have a policy check ker, that was a certain number of checking input, very clear to the agents that we're reviewing the accounts all in the languages in the geography that none of these people were yet familiar with because they were really expLoring the country. Um an immediate gains like very, very easy iteration on the first version of the assistant, two weeks to launch them into production rolled IT out to three to three human agents that were then assisted by by the assistance and their C T.

O. A sharing. Like you know we're in we seeing a north of six hundred hours uh a month. I am thinking our pricing terrible, but what i'm excited by is that is that, that case could not have been explored or discovered with a particular zed sales motion because I just don't know how you get to that fairly junior person in a set specific teen and and actually are able to pitch them and depth that quickly.

Uh, where's if you have that common infrastructure that people understand the bricks of? Not everybody knows how to do some products, not everybody knows how to do a power table, but everybody understands that they can just play round with the basic things and probably get help from somebody close to them as the other thing we've been. And the the the map of builders within companies, this heat map of people, what's a means about IT is that it's people who are just excited about iterating expLoring and and testing new stuff, which I think college well to know.

High performance or high potential in the future like this is heat seeking for potential in talent across the continent. Because the people using at the most of people who are the most comfortable saying I don't feel threatened by something is going to take the boring and repetitive side of my job way for me. I'm excited to have that go away and focus on the high value task.

I think for the first six months, I was one of the loud voices saying, what is that main use case? The two guys heard many, many times, and then eventually I realized, I, this is a primitive, were talking about sprats sheet. You could talk about freely, a word document.

You could talk about office sweet. When I interface with dust, I think about IT like slack. Except i'm not slacking my colleagues, i'm slacking assistance.

And they actually do this kind of work for me, and I can show them the kind of work. So IT feels, pat, to your point, something like a spandy eet meets the ergonomics of a slack. As IT is brought to me as supposed to. I have to go to to IT uh and that that that is IT took me a while to get there and now I see how the fragmentation is the power of what you're going after.

Give a good question on sort of the psychographic of your user because your you're common that is like heat seeking for the people who are sort of ambitious and innovative and stuff like that. I don't know if you have a name for them, but let's call them the makers.

You know the people who are not afraid to try to try new things in front to build stuff, have you come up with a systematic way to find those people? Or do they tend to find you through word of mouth or some other thing because that's not, you know, linked in profiles? Don't say, you know grio mika, right?

Like I think is a super interesting question at a couple of levels. But um emotion is is due, right? So the things that predicts a great outcome with dust coming out of a core and trying trying to think about what was most powerful about this call I had yesterday with the chief people in systems officer at the company that could not stop interrupting the five minutes.

And to my picture, yes, I did talk on, yes, I already I got a blog post on this OK. Where do I put my credit card that's called you next year? And it's the top down motion.

Is enthusiasm and optimism about this technology changing most things for most people who spend most of their days of in front of computer, you need that that that is very condition because I think IT unlocks three things. One, IT unlocks the belief in a horizontal platform for exploration. The ability for security to be in the support of business rather than a than a blocker.

And um and in genuinely sometimes uh example setting like we have founded and leadership teams, that is like how have you mention your own workforce last week and and leadership meetings are being asked din offsides, who I like how you gonna get Better answering like some of your teams queries faster with us. So once you have that, then you have the right sambo. I said that the right patch dish, I don't think we have fully cracked the build identification.

Ah so right now, it's more like bates. The products is incredible, easily use. Anybody can create an assistance even if they have not been labelled the builder by their organization. And it's just the sharing capabilities of their assistance that are somewhat thought. But we can see from the way in which people explore the product, create assistance for themselves, share them with their teammates in a limited way, a great prediction of that type of of. Personally, if you ask me to look at linton predict who are gonna be, who gonna be in, in, in that family, I take the number of discriminatory is someone to degree is a bit ages, but like people who may be earlier on in their careers, who have a mix of tasks that they obviously know they can get and assistant to help with, so they have use case ones laid out for them, people who have repetitive tasks and scripted that way out of a lot of of of repetitive things before .

just to be excluded, like we had the conversation. I think it's okay to say, like IT, is people under twenty five, like we are saying yesterday, the power users, the people are using this all the time. Um at the companies are the people under twenty five because they aren't set in their ways just to be explicit.

That doesn't an everyone. You can be seventy and constantly innovating in a new way. But in central, they don't have the pattern that they have been set to.

And by the way, that's true of a lot of the next generation Operatives ity notion, which that works really closely with. That is a under twenty five power law business. And you know the teammate here, other twenty five, keep pushing me to transfer over to notion and it's just a different type of thinking. IT feels like a very similar motion at dust.

Yeah I I, I I think that the the the one thing we had, the we have, which is useful, is the immense be to see success of ChatGPT as a now obviously world famous product. Uh, has made IT really easy to set up pilots by just telling teams, you know what, send a survey out, ask people how often they y've used ChatGPT for personal use in last seven days, like rank by descending order. And that's your pilot team.

That's the people you want to have. Poke holes at kick has. And because they have, we've asked the entire world to move from calculator technology, punch the same keys, you'll get the same result to the caster technology, ask the same question, you'll get a slightly different result.

This is not happen. This is the biggest shift, the use of the tools that we have since the advent of the computer. We're asking an entire cohort of the workforce to move to a dash mindset.

And the only way you get that is by having a risk reward ratio and that you're comfortable enough is like, you know what I am, i'm not asking you to be right hundred percent of the time. I'm asking IT give me a draft that save me time many, many, many times over and that distribution of r oise, something that I am comfortable expLoring within its rating on. And I think that, that is really one of the predictors that we see in people who've tried ChatGPT or in people who just curious with new technologies. They expect that some of it's gonna a bit broken, but the upside scenario to them is so clear and so tenets that they're willing to make that trade off for that local risk to get being started.

So you guys have a lot of very strongly health belief um internally and externally. And the good news is you ve consistently been right without this strongly help beliefs you named you of them. I mean you ve talked about this shift from the terminal tic to secure way before I was mainstreamed um talked about astorians ation and vectorially ation.

I I think about that that can be unpacked if you like. Certain ly would need big impact on the show if we go on that rabid hole you talk about no G P U versus P M F, right? Can you just walk through some of the beliefs that dust lives by? IT can either be philosophical um as a couple of these are or tactical like no D P S before P M F.

The first one is really uh uh the continued beautiful that uh focusing our products is the right thing to do because refuse to me like we only scratching the surface of what we can do with those models. Right now we are studying from the conversational interface. So that's why you use the slack energy.

And I really truly, I truly believe that that analogy, the slack energy will not sustain in time because the way we interact with that technology will change. He started with the conversation interface, but little handing a very different place motion. Basically those models are kind of the CPU of the computer, uh, the A, P, S, and and the tokens are ready to the bish interface.

What we're doing, right? Noise is merely uh, inventing bash scripts. And we have yet to invent the U I. We have yet to invent multi processing. We have yet to invent so many things. We are ready at the the, the the very beginning of what we can do from a product send point with that technology, whether IT evolves or whether IT stays legates.

Yes, one word that I think is is going to be important and and I feel recent news has has actually helped confirm as an interesting new drop in the bucket for as the ocean. So one of our product models is augmenting humanises, not replacing them. And it's not just the new version of saying like we're not here to get people fired.

It's really that we think there is the tremendous upside in giving people who will still have a job in five to ten years time the best possible exoskeleton. And that is a very different kind of company and kind of product conversation to be like, right? How many dollars are we gonna away from your apex line next year? A vers.

This is the number of latent opportunities that you are not able to explore as a business because your people are drag down and pushing like stale slide are around or not even knowing what dependencies they have on the rest of the company. This is how much friction you've imposed on the smart people you've spent so much money hiring because half of their day or part of their week is spent doing things that we should literally not be talking about. And two and twenty, that's one. And the thing that comes back to to to the droppers.

like you've been saying that from the beginning gril. And in the begin, you didn't use the productivity like you didn't want to use the word productivity. And I wonder if that shifted. And if so, the nuance around why you choose not.

do I think productivity, these two terms that I was hesitant in turn, productivity to me, sometimes feels like an an optimization when they when really um there's two ways to be productive, there's doing the same things faster and there's doing just Better things. And I think you know the effective productivity is is is is enshrined effort versus impact.

At the end of the day, your bosses is never going to be mad if you spent no time doing the things you are assigned to do, but brought in the biggest deal for the company. Nobody he's actually gonna make make any comments on on on that being the bad decision because I think the more you grow in your career and the more you are close to the leadership of the company, the more you realized it's known about the effort. It's really about the impact.

And the impact comes in sometimes unplanned hyper plenary, completely like left field ways where it's like, of course, we needed to focus on this is current hindsight, but you need to free up time, space, energy and mental cognitive space for that. Um the other one was enterprise search. I just feel like enterprise search is one that we didn't want to put on the website because retrieval of information is obviously use case that people are very excited about very quickly.

But um where is very convinced that looking for the document is a step that people are not particularly passionate about. Nobody wakes up in the morning and is like i'm so happy that i'm going to get to get the right document the first time when I do the same. People just want to get that job done.

And IT just so happens that using context from three different documents across seven data silos helps them get IT done faster or Better. So I think the search bit is just it's another jo B2Be don e. Nobody really wants to search. They want they want to complete, they they want to prove, they want to test. But the search bit is a step that we think we'll get abstracted and go back stands point.

I think that the interfaces in the experiences we have with this technology will sort of really try to forget about what the original data source was quite fast potentially once we've gone over that that the trust hodges that that exist today. Um the thing that that this all comes back to collaboration collaboration between human and non human agents and and I think projects uh by by anthropic are an amazing um an amazing example here. Um we thought about coeds tion last summer.

We have an amazing intern uh from t with this last summer who spend their time working on A A coelian interface. How do you chat to an assistance to make something that you're thinking about Better, whether it's an APP or a project or a document or script? And this is something that obviously the reason released by anthropic has has made very powerful to many more people.

That is to me, the the the the interface and the interaction that we need to get right and and that will be in the future. So we say augmentation and we'll stick to IT because I think IT really helps us focus on the interfaces that help humans and non humans make progress faster. Um it's going to be about proposals.

How do I how do I get to have a human in the loop with a proposal that's written just in the right way to decide if we swipe left or swipe rights on IT. It's going to be coalition tion. How do I have the language of the human in front of the assistant? Be as easy to interpret and as full proof as possible for the the final? Project is to move into its final form as as quickly as possible.

And so you need that, that interface, that interaction between the agent and and and the human. And you forget that when you replace too quickly, when when you focus on just replacing in removing you, you've build something that is fire and forget essentially and you'll see you'll see the games, you'll the dollar games um but I you know if you've automated a hundred percent of your customer support tickets, you still need the insights from more people are pissed off about. You still need to understand and have your singer on the pulse of why people are stuck.

Otherwise you're slowing down your product development efforts and product development efforts today live and and die by some of the comments that are coming in from support tickets. And so how you how you've made that problem go away and become like actually may be cheaper, sure. But also individual and IT is not, I think, a super long term view of how your product and business is onna serve your customers best. Um because you you still need to think about the ultimate interferences that are going to enable the decision making to make IT Better and strategic and and and and the best option for your customers in the future.

So keeping this human in the look always I mean, IT is human of one way to stay IT. But IT is p this human driven, like the whole point of all of this technology that we are building, is to serve humans Better. And as soon as you remove that, you made, you ve made a terrible mistake, because someone else is mark going to do that, and they're going to actually have a Better experience with customers and employees and stakeholders. And then they're gonna win. Know obviously.

the scenarios which are going to catch me and you me like this one, this, we know that humans get IT wrong way more. And so we should obviously place IT. This is a complex and nuance problem. So i'm sure that certain areas of IT where pure replacement has a fully understood non external like with no negative c externally value but I I adventure that we're pretty poor modeling where value is created and how it's fun through the parts of our company today. And you know economists have been greater showing that when you don't prise negative externals, well, we end up in messy situation.

Uh and so this is this is the question that that I post to leaders who are asking you watch like automate first like, well, I don't know which parts of the company do you worry about the most. And often I defined that get ceos of panicked about what their customers say on support tickets. And so making that problem go away, making that problem less visible might be great for some of bex conversations.

And and your stop Price could have enforcing consequences if you haven't a fun IT through in in the right places. But but but but also, I think there's so much more to do than to shave three percent of your baLance sheet. The the the spectrum of opportunity that you're giving your team, if these technologies in their hands and if they are able to come up with ideas is broader than just firing people out there, anything you shouldn't do that?

Like I think that I don't want us to be perceived as like nine in in in, in this ecosystem where the disruptive nature of this technology is going to take some people's jobs away because those jobs were currently being done by humans, full lack of a Better alternative. I think in certain situations, you could see those jobs having been created because we are waiting for the robots having been framed in a way that was because we were waiting for the robots. But I don't know that that's what leaders of companies are excited by. I think the the the outside, the future, the way in which we need to be resilient forwards to common and what our competition is gonna up in, those are the ways in which energy and support I fields should be shared to supporting you guys.

Um second time founders, uh you started your first company over ten years ago. Uh, you were early acquisition of strike. You guys were their student early on. What have you learned and done differently this time as second time fatter?

Um I think um really understanding that a few explosive bets are more likely to get you anywhere meaningful than over optimizing too early on on some think that is still meaningless in the market. That's one thing that I I think we think about differently. So like expLoring verses, exploiting a um and and all these frameworks.

Uh that's one. Um I think the transparency that you get the trust in in partment that you give to your team is we I don't think we were against IT. It's more that we will close this about how much more empowers you could be.

So uh the the idea one of the best words from my stripe, as with paper trail and IT, was you had two people in a corridor have a conversation and then one of them would take the time to just write a paper trail and back or in a document, say, you know what, we just had this exchange and we've moved the needle in this in this direction and IT saved and other humans the time and effort to go in a meeting room or figure out that this decision has been made and and and IT feeds a graph network of trust and respect for your coworkers that is, I think, second to none. And how you and then just achieve more as a team. So culturally, you need to sort push that to begin with because especially people who are earlier in their career will not always feel comfortable with our information should be net.

So I think that's one where where examples is um big markets that you really believe in for a long time. And we love the technology when we started our first company like twelve years ago. This is great.

This is amazing. These are Q R codes. Everybody is is going to use them. Uh and like now we have to wait for pandemic to sell q codes. Okay, do that next.

And and and so like throwing falling in love with the technology and not really fundamentally understanding how big the business could be if it's successful. And asking that question early in a abash. There is one thing that I feel is is different.

So what what do we kept? These are our experience together. I think the add vantage to uh h having built a company, I was a person because you've explored everything.

You've explored the the the the beauty, the terrible, the joy, the pain and you know entire API in an author. Enables much more efficient, 靠 funding， I mean cofounder uh action and corporation. I think it's a really big and fag and fair vintage.

Um I think the biggest one that I that I that I that I think he is completely different from me uh and that h mention is a about empery. It's a as a really as a funder. It's not you as good. I mean it's not to you early and it's to you to build them to to to build the initial Spark.

But then, uh, for the sake of the company, uh, you are not the one that has to build you, the one that has to make create an environment for people to be import, to build those scenes and explore and create new stuff. And uh, the best value you can give this, I don't like to lose that to use that world that leadership was coming combed. It's not necessary really should is really guidance and trying to do to create environment where every is as the chance to do what they they want. But yes, uh in a guide and environments with h everything works as a whole, but that won't the biggest different that we learn about tripe at this special so guys.

let's move to a lighting round. Got a couple questions for you all right?

Lighting round question number one um still ensure these predictions for where the world of A I is going on twitter from time to time at this moment. What is your top contrarian prediction for where the world that day I is going? And don't give me this buy model little bit of this little bit of that. So I see your point of view what's what's your top contrary prediction for the world of A I is going.

um I see the, uh. It's a lightning. I have just something ah it's going to it's going be tough. It's going to be we we I think we on the withdrawn turing a prety tough period how so at the excitement will go down. Maybe i'll take times to get in the next stage of the technology. This treatment is value to create, but people will not sit uh yet and it'll take a long time for to dessus for society. So there is massive review of value to create, but it's gonna a that we might have tough times in from us.

right? Short term as mist, long term optimistic. Lighting room question number two. Uh, and this is for both of you, who do you admire most in the world? V I uh .

ah ah is just is just incredible. I ve had the chance to work with him ah his my favorite people in the his he is extremely smart but is not a genius build there. He's a genius, uh, leader.

He is just a visionary. And I think that uh, been increasable capacity. I know know him uh I don't I actually don't know him by my lots and h in terms of pure genius in A I I think it's a shaman and yau by open eye. They have crazy last names, all that people look at up. But so human, any I you are .

I am impressed by those who've been around for a while and a good and acting as good resistance in condensor elements in the system. They just know I providing the friction to uh remain in optimistic but cautiously so. And and to me in one of the first I think IT was I can't ever as a tweet or a podcast or article, but the hearing Younger current to be like, you know, we can make pretty good decisions with a gas of water in a sandwich.

And these things require power station sized data sources and are not making great decisions on some things. So we've we feel something is missing and he is like elegant that he putting that back into into into perspective. IT has been interesting to me because it's hard to not cave to the hype, I think.

And so in in some ways, pushing for a simple idea like being open, which which I think yan accounts doing uh quite aggressive despite that not always probably being the easiest decision uh and also saying you we probably haven't solved everything all the time ah is uh is nice and my for my person experience the researchers that have worked for or with him, I have learned and taken from from that quite a bit and so that that and it's not french, but a you touch modesty, touch of temperance. I've i've appreciated in in in my discovery of the general sight of about vient tiles like up to ten years of just doing your prediction and classes ation from fraud at risk in and on boarding at story and and health care claims management and things like that. Um it's nice to to feel like there's a uh some people who ve seen seen a lot, done a lot and I just questioning rather than a affirming .

all right. So that brings into the third and final lighting round question. You chose a frenchman for your most admired uh, gay real and dust is proudly made in france. Paris has been in an epicenter. Certainly end up the center for all things A I um your take on the prison ecosystem and what do you want to say for the french founders listening to this podcast other than I started with an english? It's it's their fault, not ours.

Yeah I think that the french of system is awesome because um we we compared to where wage was twelve years or two fifteen years first company, no, we have tenants because there's been generation of scales that once through the market and and train all all that talents and and most recently, that kind of an explosion of I talent as well, which is this very exciting. Um so I say IT creates a pool .

of of of tenant and .

uh with the right conditions to create a create incredible companies. Uh obviously it's not uh tackling the the U S. Market from france chAllenge and so that that things to to be becoming into account.

of course. Yeah, I think this if if you have ambition, there's a lot more to do. And then as long as you're not knife where you know there is still some realities like you, you can you can fight some aspects of, uh, narratives. You can fight gravity, at least you shouldn't. You should probably work with gravity well, you should fight, but there is a ton more we can do. And I think we we have to behave a little more like um tech countries like israel I think uh uh mixing uh root example tion A A recognition for where talent is and how it's already connected and and has high trust connective tissue which think is a great a catholic and excEllent and in making great companies happen. Um but a recognition for weather markets are where people are buying, where people are paying and how quickly people are making decisions on on shifting to new technology, especially in that space.

I think the the biggest advice is uh as as a french funder, if you always been friends, you have kind of that feeling that something magical must be happening in U. S. Something special, there be something special about those people.

Well, tell you a venesta an option. I am working with ico. A these all are novel humans. You don't have any logical capabilities to just like us.

And so very important to to be be ambitious and and believe strongly that you can, you can make IT. You can do IT wherever these are from friends. This is U. S. wonderful.

That's a good place to end. Thank you, gentlemen.

like you guys.