Scale AI interim CEO Jason Droege: two $100M+ businesses, monthly growth since Meta deal, and the trough of AI disillusionment in enterprises

life in which I was living. Armani told the journal in 2012. Johning, I hate to cut you off. Let's not keep our guests waiting. Let's not keep our guest waiting, though. That is a fantastic story. Welcome to the stream. Bring in Jason. Sorry, John was getting carried away there.

Have you heard the story of Georgio Armani? He just died. And uh what a fantastic obituary in the journal today. No, I was hoping you'd keep going. That was that was great. Sorry to keep you waiting. Great to meet you, Jason. Thanks for coming on. Thank you so much for hopping on. Um yeah, nice to meet you guys.

Why don't you uh kick us off with an introduction and kind of your path to how you wind up wound up in this particular situation? Yeah, totally. So, I've been in tech for a long time. I've been at scale for a year. Um my background has been in uh starting companies, you know, early early in my career.

Um onto, you know, more notably at Uber. I started the Uber Eatats business there. Um and that grew uh you know, much larger than we ever thought it would. How big is the uh how big is that business today if you don't mind me asking?

When I left it was about $20 billion a year of GM and sorry that I think they've grown it to maybe 70 billion 70 80 billion. It just keeps going. Um yeah the um DAR just announced yesterday that uh in Australia alone they've served a billion deliveries which is about two billion meals.

So, um, you know, that's a lot of shrimp on the barbie. You know, when you when you Yeah, when you, uh, you know, when you get involved in these things, you're always surprised at how big things can get. And I think that's going to end up being true here with this current wave.

Um, you know, and, uh, that was one of the reasons why I was excited to, you know, join scale a year ago, joining Alex at the time.

Um and uh you know I had experience in marketplaces done a lot of growth stage businesses things where where demand was outstripping supply and you needed to fix a lot of things to get the business on track. Um and you know it's been a it's been an exciting uh fascinating journey over the past year.

I mean nothing like I ever expected. Yeah. What was the uh the shape of the business when you came in? Like from my perspective scale has been a very interesting business in that it hasn't like many companies have a second act. which I feel like scales at a second and third maybe even a fourth act.

Uh the way I tell the story is uh you know Alex Wang is starting with uh creating data for self-driving cars. Eventually that business maybe gets rolled into Whimo and and Tesla directly and so scale expands into the RLHF boom during the LLM boom uh and and has kind of grown.

And then my my my question was are we going to see Scale AI take on robotics data or more specific tasks and so uh describe the shape of the business when you joined and and where you were thinking about taking it. Yeah. No, for sure that I mean the history is uh the history is amazing and fascinating.

I think Alex had this insight um that data was the most important thing to models. Um and that was 9 years ago. So that's very visionary. Um and I was very impressed whenever I was talking to him about a a role at the company, you know, only a year ago. Um and you're right, we've gone from different types of data.

Um and and one interesting thing about like this entire journey is that every single um uh type of data someone has said like, "Oh, when is this going to go away? When is this going to go away? When is this going to go away?

" And this is a common thing that we hear and we've heard it so many times that uh you know we sort of are used to addressing it and there's always some other source of data and one of the reasons why is because um data is the application in many ways.

Um you know you know you're you're going from a software space where it's software organizes data for use by humans. Now you've got data driving applications of things. And so as we generate more and more data, the models are serving customers.

Um those um customers have higher and higher demands of of what the model should do which then drives uh demand for more data and the data changes.

Um but uh you know we've gone through this so many times from expert data to you know now you know um you know models are working on like how do they do things for for you right? Right. It was about knowledge capture and then now it's about skill capture. Yep. So um it's been yeah it's been an awesome journey.

Um uh and the company has been extremely agile and extremely commercial over the entire journey which is um you know which leads us to this discussion. Yeah.

It hasn't just been just like like if you zoom out it's like one smooth sort of exponential but in fact there's different pools of data that have been maybe a series of stacked S-curves. But that's the story of all technology and and all progress.

Um, so I'm interested to know uh what I like when when when Alex did invest like the best, he had a very interesting point about uh robotics data being particularly rare.

Like at least we had the internet to crawl and scrape and that was something that didn't it eventually needed to be polished with RLHF, but uh with robotics data there's just so little out there.

Um, and and I was wondering if if if that was something that you were interested in looking at or uh if it was immediately the skill capture thing of these like really really niche experts, people who know about archaeology and it's just not on the internet and you need to go get some PhD, get them on the platform and start having them answer questions so that that that can get baked into the models and and uh and we can get out of the what is it like Gellman Amnesia Dunning Krueger for these models where it sounds great until you're asking about the the thing that you're the expert in and then you're like, "Oh, actually it's it's it's only at a it's only at a college level.

It's not actually at the full expert level. " Right. Right. Right. Right. Yes. So, yes, we do robotics data today. Um we've been doing, you know, that's that's one of the newest parts of our business. Um uh that's the way beyond expert data.

Uh so, we've been doing expert data for a year and a half plus, depending on where you draw the line on experts. But everyone in the data business is in the expert data, you know, is in the expert business because that's what the models need. that's what they demand.

Um, you know, something like 15% of the people in our network are PhDs. Um, 25% have master's degrees and like a very large percentage, I think 60% or so, um, have at least a bachelor's degree. And so, you know, that is where the knowledge capture has gone. Um, robotics, yeah, it is interesting, right?

Because you can't necessarily pre-train on this like corpus of information from the past 25 years of the internet. And so, that probably just increases the size of the opportunity going forward. But the space is new, right? I mean like like you know there aren't robots you know walking around the world yet.

So a lot of this is still um in development. One other thing worth um mentioning is like all the data types um from before we still do like we still data autonomous vehicles companies. We're still providing computer vision data to the US government um and other parties.

Um and so these things do stack up as you mentioned into what is now a pretty big business. Is there still some piece of the business that's not AI, not directly AI related?

Like uh like almost like mechanical Turk like I just need an actual human to do something at the at an API endpoint uh for you know a very reliable very scalable and so I come to you and I and instead of going to an LLM I just have people doing it on demand. Is that a thing? That's not a thing for us.

I do think that is a thing in the industry like some of that work has gone into BPOS and you know um sort of like lower margin less difficulty type tasks um and so we're at the frontier and part of being at the frontier means you do the hard stuff um the margins are better um the skill level required is higher um and it pushes you to constantly be driving the business forward and and that adaptability frankly that that started with Alex in terms of like always looking for the next thing and being really close to the tech is kind of what's allowed us to here on the customer side.

How are companies buying this data? It feel the it seems like there's overwhelming demand. And when you go to them and you say like 10 data, please 10. No, no, no. I I mean I mean more so like what what are the actual like customer relationships look like?

Seems like a lot of the labs and big players are actually working with multiple providers. Um and that's not necessarily a bad thing because different providers can can specialize in different things. Yeah.

I mean um you know we have you know that is true that's like that's sort of a uh a general understanding of the market the the reality though is on the ground if you want high quality frontier data that's difficult to get at volume um the number of providers goes down dramatically um and so there's a lot of claims being made out there about doing things at at scale um you know we're one of the largest providers today we've been one of the largest providers um and so I think think that that's, you know, through all of the change, that's what's allowed us to keep the relationships that we had, which is um it's actually pretty hard to get this data right like like you really have to have the right network.

You have to know how to talk to the network cuz if you imagine like being in the contri, we call them contributors um meaning the people who provide the data to the models uh through us.

If you imagine being in their mindset, you have to have a system that can uh interface with them on what is a part-time job for them to uh give high quality data for a very technical um audience which is the researcher.

Um you know uh so that it it changes the weights of the models in a positive way and that's not a super straightforward thing to do. Um and so uh yeah I mean and and and the business is going great. I mean like I've seen a lot of press out there.

Um there's, you know, there's been a lot of coverage, a lot of speculation. Um, you know, every month since the deal happened, we've had growth, which I think a lot of people will be surprised to hear.

Um, you know, given I mean, like I read the news and then we look at what's going on inside the company and we're just like, okay, we got to start talking a little bit more. Obviously, I've been in the seat only a few months and so there's been a lot going on, but um, yeah, like things are going well.

Um we have two I mean one thing worth noting like is we have two multiund million dollar businesses. Um we have our data business which we're most known for. Yeah.

So so if you were to split either of these out and we have an applications and services business which which applies this in in customers that business alone has hundreds of millions of dollars of revenue. And so if you were to break that out it'd be its own unicorn or deck of corn or whatever they're called these days.

Um so that's wild. Do you ever uh you get a inbound maybe customer fills out a form on your website. Do you ever you ever have to check to make sure it's not like a runaway super intelligence that's like you know uh off on its own just like hoovering up data? Well, I mean great uh uh great question.

Um uh we tend to know our customers pretty deeply on you gota you gota have the KYC process in place and and frankly they spend you know these are data is not cheap to procure and so um that would be a very expensive hoovering operation.

Uh if well if AI if if AI you know progresses like the AI 2027 crowd uh was projecting at some point or another a super intelligence will come and and you'll have to choose you know humanity or or or the super intelligence. Have you run into any like super small companies that are customers?

We've talked to some small LLM uh providers like uh they're not building the mega GPT4 GPT5 level models. They say, "We're still doing, you know, transformer-based AI, but it's it runs on a gaming a gaming graphics card. " And I feel like usually they're just distilling llama or something into a smaller model.

But is there any are there any customers that have come to you and and just had some weird niche use case that's like pretty light on the data, but so differentiated that it actually provides value for them in their particular niche, or is it all just like, you know, the huge labs, the big the people taking really big swings?

Yeah, I think the big swings is most of it. We do see some of the smaller stuff um uh where you know where we actually are starting to see some data needs which um is sort of you know it goes to just following the market is actually inside of um enterprises. Yeah.

So they're not actually building their own foundation model but they're hitting the limits of their own um uh sort of the data that they have to make applications useful in their environment. And obviously the US government, we do a lot of computer vision data for the department of defense too.

So um those are the more sort of smaller customers and maybe less publicized things that scale does.

Uh are are there other analogies to the DoD where there are companies that have private uh like more like more secure data needs and maybe the the scale process is working on top of their own proprietary data and they need to transform it somehow and and and you get involved. Is that is that a thing at all? Yeah.

So that's the applications and services side of our Sure. That makes sense. Right. And so and so the history there is um uh as we were supplying data to the model builders uh obviously scale is very well known.

Alex is very well connected the company knows a lot of people started contacting us saying like hey we kind of thought this AGI thing was just going to like pop out of the box and do everything for us and the reality is is like that's not what's happening at all.

Um uh and uh so as we got pulled into that world and being asked like well how do I apply this to an accounting problem? How do I apply this to a healthcare problem? um we started to see the actual fundamental problems inside of these organizations and why they can't get there.

Um, and what's interesting is I like I've actually been on site with some customers in the past, um, in the past month, you know, as we're getting to production, um, with this latest round.

And some of the limits they're hitting is, uh, the actual human knowledge from the executive staff or in like a healthcare setting, maybe the most senior doctor or in an accounting setting, the most senior accountant.

And they, you know, they sort of went into this this thinking like, okay, we might need some data from the subject matter experts here, but but you know, um maybe a junior accountant or maybe a junior lawyer will just go in and label some data where we don't have recorded data.

Um uh and you know, but we um you know, need the models to do something that they currently aren't doing.

And what's happening is is that the senior staff is finding that they need to get involved because as you build these agent systems um which agents are basically obviously just applications running um you know that leverage AI that coordinate with each other uh in order to align them you which goes to your AI 2027 comment which is like are they going to be aligned to like human values and will they do the things that we want them to do the way that we want them to do them.

You actually need the question becomes who do you align it to in an organization? Like do you align it to you know someone who's straight out of college? Maybe. Do you align it to the most senior person in the organization?

Maybe they need a partner to help them figure out like which do you align it do you align it to compliance? Yeah. The or you know somebody more on the operating side. Yeah. Yeah. Yeah. What are you guys seeing on the like smaller uh model builder side of things?

are you seeing like um the you know the things you're asking me about in terms of like like their needs and specialization? Yeah, I mean I I I think there's like this there's this big question right now about uh inference costs and and do you need to throw a reasoning model at everything?

token costs are coming down, but people just keep moving on to the they're like, "Oh, the the the better model is the same price as the last model. " So, I'm not even capturing that 10% uh or that 10x gain or whatever is happening in the in the cost of inference.

Uh the most recent data point that we actually got that was hard was from Notion in the Wall Street Journal. Uh Ivan, the founder, said that their gross margins dropped from 90% to 80%. You know, not not that bad. I think that that's a trade that they should take all day long.

Um but but you could imagine that uh yeah I mean I think I think go oh yeah you could imagine that so basically notion's probably paying for the the best-in-class intelligence uh and that's and that's having a material cost on their gross margins but if they figure out that okay what users are really excited to pay for is you know a report that summarizes everything that's happening in notion across all your documents or they really love templating with AI I or they really love having just a a a chat interface on the on the right that you know is somewhat aware of the documents and they can just kind of ask hey I I you know I I remember putting all of my medical records in one notion thing or I remember putting all my financial or my CRM over here like can you find that for me?

um over time I would bet that they're able to distill that into smaller and smaller models, bring down the per token cost or token per dollar per million token cost uh to a place where the margins go back up.

But it's kind of a broad question in the market as uh as more and more companies lean into AI and want to deliver like the most frontier model possible to their customers. Uh no one wants to say like oh yeah we have yeah we did we we added AI to our product. It's uh GPT 3. 5. People will be like, why?

What am I paying for? What can you share about like more general business strategy at scale going forward? You're 49% of the company is owned by Meta, but you guys still have a bunch of customers, tons of revenue. You talked about two ninefigure businesses. You got a bunch of people on your team.

like h how are you kind of like rallying the team and and what are what are um what are reasons that people should join scale today and and be a part of the the opportunity you guys are going after? Yeah. Yeah. Yeah. Of course. Yeah.

It's worth taking a step back and talking about that deal for a second which is um you know Meta put in over 14 billion for 49% of the company as part of that. Alex and a and um some others went to Meta. um uh scale still has over,00 employees um and we're and and a billion dollars on the balance sheet.

Um and so we have a lot of and those employees to be clear are not none of those are contributors. Those are like people that work the contributor network. Yeah. Yeah. The contributor network has hundreds of thousands of people. Um and uh so those are people that are working at scale.

Um uh uh when I joined the company actually I joined the application side of the business. I didn't join the data side of the business when I joined as a chief strategy officer and I sort of saw the opportunity there. Um and so we're investing heavily there and customers are reserving budget heavily there.

Um and you know as an example we just uh uh got a $100 million contract with the army um to provide services and data and people to them.

Um and so you've got the data business which is supplying data to the models and then you've got the applications and services business which is making those things make making those models do things inside of large complex organizations fortune 500s US government international governments um and that this opportunity is going to be huge for scale and very few people actually know about it.

This is a multiund million business. As I said, today we've got contracts with, you know, Qar is a huge customer. The US government's a huge customer. We've got dozens of Fortune 500 companies that we work with.

Um, and so, uh, uh, I think the opportunity for scale right now is just to like capture this wave going forward because we're at the very very beginning of actually making AI work inside these big organizations.

Like yeah, I was going to ask you where where in big organizations right now, where do you think AI is overhyped and where do you think it's underhyped? We asked this to Karp yesterday. Yeah. I mean, I think I think it's very overhyped that it's going to eliminate jobs in the next like one to two years.

Like I think this line of thinking would require a change in the curve of capabilities and the change management inside of these organizations where I think it's like way overblown.

like the promise is just so high and the reality of what's going on on the ground is there's value but the value needs to be extracted and that requires a ton of work. Yeah.

Um, and so I think what's going to happen with these customers in the next year, if I had to make a prediction, is you're going to go from like having a certain amount of money, which is usually a lot in these companies, um, allocated for like AI initiatives to what is the ROI?

Like, you know, the grim reaper is going to come for every AI company that is not delivering value to these customers because because what's happened is they've all been sucked into this idea that like, oh, if we don't invest in this, we're going to miss out.

but because AGI is coming and the reality is is that if that doesn't pay off that expectation is very high. Um and if you're not finding a way to like give ground truth to these customers, they're going to be very disappointed.

So do you think we're approaching the trough of disillusionment in some of these big enterprises? I mean like I I think the gap Yeah. I mean my short answer is yeah. Um I think that the the the I've seen this in many tech cycles before. The promises are way out of ours. piece. There's still a ton of value.

Like it's not that it's like vapor. Like we are delivering value, but the effort it takes and the amount of data you need to like organize in a company and get from human beings inside the company to make sophisticated applications work is way harder than people are are selling today except for us. Yeah, makes sense.

Um h has the contributor has like the geopolitical or geo geographic footprint of the contributor network changed as you've moved towards more like expert focused uh criteria like the PhDs um is it still spread all over the world are people onboarding in Europe and America like what what is the what does the geo uh the geographic footprint look like right now?

Yeah, I mean the footprint as you would expect uh uh you know centers around areas where um PhD and masters and advanced degree um uh you know people are um and so you have a very high US footprint um you know parts of Europe uh India and Latin America there's actually I mean one interesting fact is like um there's um a lot of contributors who are who are contributing at a high level who've actually grown themselves like Like you would think it would be like if you don't have a PhD you can't do this type of work.

Yeah. And you know maybe PhD is too far of a stretch here but like you you like like you don't have domain expertise. What we've actually seen is contributors go through this whole cycle from a from a from a sophistication standpoint and we think that they're selfeaching. Yeah. Basically the find method the find teach.

Yeah. Which is kind of amazing if you think about labor markets and like you would just assume like oh the technology would move on but the people are adapting. Yeah. Uh Jordy, anything else? No, this was super super insightful. Yeah, this is great. Thank you so much for hopping on. Always welcome to join you.

You just give us a few minutes warning and we'll send you the link and you drop on. So glad glad you guys are out telling your story and congrats on all the progress. Yeah, this was fun. Thank you for the opportunity. Yeah, of course. We'll talk to you soon. Cheers, Jason. All right, cool. Take care. Bye.

Let me tell you about adquick. com. Out ofome advertising made easy and measurable. Say goodbye to the headaches of out of home advertising. Only Adqubines technology, out of home expertise, and data to enable efficient,