AI transcript
0:00:03 There will be a lot more divergence
0:00:04 between the law of the labs
0:00:07 in terms of what research directions they choose to explore
0:00:11 and which ones ultimately have breakthroughs at various times.
0:00:12 One of the hallmarks of this next phase
0:00:15 is actually going to be data production.
0:00:17 Basically no agent really works.
0:00:20 Well, it turns out there’s just no agent data on the internet.
0:00:25 The pricing for model imprints fall dramatically,
0:00:26 dramatically, dramatically, like order magnitude.
0:00:29 Yeah, orders, orders, yeah, two orders of magnitude.
0:00:30 Over two years.
0:00:35 If you’ve been listening to the A16Z podcast for a while,
0:00:37 you’ll know we talk a lot about AI.
0:00:39 We’ve covered the algorithms of power LLMs
0:00:42 and the compute required to run them.
0:00:44 But equally important is data.
0:00:46 Our guest today is as deep as you can get
0:00:50 in this world of data, the fuel behind LLMs.
0:00:52 In fact, he even recently said, quote,
0:00:55 “As an industry, we can either choose data abundance
0:00:57 or data scarcity.”
0:01:00 So what data exists today and what needs to be created,
0:01:02 either measured or synthesized?
0:01:04 Listen in to find out as I pass it over
0:01:07 to A16Z growth general partner, Sarah Wang,
0:01:09 to properly introduce this episode.
0:01:12 Hey guys, I’m Sarah Wang,
0:01:15 general partner on the A16Z growth team.
0:01:17 Welcome back to our AI revolution series
0:01:19 where we talk to industry leaders
0:01:22 about how they’re harnessing the power of generative AI.
0:01:25 Our guest this episode is Alexander Wang,
0:01:27 the founder and CEO of Scale AI,
0:01:30 a company that has become synonymous with gen AI
0:01:32 and the data needed to power advances
0:01:34 in large language models and beyond.
0:01:36 With scale’s work across enterprise,
0:01:38 automotive and the public sector,
0:01:40 Alex is also building the critical infrastructure
0:01:42 that will allow any organization
0:01:44 to use their proprietary data
0:01:47 to build the spoke gen AI applications.
0:01:48 For those of you who don’t know Alex,
0:01:51 he is one of the most impressive CEOs we’ve ever met.
0:01:53 And that’s saying something,
0:01:56 given A16Z first met Alex when he was 21
0:01:57 and already the CEO
0:02:00 of one of the fastest growing companies at its scale,
0:02:04 which he founded right before dropping out of MIT in 2016.
0:02:05 In this conversation
0:02:07 with A16Z general partner, David George,
0:02:10 Alex discusses the three pillars of AI,
0:02:12 models, compute and data,
0:02:14 and how creating abundant data
0:02:17 is core to the evolution of gen AI.
0:02:20 Alex also shares his learnings from the growth of scale,
0:02:22 his approach to leadership,
0:02:25 and what he thinks growth stage founder CEOs
0:02:27 tend to get wrong about hiring.
0:02:28 Let’s get starting.
0:02:33 As a reminder, the content here
0:02:35 is for informational purposes only,
0:02:37 should not be taken as legal, business, tax
0:02:38 or investment advice,
0:02:40 or be used to evaluate any investment or security,
0:02:42 and is not directed at any investors
0:02:45 or potential investors in any A16Z fund.
0:02:47 Please note that A16Z and its affiliates
0:02:48 may also maintain investments
0:02:51 in the companies discussed in this podcast.
0:02:53 For more details, including a link to our investments,
0:02:57 please see a16z.com/disclosures.
0:02:59 (upbeat music)
0:03:04 – We’re very excited today to have Alex Wang,
0:03:07 the founder and CEO of Scale AI with us.
0:03:08 Alex, thanks for being here.
0:03:09 – Thanks for having me.
0:03:12 – I always love talking to you and I always learn a ton,
0:03:13 but maybe to start,
0:03:14 why don’t you just tell us a little bit about
0:03:17 what you’re building at Scale AI and then we’ll dive in.
0:03:18 – Yeah, so at Scale,
0:03:20 we’re building the data foundry for AI.
0:03:21 So, taking a step back,
0:03:24 AI boils down to three pillars.
0:03:26 All the progress we’ve seen has come from
0:03:28 compute data and algorithms,
0:03:30 and the progress among all three of these pillars.
0:03:32 Compute has been powered by folks like NVIDIA,
0:03:34 the algorithmic advancements have been led
0:03:36 by the large labs like OpenAI and others,
0:03:38 and data is fueled by Scale.
0:03:42 And so our goal is to produce the frontier data necessary
0:03:44 to fuel frontier level advancements
0:03:46 in partnership with all the large labs,
0:03:49 as well as enable every enterprise and government
0:03:51 to make use of their own proprietary data
0:03:53 to fuel their frontier AI development.
0:03:55 – So on this topic of frontier data,
0:03:58 practically, but how do you actually get it?
0:04:01 – Yeah, I think this will be one of the great
0:04:04 human projects of our time, if that makes sense.
0:04:08 And I think that the only model that we have in the world
0:04:10 for the level of intelligence that we seek to create
0:04:11 is humanity.
0:04:14 And so the production of frontier data looks a lot like
0:04:17 a sort of marriage between human experts and humanity
0:04:20 with technical and algorithmic techniques around the models
0:04:23 to produce huge amounts of this kind of data.
0:04:26 And by the way, all the data that we’ve produced today,
0:04:28 the internet has looked like that too.
0:04:30 The internet in many ways is this like collaboration
0:04:32 between machines and humans to produce
0:04:35 large amounts of content and data.
0:04:37 It’ll look like the internet on steroids.
0:04:40 What happens if the internet basically,
0:04:43 instead of just being a human entertainment device
0:04:46 with this like byproduct of data generation,
0:04:48 what if it were just this large scale
0:04:49 data generation experiment?
0:04:52 – So you have a very unique perspective
0:04:53 into the state of the industry.
0:04:56 So how would you characterize the state of models,
0:04:58 the language models right now?
0:05:00 And I’d love to sort of get into things
0:05:01 like market structure,
0:05:04 but just sort of what’s the state of the industry right now?
0:05:06 – Yeah, I think we’re sort of closing in
0:05:11 at the end of maybe phase two of language model development.
0:05:14 I think phase one was the early years
0:05:15 of almost like pure research.
0:05:19 So phase one hallmarks are the original transformer paper,
0:05:22 the original small scale experiments on GPTs
0:05:25 all the way leading up probably until like GPT three
0:05:27 was this sort of phase one all research,
0:05:30 very, very focused on sort of like small scale tinkering
0:05:32 and algorithmic advancements.
0:05:36 And then phase two, which is sort of maybe GPT three till now
0:05:39 is really the sort of like initial scaling phase.
0:05:42 So we had GPT three that worked pretty well
0:05:44 and then opening eyes to start with really scaled up
0:05:46 these models to GPT four and beyond.
0:05:51 And then many companies, Google, Anthropic, Meta, XAI now,
0:05:56 many, many companies have also joined on this sort of race
0:05:58 to scale up these models to incredible capabilities.
0:06:01 So I think for the past, let’s say three years,
0:06:04 it’s almost been more about execution than anything.
0:06:05 It’s a lot of just engineering,
0:06:07 like how do you actually have large scale training,
0:06:08 work well?
0:06:10 How do you make sure there aren’t weird bugs in your code?
0:06:12 How do you set up the larger clusters?
0:06:14 A lot of execution will work to get to where we are now,
0:06:17 where we have kind of a number of very advanced models.
0:06:19 And then I think we’re entering a phase
0:06:21 where the research is gonna start mattering a lot more.
0:06:24 Like I think there will be a lot more divergence
0:06:25 between a lot of the labs
0:06:28 in terms of what research directions they choose to explore
0:06:31 and which ones ultimately have breakthroughs at various times.
0:06:33 And it’s sort of an exciting alternating phase
0:06:35 between maybe just raw execution
0:06:39 versus sort of a more innovation powered cycle.
0:06:41 – They’ve kind of gotten to a point
0:06:44 where I wouldn’t say there’s like abundant compute
0:06:46 but they’ve had enough compute that they’ve needed
0:06:48 in order to get to the models where they’re at.
0:06:50 That’s not a constraint necessarily.
0:06:53 They’ve kind of exhausted as much data as they possibly can
0:06:55 all of the frontier labs.
0:06:58 And so the next thing will be breakthroughs on that
0:07:00 and then advancing the ball on the data side.
0:07:01 Is that fair?
0:07:02 – Yeah, and I think basically,
0:07:03 yeah, if you look at the pillars,
0:07:05 compute were obviously continuing
0:07:06 to scale up the training clusters.
0:07:10 So I think that direction is pretty clear on the algorithms.
0:07:12 I think there has to be a lot of innovation there.
0:07:14 Frankly, I think that’s where a lot of the labs
0:07:15 are really working hard,
0:07:16 I think on the pure research of that.
0:07:18 And then data, you can have alluded to it,
0:07:22 we’ve kind of run out of all the easily accessible
0:07:24 and easily available data out there.
0:07:25 And–
0:07:25 – Yeah, Common Core is all done.
0:07:27 Everybody’s got the same access to it.
0:07:28 – Yeah, exactly.
0:07:30 And so a lot of people talking about this is the data wall.
0:07:31 You know, we’re kind of hitting this wall
0:07:33 where we’ve leveraged all the publicly available data.
0:07:35 And so one of the hallmarks of this next phase
0:07:37 is actually going to be data production.
0:07:40 And what is the method that each of these labs
0:07:42 is going to use to actually generate the data necessary
0:07:45 to get you to the next levels of intelligence
0:07:47 and how do we get towards data abundance?
0:07:50 And I think this is going to require a number of fields
0:07:53 of sort of advanced work and advanced study.
0:07:55 I think the first is really pushing
0:07:57 on the complexity of the data.
0:07:59 So moving towards frontier data.
0:08:01 So a lot of the capabilities
0:08:03 that we want to build into the models,
0:08:05 the biggest blocker is actually a lack of data.
0:08:08 So for example, agents has been the buzzword
0:08:10 for the past two years
0:08:12 and basically no agent really works.
0:08:14 Well, it turns out there’s just no agent data
0:08:15 on the internet.
0:08:17 There’s no just pool of really valuable agent data
0:08:18 that’s just sitting around anywhere.
0:08:20 And so we have to figure out how to produce
0:08:21 really high quality agent data.
0:08:23 – Give an example of like, what would you have to produce?
0:08:26 – So we have some work coming out on this soon
0:08:28 which demonstrates that right now,
0:08:30 if you look at all the frontier models,
0:08:32 they suck at composing tools.
0:08:36 So if they have to use one tool and then another tool,
0:08:38 let’s say they have to look something up
0:08:39 and then write a little Python script
0:08:41 and then chart something.
0:08:43 They use multiple tools in a row, they just suck at that.
0:08:45 They just are really, really bad
0:08:47 at utilizing multiple tools in a row.
0:08:49 And that’s something that’s actually very natural
0:08:50 for humans to do.
0:08:52 – Yeah, but it’s not captured anywhere, right?
0:08:53 That’s the point, right?
0:08:53 – Exactly.
0:08:55 – So you can’t actually go take the capture
0:08:57 of somebody going from one window to another
0:08:59 into a different application
0:09:01 and then feed that to the model so it learns, right?
0:09:02 – Exactly, yeah, yeah.
0:09:05 So these sort of reasoning chains through
0:09:07 when humans are solving complex problems,
0:09:09 we naturally will use a bunch of tools,
0:09:10 we’ll think about things,
0:09:12 we’ll reason through what needs to happen next,
0:09:14 we’ll hit errors and failures
0:09:16 and then we’ll go back and sort of like reconsider.
0:09:17 You know, a lot of these reasoning chains,
0:09:19 these agentic chains are,
0:09:21 the data just doesn’t exist today.
0:09:23 So that’s an example of something that needs to be produced,
0:09:24 but taking a big step back,
0:09:26 what needs to happen on data.
0:09:28 First is increasing data complexity,
0:09:29 so moving towards frontier data.
0:09:31 The second is just data abundance,
0:09:32 increasing the data production.
0:09:35 – Capturing more of what humans actually do
0:09:36 in the field of work.
0:09:37 – Yeah, both capturing more of what humans do
0:09:40 and I think investing into things like synthetic data,
0:09:42 hybrid data, so utilizing synthetic data,
0:09:44 but having humans be a part of that loop
0:09:46 so that you can generate much more high quality data.
0:09:48 We need basically just in the same way,
0:09:51 I think with chips, we talk a lot about chip foundries
0:09:52 and how do we ensure that we have like
0:09:54 enough means of production chips.
0:09:55 And the same thing is true for data,
0:09:57 we need to have effectively data foundries
0:10:00 and the ability to generate huge amounts of data
0:10:02 to fuel the training of these models.
0:10:05 And then I think the last leg of the stool,
0:10:08 which is often rated as measurement of the models
0:10:10 and ensuring that we actually have,
0:10:13 I think for a while the industry is just sort of like,
0:10:14 oh yeah, we just add a bunch more data
0:10:15 and we see how good the model is
0:10:16 and we add a bunch more data
0:10:17 and we see how good the model is,
0:10:19 but we’re gonna have to get pretty scientific
0:10:22 around exactly what is the model not capable of today
0:10:24 and therefore what are the exact kinds of data
0:10:26 that need to be added to improve the model’s performance.
0:10:29 – How much of an advantage do the big tech companies have
0:10:33 with their corpus of data versus the independent labs?
0:10:35 – Yeah, well there’s a lot of regulatory issues
0:10:38 that they have with utilizing their existing data corpuses.
0:10:40 You know, you can look through,
0:10:41 this is before all this generative AI work,
0:10:44 but at one point Meta did some research
0:10:47 that utilized basically all the public Instagram photos
0:10:49 along with their hashtags
0:10:51 to train really good image recognition algorithms.
0:10:54 They had a lot of regulatory problems with that in Europe.
0:10:56 Like it turned out to be a huge pain in the ass.
0:10:58 So I think that that’s one thing
0:11:01 that’s kind of difficult to reason through,
0:11:03 which is to what degree from a regulatory perspective,
0:11:04 particularly in Europe,
0:11:05 these companies are going to be able
0:11:07 to utilize their data advantages.
0:11:09 So I think that was kind of TBD.
0:11:12 I think that the real way in which a lot of the large labs
0:11:16 have just dramatic advantages is just,
0:11:18 they have very profitable businesses
0:11:22 that can provide near infinite sources of capital
0:11:24 for these AI efforts.
0:11:25 And I think that that’s something
0:11:27 that I’m watching pretty intently
0:11:29 or I’m very curious to see how it plays out.
0:11:31 – Hey, it’s Steph.
0:11:34 You might know that before my time at A16C,
0:11:36 I used to work at a company called The Hustle.
0:11:38 And then we were acquired by HubSpot
0:11:41 where I helped build their podcast network.
0:11:42 While I’m not there anymore,
0:11:45 I’m still a big fan of HubSpot podcasts,
0:11:47 especially My First Million.
0:11:49 In fact, I’ve listened to pretty much
0:11:51 all 600 of their episodes.
0:11:53 My First Million is perfect for those of you
0:11:55 who are always trying to stay ahead of the curve
0:11:56 or in some cases,
0:11:58 take matters into your own hands
0:12:00 by building the future yourself.
0:12:03 Posted by my friends, Sam Parr and Sean Curry,
0:12:05 who have each built and sold eight-figure businesses
0:12:07 to Amazon and HubSpot,
0:12:08 the show explores business ideas
0:12:10 that you can start tomorrow.
0:12:12 Plus, Sam and Sean jam alongside guests
0:12:15 like Mr. Beast, Rob Dyrdek, Tim Ferriss,
0:12:18 and every so often, you’ll even find me there.
0:12:20 From gas station pizza and egg carton businesses
0:12:23 doing millions all the way up to several guests
0:12:25 making their first billion.
0:12:27 Go check out My First Million
0:12:28 wherever you get your podcasts.
0:12:31 (upbeat music)
0:12:37 – There’s this whole question,
0:12:39 the industry is like, are they over-investing?
0:12:40 And if you listen to their earnings calls
0:12:41 of the big tech companies,
0:12:43 they’re like, look, our risk is under-investing,
0:12:44 not over-investing.
0:12:45 What do you make of that?
0:12:46 – Yeah, I mean, if you think about,
0:12:50 let’s take the incentives of any one of the CEOs of the,
0:12:53 put yourself in the shoes of Sundar Pachai
0:12:54 or Mark Zuckerberg or whatnot.
0:12:55 – Or Satya, yeah.
0:12:56 – Or Satya.
0:13:00 And to your point, if they really nail this AI thing,
0:13:03 they could generate another trillion dollars of market cap
0:13:04 probably very easily.
0:13:06 If they really are ahead of the competition
0:13:07 and they productize in a good way,
0:13:10 like trillion dollars of market cap, kind of no-brainer.
0:13:12 And if they don’t invest the extra,
0:13:15 whatever it is, 20 or 30 billion of CapEx per year,
0:13:16 and they miss out on that.
0:13:19 And then there’s some real existential risk, I think too,
0:13:20 for each of the large–
0:13:21 – Yeah, in their each formula.
0:13:24 – Yeah, all their businesses are potentially
0:13:26 deeply disruptible by AI technology.
0:13:29 So the risk reward for them is very obvious.
0:13:31 So that’s, I think, the big picture thinking.
0:13:32 And then from a more tactical level,
0:13:35 I think all of them are gonna be able to
0:13:38 pretty easily recoup their capital investments
0:13:41 just by worst case making their core businesses
0:13:42 more efficient and effective.
0:13:44 So for example, like, you know.
0:13:47 – Yeah, GPU utilization for Facebook advertising.
0:13:48 – Yeah, Facebook, Google,
0:13:50 they make their advertising systems a little bit better.
0:13:52 They can recoup billions of dollars just by–
0:13:53 – Yeah, better performance.
0:13:55 – Yeah, better performance there.
0:13:56 Apple can easily recoup the investments
0:13:58 if it drives an upgrade cycle.
0:13:59 I mean, these are things
0:14:00 that I think are pretty clear.
0:14:02 – Look, it’s generally great for the industry
0:14:04 that they are investing so much capital
0:14:05 because they also are in the business
0:14:06 of renting this compute out,
0:14:08 or at least in the case of Google and Microsoft, they are.
0:14:10 – And the models are making their way,
0:14:12 like, Lama 3.1 is open source.
0:14:16 And so even the literal fruits of all the investment
0:14:18 are becoming broadly accessible.
0:14:19 And so the surplus generated
0:14:23 from the open source and these models is kind of insane.
0:14:23 – It’s insane.
0:14:26 And that’s a great segue into market structure
0:14:27 at the model layer.
0:14:29 So what do you think actually happens?
0:14:32 Are there the few players that we’ve all identified now,
0:14:34 the handful, and they all compete?
0:14:37 Do you think it’s a profitable business?
0:14:39 What impact does open source have
0:14:40 on the quality of the businesses?
0:14:43 Take us a couple of years ahead and give us your forecast.
0:14:45 – Yes, we’ve seen over the past,
0:14:47 even just like year and a half,
0:14:50 the pricing for model inference
0:14:53 fall dramatically dramatically dramatically.
0:14:54 – Right, order of magnitude.
0:14:55 – Yeah, two orders of magnitude.
0:14:56 – Two orders of magnitude.
0:14:57 – Over two years.
0:14:58 And so it’s this shocking thing
0:15:01 that it turns out intelligence might be a commodity.
0:15:04 But no, I mean, I think that this huge sort of lack
0:15:07 of pricing power, let’s say, on the pure model layer,
0:15:12 certainly indicates that renting models out on their own
0:15:14 may or may not be the best long-term business.
0:15:17 I think it’s likely to be a relatively mediocre
0:15:18 long-term business.
0:15:20 – Well, I guess it depends on the breakthrough thing,
0:15:22 which is the earlier point, right?
0:15:24 To the extent that someone actually has a durable breakthrough
0:15:26 or multiple people have durable breakthroughs,
0:15:28 like then potentially market structure is different.
0:15:31 – So two things, if meta continues open sourcing,
0:15:34 that puts a pretty strong cap as to the value
0:15:35 that you can get from the model layer.
0:15:39 And then two, if at least a handful of the labs
0:15:42 are able to have similar performance over time,
0:15:45 then that also dramatically changes the pricing equation.
0:15:47 So we think that it’s not 100%,
0:15:51 but chances are the pure model renting business
0:15:52 is not the highest quality business,
0:15:54 where there are much higher quality businesses
0:15:56 are going to be above and below.
0:16:01 So below, I mean, NVIDIA is obviously an incredible business,
0:16:03 but the clouds also have really great businesses too,
0:16:05 because it turns out it’s pretty hard logistically
0:16:08 to actually set up large clusters of GPUs.
0:16:10 And so the cloud providers actually have pretty good margins
0:16:11 when they rent out.
0:16:13 – And the traditional data center business
0:16:14 is very much a scale game.
0:16:15 – Yep.
0:16:17 – So they are massively benefited
0:16:18 relative to smaller players.
0:16:19 – Yeah, exactly.
0:16:20 So I think picks and shovels,
0:16:22 so if you’re under the model layer,
0:16:23 I think there’s great businesses there.
0:16:24 And if you’re above the model layer,
0:16:26 if you’re building applications,
0:16:27 ChatJPT is a great business.
0:16:31 And a lot of the apps in the startup realm
0:16:32 actually are working pretty well.
0:16:34 I mean, none of them are quite as big as ChatJPT, obviously,
0:16:36 but a lot of apps,
0:16:38 if they nail the early product market fit,
0:16:40 end up being pretty good businesses,
0:16:41 great businesses as well,
0:16:44 because the value that they generate for customers,
0:16:47 if they get the whole user experience correct,
0:16:49 far exceeds the inference cost of the models.
0:16:50 – There’s some cool stuff here, right?
0:16:54 I think an Anthropics launch of artifacts in Claude,
0:16:57 it’s like the first pin drop of this major theme
0:17:00 of all the labs are gonna be pushing
0:17:02 much deeper product integrations
0:17:04 to be able to drive higher quality businesses.
0:17:05 So that’ll be the other story is,
0:17:07 I think we’re gonna see a lot of iteration
0:17:10 at the product layer and the product level.
0:17:13 The sort of boring chatbots is not gonna be the end product.
0:17:14 That’s not the end all be all.
0:17:15 – Disappointing outcome.
0:17:16 – Yeah, exactly.
0:17:19 And product iteration and the product innovation cycles
0:17:21 is very hard to predict because,
0:17:23 I mean, OpenAI was surprised how popular chatGPD was.
0:17:25 I don’t think it’s like super obvious to me
0:17:27 or anyone in the industry, frankly,
0:17:30 what exact products are gonna be the ones that hit
0:17:32 and what’s gonna provide the next legs of growth.
0:17:36 But you have to believe that an OpenAI or an Anthropic
0:17:38 can build great applications businesses
0:17:41 to for them to be long-term independent and sustainable.
0:17:43 – Yeah, for sure.
0:17:45 Yeah, and then it’s what drives competitive advantage.
0:17:46 Obviously you have the model,
0:17:48 a tightly integrated product on top of it,
0:17:51 and then the good old fashioned modes from there.
0:17:51 – Yeah.
0:17:54 – Workflows, integrations, all that stuff.
0:17:56 – I think you can clearly see they’re thinking on it.
0:17:58 I mean, both OpenAI and Anthropic
0:17:59 hired chief product officers
0:18:01 within two months of each other.
0:18:02 – Yeah, they’re figuring it out.
0:18:04 – And then it’s sort of a change of tune
0:18:07 where they’re like, “Oh no, we’re very purely focused on this
0:18:09 “and it’s okay, I think there’s the realization to it.”
0:18:11 So, yeah, exactly, it makes a whole sense.
0:18:12 You’ve got an application business
0:18:14 with some really interesting customers.
0:18:15 What are you hearing from enterprises
0:18:18 as to how they’re actually putting this into place?
0:18:20 – I think what we’ve seen is
0:18:22 there was a huge amount of excitement from the enterprise.
0:18:24 A lot of enterprises were like,
0:18:26 “Shit, we have to start doing something.
0:18:27 “We have to get ahead of this.
0:18:29 “We have to start experimenting with AI.”
0:18:33 I think that that led them to this fast POC cycle
0:18:35 where they’re like, “Okay, where are all the low hanging
0:18:36 “fruit ideas that we have?”
0:18:37 – Go buy AI stuff.
0:18:38 – Yeah, yeah.
0:18:39 – And let’s go try all of it.
0:18:42 And some of those things are good, some of them aren’t good,
0:18:45 but I think regardless, it’s been this big frenzy,
0:18:48 much fewer of the POCs have made it to production
0:18:50 than I think the industry overall expected.
0:18:52 And I think a lot of enterprises are looking at now
0:18:55 and the doomsday that they thought might have happened
0:18:56 hasn’t really happened.
0:19:00 AI has not fully terraformed and transformed
0:19:01 most of the major industries.
0:19:03 Like it’s not like totally, you know–
0:19:04 – It’s sort of marginal stuff.
0:19:06 It’s like efficiency gains and support
0:19:08 and then some of the creative tasks and things like that.
0:19:09 – Yeah, exactly.
0:19:10 – Otherwise it’s pretty light.
0:19:12 The thing that we think a lot about is like,
0:19:15 what AI improvements or AI transformations or AI efforts
0:19:17 that we’re working on actually can
0:19:20 meaningfully drive the stock price
0:19:22 of the companies that we’re working on.
0:19:24 And so that’s what we encourage all of our customers
0:19:25 to really be thinking about
0:19:28 because at the end of the day, the potential is there.
0:19:31 There’s latent potential for almost every enterprise
0:19:33 to implement AI at a level
0:19:35 that would meaningfully boost their stock price.
0:19:37 – Mostly in the form of cost savings.
0:19:38 – Efficiency gains.
0:19:39 – Well today in the form of cost savings,
0:19:41 but then also much better customer experiences.
0:19:43 Like I think in a lot of industries
0:19:46 where there’s a lot more manual interaction with customers,
0:19:48 you should be able to drive much better customer interactions
0:19:49 if you have more standardization
0:19:51 and you were able to use more automation.
0:19:54 And then those eventually would make their way to gains
0:19:56 of market share with respect to competitors.
0:19:59 So that’s what we’re pushing our customers towards.
0:20:01 And I see it, some of the CEOs that we work with,
0:20:02 they’re all on board.
0:20:04 And they understand that it’s gonna be
0:20:05 a multi-year investment cycle.
0:20:08 They might not see gains the next quarter,
0:20:10 but if they actually pull through the other side,
0:20:12 they’re gonna see massive transformations.
0:20:15 I think that a lot of the frenzy around small use cases
0:20:17 and sort of the more marginal use cases,
0:20:18 I think that’s good.
0:20:19 I think it’s exciting.
0:20:20 I think they should be doing it.
0:20:23 But to me, that’s not what we’re all here to do.
0:20:24 – Yeah, it’s very much like the application layer
0:20:26 is like very much like phase one right now,
0:20:28 which is, I mean, yeah, there’s some automation,
0:20:30 but it’s largely like chatbots.
0:20:34 My hope as a startup investor is that over time,
0:20:36 there’s a window that opens for the startups
0:20:39 where product innovation will help them to win
0:20:40 and beat the incumbents.
0:20:42 Like my partner, Alex Rampel, has this phrase,
0:20:44 which is, is the startup gonna get to distribution
0:20:46 before the incumbent finds innovation.
0:20:48 And I think there’s an opportunity for it,
0:20:50 but it’s like the tech is too early right now.
0:20:51 – Yeah.
0:20:52 – I don’t know if you would agree with that, but…
0:20:55 – I think the tech is too early to imagine,
0:20:57 yeah, again, because it’s mostly cost saving.
0:20:59 I think if most of the benefit is on the cost saving side,
0:21:02 then that’s not really enough to disrupt large incumbents
0:21:04 that has already kind of like pushed their way through
0:21:07 all the costs of growing and distribution.
0:21:09 – How valuable do you think is the data inside of enterprises?
0:21:11 Like you said, JP Morgan has whatever,
0:21:14 15 petabytes of data or something like that.
0:21:17 I never wrote the numbers, but like, is that overrated?
0:21:19 How much of it is actually useful?
0:21:22 Because to date, most of that data has not given them
0:21:24 some meaningful competitive advantage.
0:21:26 So do you think that actually changes?
0:21:28 – I think AI is the first time you could see
0:21:30 that potentially change because basically,
0:21:32 obviously there’s the whole big data wave.
0:21:34 Big data boils down to better analytics,
0:21:36 which is helpful, like marginally helpful
0:21:39 for business decision making, but not deeply transformation.
0:21:41 – It does massively change the way the products work.
0:21:42 – Yeah, exactly.
0:21:45 Whereas now you actually can imagine
0:21:47 some massive transformation in the way the products work.
0:21:49 So let’s take any big bank.
0:21:54 A lot of the valuable interactions between a user
0:21:57 and a large bank like a JP Morgan or Morgan Stanley or whatnot
0:21:59 are human driven, are people driven.
0:22:02 And they try their best to ensure that the quality
0:22:03 of experience is very high across the board,
0:22:06 but obviously with any large process,
0:22:07 there’s only so much you can do to assure that.
0:22:10 But all of your prior customer interactions
0:22:12 and all the ways in which your business
0:22:16 has worked historically is the only available data
0:22:18 to be able to train models to do well
0:22:19 at this particular task.
0:22:21 And if you think about like wealth management,
0:22:23 there’s very little indistribution data of that
0:22:25 on the internet that you could trade a model off of.
0:22:26 – So there’s behind the walls,
0:22:28 there’s actually quite a bit, it’s very rich.
0:22:29 – Yeah, huge amounts of data.
0:22:31 So I think that a lot of the data is probably
0:22:33 not super relevant to actually transforming your business,
0:22:35 but some of the data is hyper valuable.
0:22:38 So I think enterprises have a lot of trouble
0:22:41 and challenge around actually utilizing
0:22:43 any amount of data that they have.
0:22:46 It’s poorly organized, it’s sort of all over the place,
0:22:49 they pay consulting firms tens of millions of dollars,
0:22:51 hundreds of millions of dollars to do these data migrations.
0:22:53 And it’s even after that no change in results.
0:22:54 Yeah, no change results.
0:22:57 So I think it’s historically very difficult place
0:22:59 for enterprises to really drive transformation.
0:23:01 And so in some ways this is the race,
0:23:04 are they gonna be able to figure out how to utilize
0:23:08 and leverage their data faster than some startup
0:23:10 figures out how to somehow get access to data?
0:23:11 – Create a massively different product
0:23:13 with a little bit subset of the data.
0:23:14 – Yeah, exactly.
0:23:17 – Shifting gears to how you run your company
0:23:19 and how you built your company.
0:23:21 One of the things that you’ve talked about
0:23:25 is a mistake that you made during the go times
0:23:29 of 2020 and 2021 around hiring.
0:23:31 And this notion that in order to scale,
0:23:32 you had to hire a ton.
0:23:35 And it’s something we saw with all of our portfolio companies
0:23:37 was like, hey, there’s war for talent.
0:23:39 And it meant that we gotta go higher,
0:23:40 we gotta go higher, we gotta go higher.
0:23:42 So what were the lessons that you learned
0:23:44 through that process and then how have you changed
0:23:45 how you’ve done things afterwards?
0:23:48 – So over the past few years,
0:23:51 we’ve basically kept our headcount flat.
0:23:54 I mean, we’ve grown it very slightly as the business grown,
0:23:57 but the business itself is five X, well, six X.
0:23:58 You know, the business has grown dramatically.
0:24:02 And the takeaway from this entire process is
0:24:07 it feels very logical that more people equals better results
0:24:10 and more people equals more stuff being done,
0:24:12 but rather paradoxically, I think,
0:24:14 if you have a very high performing team
0:24:15 and a very high performing org,
0:24:18 it’s nearly impossible to grow it dramatically
0:24:21 without losing all of that high performance
0:24:22 and all of the winning culture.
0:24:24 – Yeah, reducing the communication and coordination overhead
0:24:26 actually increases productivity.
0:24:27 – That’s definitely true.
0:24:29 And I think it’s actually something even deeper,
0:24:31 which is that a very high performing team
0:24:33 of a certain size is almost like
0:24:36 this very intricate sculpture
0:24:38 and this interplay between all the people in the team.
0:24:40 And if you just add a bunch of people into that,
0:24:42 even if the people are great,
0:24:43 like it just screws the whole thing up.
0:24:46 And no matter what, as you add people,
0:24:47 you’re gonna have regression to the mean.
0:24:49 You know, if you kind of observe companies
0:24:50 that do scale healthcare a lot
0:24:52 and that’s pretty core to their financial results,
0:24:54 I think they acknowledge that regression,
0:24:55 that mean regression.
0:24:57 So if you think about like the scaling
0:24:58 of large sales teams, for example.
0:24:59 – Yeah, sure.
0:25:00 – You acknowledge that you’re going to have
0:25:02 that mean regression, but you just operationalize
0:25:04 so that you’re like a little bit above the mean.
0:25:05 And if you’re able to do that,
0:25:07 then the whole equation still works financially.
0:25:09 – Yeah, I’d say sales is different than product.
0:25:11 – Yeah, totally, of course.
0:25:13 But our observation is just startups work
0:25:15 because you have very high performing teams
0:25:17 and you wanna keep those high performing teams
0:25:19 intact as long as you possibly can.
0:25:22 You know, I think a common startup failure mode
0:25:24 is that you have something that works,
0:25:26 but everybody in the company is really junior.
0:25:28 So then things are scaling,
0:25:30 but all the wheels are kind of falling off.
0:25:32 Your investors tell you how you should hire
0:25:34 some executives, you go through these searches
0:25:37 that are somehow uniquely soul crushing every time.
0:25:40 But you go through, if you’re graded in,
0:25:41 it works half the time.
0:25:43 – Yeah, yeah.
0:25:44 – So you go through the exact searches,
0:25:47 you’re bringing exact and then you give the exact
0:25:49 a lot of rope and your exact say,
0:25:51 hey, we need to hire a massive team
0:25:52 for us to hit our results.
0:25:55 And you’re like, yeah, I mean, I’m pretty experienced,
0:25:57 you seem really experienced, let’s do what you say.
0:25:59 And you let these big teams sort of be built.
0:26:02 And the reality is I think this almost always
0:26:03 results in ruin.
0:26:05 I think that this isn’t to say that you can’t hire executives
0:26:07 from the outside, but I think what you need to do
0:26:09 when you hire executives from the outside
0:26:11 is they really get steeped in how the company works.
0:26:15 And before they make any major sweeping suggestions,
0:26:18 they get into the rhythm and the operations of the company.
0:26:20 And they understand why does the whole thing
0:26:21 work in the first place?
0:26:23 Why are the things that are working working?
0:26:25 And then they make thoughtful suggestions.
0:26:27 Initially, they take small steps
0:26:29 and you sort of like you trust and verify
0:26:30 each of these small steps.
0:26:33 And eventually maybe they can make more sweeping suggestions,
0:26:35 but it should be at a point where they have
0:26:38 a clear track record of making small steps
0:26:38 that have been really beneficial.
0:26:39 – Oh, that’s interesting.
0:26:41 That’s interesting and very tangible, right?
0:26:43 It starts small when you hire a big executive
0:26:44 and it’s a little bit counterintuitive.
0:26:46 And it’s not the way that any of those executives
0:26:48 want to go.
0:26:50 – Yeah, I think that there’s kind of an exec fantasy
0:26:52 that I’ve noticed, which is, and by the way,
0:26:54 I think executives are great people
0:26:55 and they’re like, they’re incredible.
0:26:58 But there is a tendency for an executive fantasy,
0:27:00 particularly for like Silicon Valley companies
0:27:02 with young founders and whatnot, which is,
0:27:05 oh, I’m gonna come in and I’m gonna fix this whole thing.
0:27:08 I’m gonna make this a professional operation.
0:27:10 You’re recruiting teammates at the end of the day.
0:27:12 You’re not recruiting like some magic wand.
0:27:14 You’re recruiting a teammate who you believe
0:27:17 over an extended period of time is gonna have great judgment
0:27:20 in making repeated decisions about the business.
0:27:21 But, and this is where we’ve made mistakes.
0:27:24 It’s like, you’re not buying some magical bag of goods
0:27:27 that is gonna bring this magic formula into your business
0:27:29 that will all of a sudden make the whole thing work.
0:27:31 On the flip side, there’s a founder fantasy.
0:27:34 The founder fantasy, or the founder CEO fantasy,
0:27:37 which is, oh, I’m gonna just hire a bunch
0:27:39 of incredible execs, throw over me fucking pros.
0:27:41 And then I’m gonna go–
0:27:43 – They’ll do the stuff I don’t wanna do.
0:27:47 They’ll do all the stuff I don’t wanna do.
0:27:48 And I’m gonna be able to sit back
0:27:50 and watch the machine work.
0:27:52 And that’s also extremely unrealistic
0:27:53 because the flip side is also true.
0:27:55 The reason that you are a good founder CEO
0:27:57 is because you make very good decisions
0:27:59 over and over again over an extended period of time.
0:28:02 And to pull yourself out of those decision-making loops
0:28:04 would be kind of crazy.
0:28:05 – That’s a pattern we’ve seen a lot,
0:28:07 which is I’m gonna hire executives.
0:28:08 I’m gonna step back a bit.
0:28:10 And then it’s, oh, shit, realization
0:28:11 that like, hey, some big decisions go wrong
0:28:13 and wait, this is the point of me being here.
0:28:15 – Yeah, I think it can work
0:28:19 if your industry is very stable, potentially.
0:28:20 – Well, look at any public company
0:28:24 when they change CEOs and the stock price moves like 2%.
0:28:26 And it’s like, oh, okay, well, actually,
0:28:27 it doesn’t really matter.
0:28:29 That is a cog, but that is very different
0:28:32 from a high-growth startup that’s run by a founder.
0:28:33 – Exactly, yeah, yeah.
0:28:35 And I think that a lot of startups
0:28:36 and a lot of companies are valuable
0:28:38 because of an innovation premium.
0:28:39 – 100%.
0:28:41 – Investors believe that founder-led companies
0:28:43 are going to out-innovate the market.
0:28:47 And so your job is to out-innovate the market, so.
0:28:49 – You better be in the strategic decisions.
0:28:50 Yeah, for sure.
0:28:51 How about MEI?
0:28:55 So you recently rolled out this concept.
0:28:59 I think like half of my ex-feed was praising you
0:29:01 and that’s probably more than half.
0:29:03 Some portion of my ex-feed was yelling at you,
0:29:04 talking about the concept
0:29:06 and what are your observations of rolling it out so far?
0:29:10 – Yeah, so MEI, we basically rolled out this idea
0:29:12 of merit, excellence, and intelligence.
0:29:15 And the basic idea is in every role
0:29:18 we’re gonna hire the best possible person
0:29:20 regardless of their demographics.
0:29:23 And we’re not going to do any sort of quota-based
0:29:25 optimization of our workforce
0:29:28 to meet certain demographic targets.
0:29:29 That doesn’t mean we don’t care about diversity.
0:29:33 We actually care about having diverse pipelines
0:29:35 and diverse top-of-funnel for all of our roles.
0:29:37 But at the end of the day,
0:29:39 the best, most capable person for every job
0:29:41 is going to be the one that we hire.
0:29:45 It’s one of these things that was mildly controversial,
0:29:50 but I think it’s also, if we were to just take a big step back
0:29:53 as to who should companies be hiring,
0:29:54 I think it’s kind of an obvious-
0:29:56 – Sort of common sense, yeah.
0:29:57 – Yeah, it feels kind of obvious that-
0:29:58 – Lost the plot, yeah.
0:30:00 – Companies should hire the most talented people.
0:30:02 And I think there’s obviously this became this big question
0:30:05 of like how much social responsibility do companies have
0:30:06 in what they do?
0:30:09 My take is I operate in a very competitive industry.
0:30:13 Scale’s role is to help fuel artificial intelligence.
0:30:14 It’s very important technology.
0:30:17 We need incredibly smart people to be able to do this,
0:30:20 and we need the best people to be able to accomplish this.
0:30:22 I think that this is something that,
0:30:24 I think most people at scale would say
0:30:26 was sort of like implicitly true,
0:30:27 or sort of it wasn’t like a departure
0:30:31 from how many of us thought of what we do at scale.
0:30:33 But it was really valuable for us to codify it
0:30:34 because it gives everybody confidence
0:30:37 that even if this is how we operate today,
0:30:38 companies change over time,
0:30:40 we’re not gonna change this quality.
0:30:41 – Well, this has been awesome.
0:30:45 I wanna close with an optimistic question and forecast,
0:30:49 which is what is your sort of own view of,
0:30:50 or definition of AGI?
0:30:54 And what is your expected timeline to when we reach that?
0:30:56 – Yeah, I like the definition of this that’s sort of like,
0:31:00 let’s say 80 plus percent of jobs
0:31:02 that people can do purely at computers,
0:31:03 so digital focused jobs.
0:31:05 AI can accomplish those jobs.
0:31:07 It’s not like imminent.
0:31:09 It’s not like immediately on the horizon,
0:31:12 so on the order of four plus years,
0:31:14 but you can see the glimmers,
0:31:17 and depending on the algorithmic innovation cycle
0:31:18 that we talked about before,
0:31:20 could make that much sooner.
0:31:21 Yeah, that’s awesome.
0:31:22 – Very exciting.
0:31:23 Well, Alex, thanks for being here.
0:31:24 Great to chat with you as always.
0:31:26 Learned a ton.
0:31:27 Really appreciate it.
0:31:28 Yeah, thanks for having me.
0:31:32 – All right, that is all for today.
0:31:34 If you did make it this far, first of all, thank you.
0:31:36 We put a lot of thought into each of these episodes,
0:31:38 whether it’s guests, the calendar Tetris,
0:31:40 the cycles with our amazing editor, Tommy,
0:31:42 until the music is just right.
0:31:44 So if you like what we put together,
0:31:49 consider dropping us a line at ratethespottcast.com/a16z,
0:31:51 and let us know what your favorite episode is.
0:31:54 It’ll make my day, and I’m sure Tommy’s too.
0:31:56 We’ll catch you on the flip side.
0:31:59 (gentle music)
0:32:01 (soft music)
0:32:11 [BLANK_AUDIO]
0:00:04 between the law of the labs
0:00:07 in terms of what research directions they choose to explore
0:00:11 and which ones ultimately have breakthroughs at various times.
0:00:12 One of the hallmarks of this next phase
0:00:15 is actually going to be data production.
0:00:17 Basically no agent really works.
0:00:20 Well, it turns out there’s just no agent data on the internet.
0:00:25 The pricing for model imprints fall dramatically,
0:00:26 dramatically, dramatically, like order magnitude.
0:00:29 Yeah, orders, orders, yeah, two orders of magnitude.
0:00:30 Over two years.
0:00:35 If you’ve been listening to the A16Z podcast for a while,
0:00:37 you’ll know we talk a lot about AI.
0:00:39 We’ve covered the algorithms of power LLMs
0:00:42 and the compute required to run them.
0:00:44 But equally important is data.
0:00:46 Our guest today is as deep as you can get
0:00:50 in this world of data, the fuel behind LLMs.
0:00:52 In fact, he even recently said, quote,
0:00:55 “As an industry, we can either choose data abundance
0:00:57 or data scarcity.”
0:01:00 So what data exists today and what needs to be created,
0:01:02 either measured or synthesized?
0:01:04 Listen in to find out as I pass it over
0:01:07 to A16Z growth general partner, Sarah Wang,
0:01:09 to properly introduce this episode.
0:01:12 Hey guys, I’m Sarah Wang,
0:01:15 general partner on the A16Z growth team.
0:01:17 Welcome back to our AI revolution series
0:01:19 where we talk to industry leaders
0:01:22 about how they’re harnessing the power of generative AI.
0:01:25 Our guest this episode is Alexander Wang,
0:01:27 the founder and CEO of Scale AI,
0:01:30 a company that has become synonymous with gen AI
0:01:32 and the data needed to power advances
0:01:34 in large language models and beyond.
0:01:36 With scale’s work across enterprise,
0:01:38 automotive and the public sector,
0:01:40 Alex is also building the critical infrastructure
0:01:42 that will allow any organization
0:01:44 to use their proprietary data
0:01:47 to build the spoke gen AI applications.
0:01:48 For those of you who don’t know Alex,
0:01:51 he is one of the most impressive CEOs we’ve ever met.
0:01:53 And that’s saying something,
0:01:56 given A16Z first met Alex when he was 21
0:01:57 and already the CEO
0:02:00 of one of the fastest growing companies at its scale,
0:02:04 which he founded right before dropping out of MIT in 2016.
0:02:05 In this conversation
0:02:07 with A16Z general partner, David George,
0:02:10 Alex discusses the three pillars of AI,
0:02:12 models, compute and data,
0:02:14 and how creating abundant data
0:02:17 is core to the evolution of gen AI.
0:02:20 Alex also shares his learnings from the growth of scale,
0:02:22 his approach to leadership,
0:02:25 and what he thinks growth stage founder CEOs
0:02:27 tend to get wrong about hiring.
0:02:28 Let’s get starting.
0:02:33 As a reminder, the content here
0:02:35 is for informational purposes only,
0:02:37 should not be taken as legal, business, tax
0:02:38 or investment advice,
0:02:40 or be used to evaluate any investment or security,
0:02:42 and is not directed at any investors
0:02:45 or potential investors in any A16Z fund.
0:02:47 Please note that A16Z and its affiliates
0:02:48 may also maintain investments
0:02:51 in the companies discussed in this podcast.
0:02:53 For more details, including a link to our investments,
0:02:57 please see a16z.com/disclosures.
0:02:59 (upbeat music)
0:03:04 – We’re very excited today to have Alex Wang,
0:03:07 the founder and CEO of Scale AI with us.
0:03:08 Alex, thanks for being here.
0:03:09 – Thanks for having me.
0:03:12 – I always love talking to you and I always learn a ton,
0:03:13 but maybe to start,
0:03:14 why don’t you just tell us a little bit about
0:03:17 what you’re building at Scale AI and then we’ll dive in.
0:03:18 – Yeah, so at Scale,
0:03:20 we’re building the data foundry for AI.
0:03:21 So, taking a step back,
0:03:24 AI boils down to three pillars.
0:03:26 All the progress we’ve seen has come from
0:03:28 compute data and algorithms,
0:03:30 and the progress among all three of these pillars.
0:03:32 Compute has been powered by folks like NVIDIA,
0:03:34 the algorithmic advancements have been led
0:03:36 by the large labs like OpenAI and others,
0:03:38 and data is fueled by Scale.
0:03:42 And so our goal is to produce the frontier data necessary
0:03:44 to fuel frontier level advancements
0:03:46 in partnership with all the large labs,
0:03:49 as well as enable every enterprise and government
0:03:51 to make use of their own proprietary data
0:03:53 to fuel their frontier AI development.
0:03:55 – So on this topic of frontier data,
0:03:58 practically, but how do you actually get it?
0:04:01 – Yeah, I think this will be one of the great
0:04:04 human projects of our time, if that makes sense.
0:04:08 And I think that the only model that we have in the world
0:04:10 for the level of intelligence that we seek to create
0:04:11 is humanity.
0:04:14 And so the production of frontier data looks a lot like
0:04:17 a sort of marriage between human experts and humanity
0:04:20 with technical and algorithmic techniques around the models
0:04:23 to produce huge amounts of this kind of data.
0:04:26 And by the way, all the data that we’ve produced today,
0:04:28 the internet has looked like that too.
0:04:30 The internet in many ways is this like collaboration
0:04:32 between machines and humans to produce
0:04:35 large amounts of content and data.
0:04:37 It’ll look like the internet on steroids.
0:04:40 What happens if the internet basically,
0:04:43 instead of just being a human entertainment device
0:04:46 with this like byproduct of data generation,
0:04:48 what if it were just this large scale
0:04:49 data generation experiment?
0:04:52 – So you have a very unique perspective
0:04:53 into the state of the industry.
0:04:56 So how would you characterize the state of models,
0:04:58 the language models right now?
0:05:00 And I’d love to sort of get into things
0:05:01 like market structure,
0:05:04 but just sort of what’s the state of the industry right now?
0:05:06 – Yeah, I think we’re sort of closing in
0:05:11 at the end of maybe phase two of language model development.
0:05:14 I think phase one was the early years
0:05:15 of almost like pure research.
0:05:19 So phase one hallmarks are the original transformer paper,
0:05:22 the original small scale experiments on GPTs
0:05:25 all the way leading up probably until like GPT three
0:05:27 was this sort of phase one all research,
0:05:30 very, very focused on sort of like small scale tinkering
0:05:32 and algorithmic advancements.
0:05:36 And then phase two, which is sort of maybe GPT three till now
0:05:39 is really the sort of like initial scaling phase.
0:05:42 So we had GPT three that worked pretty well
0:05:44 and then opening eyes to start with really scaled up
0:05:46 these models to GPT four and beyond.
0:05:51 And then many companies, Google, Anthropic, Meta, XAI now,
0:05:56 many, many companies have also joined on this sort of race
0:05:58 to scale up these models to incredible capabilities.
0:06:01 So I think for the past, let’s say three years,
0:06:04 it’s almost been more about execution than anything.
0:06:05 It’s a lot of just engineering,
0:06:07 like how do you actually have large scale training,
0:06:08 work well?
0:06:10 How do you make sure there aren’t weird bugs in your code?
0:06:12 How do you set up the larger clusters?
0:06:14 A lot of execution will work to get to where we are now,
0:06:17 where we have kind of a number of very advanced models.
0:06:19 And then I think we’re entering a phase
0:06:21 where the research is gonna start mattering a lot more.
0:06:24 Like I think there will be a lot more divergence
0:06:25 between a lot of the labs
0:06:28 in terms of what research directions they choose to explore
0:06:31 and which ones ultimately have breakthroughs at various times.
0:06:33 And it’s sort of an exciting alternating phase
0:06:35 between maybe just raw execution
0:06:39 versus sort of a more innovation powered cycle.
0:06:41 – They’ve kind of gotten to a point
0:06:44 where I wouldn’t say there’s like abundant compute
0:06:46 but they’ve had enough compute that they’ve needed
0:06:48 in order to get to the models where they’re at.
0:06:50 That’s not a constraint necessarily.
0:06:53 They’ve kind of exhausted as much data as they possibly can
0:06:55 all of the frontier labs.
0:06:58 And so the next thing will be breakthroughs on that
0:07:00 and then advancing the ball on the data side.
0:07:01 Is that fair?
0:07:02 – Yeah, and I think basically,
0:07:03 yeah, if you look at the pillars,
0:07:05 compute were obviously continuing
0:07:06 to scale up the training clusters.
0:07:10 So I think that direction is pretty clear on the algorithms.
0:07:12 I think there has to be a lot of innovation there.
0:07:14 Frankly, I think that’s where a lot of the labs
0:07:15 are really working hard,
0:07:16 I think on the pure research of that.
0:07:18 And then data, you can have alluded to it,
0:07:22 we’ve kind of run out of all the easily accessible
0:07:24 and easily available data out there.
0:07:25 And–
0:07:25 – Yeah, Common Core is all done.
0:07:27 Everybody’s got the same access to it.
0:07:28 – Yeah, exactly.
0:07:30 And so a lot of people talking about this is the data wall.
0:07:31 You know, we’re kind of hitting this wall
0:07:33 where we’ve leveraged all the publicly available data.
0:07:35 And so one of the hallmarks of this next phase
0:07:37 is actually going to be data production.
0:07:40 And what is the method that each of these labs
0:07:42 is going to use to actually generate the data necessary
0:07:45 to get you to the next levels of intelligence
0:07:47 and how do we get towards data abundance?
0:07:50 And I think this is going to require a number of fields
0:07:53 of sort of advanced work and advanced study.
0:07:55 I think the first is really pushing
0:07:57 on the complexity of the data.
0:07:59 So moving towards frontier data.
0:08:01 So a lot of the capabilities
0:08:03 that we want to build into the models,
0:08:05 the biggest blocker is actually a lack of data.
0:08:08 So for example, agents has been the buzzword
0:08:10 for the past two years
0:08:12 and basically no agent really works.
0:08:14 Well, it turns out there’s just no agent data
0:08:15 on the internet.
0:08:17 There’s no just pool of really valuable agent data
0:08:18 that’s just sitting around anywhere.
0:08:20 And so we have to figure out how to produce
0:08:21 really high quality agent data.
0:08:23 – Give an example of like, what would you have to produce?
0:08:26 – So we have some work coming out on this soon
0:08:28 which demonstrates that right now,
0:08:30 if you look at all the frontier models,
0:08:32 they suck at composing tools.
0:08:36 So if they have to use one tool and then another tool,
0:08:38 let’s say they have to look something up
0:08:39 and then write a little Python script
0:08:41 and then chart something.
0:08:43 They use multiple tools in a row, they just suck at that.
0:08:45 They just are really, really bad
0:08:47 at utilizing multiple tools in a row.
0:08:49 And that’s something that’s actually very natural
0:08:50 for humans to do.
0:08:52 – Yeah, but it’s not captured anywhere, right?
0:08:53 That’s the point, right?
0:08:53 – Exactly.
0:08:55 – So you can’t actually go take the capture
0:08:57 of somebody going from one window to another
0:08:59 into a different application
0:09:01 and then feed that to the model so it learns, right?
0:09:02 – Exactly, yeah, yeah.
0:09:05 So these sort of reasoning chains through
0:09:07 when humans are solving complex problems,
0:09:09 we naturally will use a bunch of tools,
0:09:10 we’ll think about things,
0:09:12 we’ll reason through what needs to happen next,
0:09:14 we’ll hit errors and failures
0:09:16 and then we’ll go back and sort of like reconsider.
0:09:17 You know, a lot of these reasoning chains,
0:09:19 these agentic chains are,
0:09:21 the data just doesn’t exist today.
0:09:23 So that’s an example of something that needs to be produced,
0:09:24 but taking a big step back,
0:09:26 what needs to happen on data.
0:09:28 First is increasing data complexity,
0:09:29 so moving towards frontier data.
0:09:31 The second is just data abundance,
0:09:32 increasing the data production.
0:09:35 – Capturing more of what humans actually do
0:09:36 in the field of work.
0:09:37 – Yeah, both capturing more of what humans do
0:09:40 and I think investing into things like synthetic data,
0:09:42 hybrid data, so utilizing synthetic data,
0:09:44 but having humans be a part of that loop
0:09:46 so that you can generate much more high quality data.
0:09:48 We need basically just in the same way,
0:09:51 I think with chips, we talk a lot about chip foundries
0:09:52 and how do we ensure that we have like
0:09:54 enough means of production chips.
0:09:55 And the same thing is true for data,
0:09:57 we need to have effectively data foundries
0:10:00 and the ability to generate huge amounts of data
0:10:02 to fuel the training of these models.
0:10:05 And then I think the last leg of the stool,
0:10:08 which is often rated as measurement of the models
0:10:10 and ensuring that we actually have,
0:10:13 I think for a while the industry is just sort of like,
0:10:14 oh yeah, we just add a bunch more data
0:10:15 and we see how good the model is
0:10:16 and we add a bunch more data
0:10:17 and we see how good the model is,
0:10:19 but we’re gonna have to get pretty scientific
0:10:22 around exactly what is the model not capable of today
0:10:24 and therefore what are the exact kinds of data
0:10:26 that need to be added to improve the model’s performance.
0:10:29 – How much of an advantage do the big tech companies have
0:10:33 with their corpus of data versus the independent labs?
0:10:35 – Yeah, well there’s a lot of regulatory issues
0:10:38 that they have with utilizing their existing data corpuses.
0:10:40 You know, you can look through,
0:10:41 this is before all this generative AI work,
0:10:44 but at one point Meta did some research
0:10:47 that utilized basically all the public Instagram photos
0:10:49 along with their hashtags
0:10:51 to train really good image recognition algorithms.
0:10:54 They had a lot of regulatory problems with that in Europe.
0:10:56 Like it turned out to be a huge pain in the ass.
0:10:58 So I think that that’s one thing
0:11:01 that’s kind of difficult to reason through,
0:11:03 which is to what degree from a regulatory perspective,
0:11:04 particularly in Europe,
0:11:05 these companies are going to be able
0:11:07 to utilize their data advantages.
0:11:09 So I think that was kind of TBD.
0:11:12 I think that the real way in which a lot of the large labs
0:11:16 have just dramatic advantages is just,
0:11:18 they have very profitable businesses
0:11:22 that can provide near infinite sources of capital
0:11:24 for these AI efforts.
0:11:25 And I think that that’s something
0:11:27 that I’m watching pretty intently
0:11:29 or I’m very curious to see how it plays out.
0:11:31 – Hey, it’s Steph.
0:11:34 You might know that before my time at A16C,
0:11:36 I used to work at a company called The Hustle.
0:11:38 And then we were acquired by HubSpot
0:11:41 where I helped build their podcast network.
0:11:42 While I’m not there anymore,
0:11:45 I’m still a big fan of HubSpot podcasts,
0:11:47 especially My First Million.
0:11:49 In fact, I’ve listened to pretty much
0:11:51 all 600 of their episodes.
0:11:53 My First Million is perfect for those of you
0:11:55 who are always trying to stay ahead of the curve
0:11:56 or in some cases,
0:11:58 take matters into your own hands
0:12:00 by building the future yourself.
0:12:03 Posted by my friends, Sam Parr and Sean Curry,
0:12:05 who have each built and sold eight-figure businesses
0:12:07 to Amazon and HubSpot,
0:12:08 the show explores business ideas
0:12:10 that you can start tomorrow.
0:12:12 Plus, Sam and Sean jam alongside guests
0:12:15 like Mr. Beast, Rob Dyrdek, Tim Ferriss,
0:12:18 and every so often, you’ll even find me there.
0:12:20 From gas station pizza and egg carton businesses
0:12:23 doing millions all the way up to several guests
0:12:25 making their first billion.
0:12:27 Go check out My First Million
0:12:28 wherever you get your podcasts.
0:12:31 (upbeat music)
0:12:37 – There’s this whole question,
0:12:39 the industry is like, are they over-investing?
0:12:40 And if you listen to their earnings calls
0:12:41 of the big tech companies,
0:12:43 they’re like, look, our risk is under-investing,
0:12:44 not over-investing.
0:12:45 What do you make of that?
0:12:46 – Yeah, I mean, if you think about,
0:12:50 let’s take the incentives of any one of the CEOs of the,
0:12:53 put yourself in the shoes of Sundar Pachai
0:12:54 or Mark Zuckerberg or whatnot.
0:12:55 – Or Satya, yeah.
0:12:56 – Or Satya.
0:13:00 And to your point, if they really nail this AI thing,
0:13:03 they could generate another trillion dollars of market cap
0:13:04 probably very easily.
0:13:06 If they really are ahead of the competition
0:13:07 and they productize in a good way,
0:13:10 like trillion dollars of market cap, kind of no-brainer.
0:13:12 And if they don’t invest the extra,
0:13:15 whatever it is, 20 or 30 billion of CapEx per year,
0:13:16 and they miss out on that.
0:13:19 And then there’s some real existential risk, I think too,
0:13:20 for each of the large–
0:13:21 – Yeah, in their each formula.
0:13:24 – Yeah, all their businesses are potentially
0:13:26 deeply disruptible by AI technology.
0:13:29 So the risk reward for them is very obvious.
0:13:31 So that’s, I think, the big picture thinking.
0:13:32 And then from a more tactical level,
0:13:35 I think all of them are gonna be able to
0:13:38 pretty easily recoup their capital investments
0:13:41 just by worst case making their core businesses
0:13:42 more efficient and effective.
0:13:44 So for example, like, you know.
0:13:47 – Yeah, GPU utilization for Facebook advertising.
0:13:48 – Yeah, Facebook, Google,
0:13:50 they make their advertising systems a little bit better.
0:13:52 They can recoup billions of dollars just by–
0:13:53 – Yeah, better performance.
0:13:55 – Yeah, better performance there.
0:13:56 Apple can easily recoup the investments
0:13:58 if it drives an upgrade cycle.
0:13:59 I mean, these are things
0:14:00 that I think are pretty clear.
0:14:02 – Look, it’s generally great for the industry
0:14:04 that they are investing so much capital
0:14:05 because they also are in the business
0:14:06 of renting this compute out,
0:14:08 or at least in the case of Google and Microsoft, they are.
0:14:10 – And the models are making their way,
0:14:12 like, Lama 3.1 is open source.
0:14:16 And so even the literal fruits of all the investment
0:14:18 are becoming broadly accessible.
0:14:19 And so the surplus generated
0:14:23 from the open source and these models is kind of insane.
0:14:23 – It’s insane.
0:14:26 And that’s a great segue into market structure
0:14:27 at the model layer.
0:14:29 So what do you think actually happens?
0:14:32 Are there the few players that we’ve all identified now,
0:14:34 the handful, and they all compete?
0:14:37 Do you think it’s a profitable business?
0:14:39 What impact does open source have
0:14:40 on the quality of the businesses?
0:14:43 Take us a couple of years ahead and give us your forecast.
0:14:45 – Yes, we’ve seen over the past,
0:14:47 even just like year and a half,
0:14:50 the pricing for model inference
0:14:53 fall dramatically dramatically dramatically.
0:14:54 – Right, order of magnitude.
0:14:55 – Yeah, two orders of magnitude.
0:14:56 – Two orders of magnitude.
0:14:57 – Over two years.
0:14:58 And so it’s this shocking thing
0:15:01 that it turns out intelligence might be a commodity.
0:15:04 But no, I mean, I think that this huge sort of lack
0:15:07 of pricing power, let’s say, on the pure model layer,
0:15:12 certainly indicates that renting models out on their own
0:15:14 may or may not be the best long-term business.
0:15:17 I think it’s likely to be a relatively mediocre
0:15:18 long-term business.
0:15:20 – Well, I guess it depends on the breakthrough thing,
0:15:22 which is the earlier point, right?
0:15:24 To the extent that someone actually has a durable breakthrough
0:15:26 or multiple people have durable breakthroughs,
0:15:28 like then potentially market structure is different.
0:15:31 – So two things, if meta continues open sourcing,
0:15:34 that puts a pretty strong cap as to the value
0:15:35 that you can get from the model layer.
0:15:39 And then two, if at least a handful of the labs
0:15:42 are able to have similar performance over time,
0:15:45 then that also dramatically changes the pricing equation.
0:15:47 So we think that it’s not 100%,
0:15:51 but chances are the pure model renting business
0:15:52 is not the highest quality business,
0:15:54 where there are much higher quality businesses
0:15:56 are going to be above and below.
0:16:01 So below, I mean, NVIDIA is obviously an incredible business,
0:16:03 but the clouds also have really great businesses too,
0:16:05 because it turns out it’s pretty hard logistically
0:16:08 to actually set up large clusters of GPUs.
0:16:10 And so the cloud providers actually have pretty good margins
0:16:11 when they rent out.
0:16:13 – And the traditional data center business
0:16:14 is very much a scale game.
0:16:15 – Yep.
0:16:17 – So they are massively benefited
0:16:18 relative to smaller players.
0:16:19 – Yeah, exactly.
0:16:20 So I think picks and shovels,
0:16:22 so if you’re under the model layer,
0:16:23 I think there’s great businesses there.
0:16:24 And if you’re above the model layer,
0:16:26 if you’re building applications,
0:16:27 ChatJPT is a great business.
0:16:31 And a lot of the apps in the startup realm
0:16:32 actually are working pretty well.
0:16:34 I mean, none of them are quite as big as ChatJPT, obviously,
0:16:36 but a lot of apps,
0:16:38 if they nail the early product market fit,
0:16:40 end up being pretty good businesses,
0:16:41 great businesses as well,
0:16:44 because the value that they generate for customers,
0:16:47 if they get the whole user experience correct,
0:16:49 far exceeds the inference cost of the models.
0:16:50 – There’s some cool stuff here, right?
0:16:54 I think an Anthropics launch of artifacts in Claude,
0:16:57 it’s like the first pin drop of this major theme
0:17:00 of all the labs are gonna be pushing
0:17:02 much deeper product integrations
0:17:04 to be able to drive higher quality businesses.
0:17:05 So that’ll be the other story is,
0:17:07 I think we’re gonna see a lot of iteration
0:17:10 at the product layer and the product level.
0:17:13 The sort of boring chatbots is not gonna be the end product.
0:17:14 That’s not the end all be all.
0:17:15 – Disappointing outcome.
0:17:16 – Yeah, exactly.
0:17:19 And product iteration and the product innovation cycles
0:17:21 is very hard to predict because,
0:17:23 I mean, OpenAI was surprised how popular chatGPD was.
0:17:25 I don’t think it’s like super obvious to me
0:17:27 or anyone in the industry, frankly,
0:17:30 what exact products are gonna be the ones that hit
0:17:32 and what’s gonna provide the next legs of growth.
0:17:36 But you have to believe that an OpenAI or an Anthropic
0:17:38 can build great applications businesses
0:17:41 to for them to be long-term independent and sustainable.
0:17:43 – Yeah, for sure.
0:17:45 Yeah, and then it’s what drives competitive advantage.
0:17:46 Obviously you have the model,
0:17:48 a tightly integrated product on top of it,
0:17:51 and then the good old fashioned modes from there.
0:17:51 – Yeah.
0:17:54 – Workflows, integrations, all that stuff.
0:17:56 – I think you can clearly see they’re thinking on it.
0:17:58 I mean, both OpenAI and Anthropic
0:17:59 hired chief product officers
0:18:01 within two months of each other.
0:18:02 – Yeah, they’re figuring it out.
0:18:04 – And then it’s sort of a change of tune
0:18:07 where they’re like, “Oh no, we’re very purely focused on this
0:18:09 “and it’s okay, I think there’s the realization to it.”
0:18:11 So, yeah, exactly, it makes a whole sense.
0:18:12 You’ve got an application business
0:18:14 with some really interesting customers.
0:18:15 What are you hearing from enterprises
0:18:18 as to how they’re actually putting this into place?
0:18:20 – I think what we’ve seen is
0:18:22 there was a huge amount of excitement from the enterprise.
0:18:24 A lot of enterprises were like,
0:18:26 “Shit, we have to start doing something.
0:18:27 “We have to get ahead of this.
0:18:29 “We have to start experimenting with AI.”
0:18:33 I think that that led them to this fast POC cycle
0:18:35 where they’re like, “Okay, where are all the low hanging
0:18:36 “fruit ideas that we have?”
0:18:37 – Go buy AI stuff.
0:18:38 – Yeah, yeah.
0:18:39 – And let’s go try all of it.
0:18:42 And some of those things are good, some of them aren’t good,
0:18:45 but I think regardless, it’s been this big frenzy,
0:18:48 much fewer of the POCs have made it to production
0:18:50 than I think the industry overall expected.
0:18:52 And I think a lot of enterprises are looking at now
0:18:55 and the doomsday that they thought might have happened
0:18:56 hasn’t really happened.
0:19:00 AI has not fully terraformed and transformed
0:19:01 most of the major industries.
0:19:03 Like it’s not like totally, you know–
0:19:04 – It’s sort of marginal stuff.
0:19:06 It’s like efficiency gains and support
0:19:08 and then some of the creative tasks and things like that.
0:19:09 – Yeah, exactly.
0:19:10 – Otherwise it’s pretty light.
0:19:12 The thing that we think a lot about is like,
0:19:15 what AI improvements or AI transformations or AI efforts
0:19:17 that we’re working on actually can
0:19:20 meaningfully drive the stock price
0:19:22 of the companies that we’re working on.
0:19:24 And so that’s what we encourage all of our customers
0:19:25 to really be thinking about
0:19:28 because at the end of the day, the potential is there.
0:19:31 There’s latent potential for almost every enterprise
0:19:33 to implement AI at a level
0:19:35 that would meaningfully boost their stock price.
0:19:37 – Mostly in the form of cost savings.
0:19:38 – Efficiency gains.
0:19:39 – Well today in the form of cost savings,
0:19:41 but then also much better customer experiences.
0:19:43 Like I think in a lot of industries
0:19:46 where there’s a lot more manual interaction with customers,
0:19:48 you should be able to drive much better customer interactions
0:19:49 if you have more standardization
0:19:51 and you were able to use more automation.
0:19:54 And then those eventually would make their way to gains
0:19:56 of market share with respect to competitors.
0:19:59 So that’s what we’re pushing our customers towards.
0:20:01 And I see it, some of the CEOs that we work with,
0:20:02 they’re all on board.
0:20:04 And they understand that it’s gonna be
0:20:05 a multi-year investment cycle.
0:20:08 They might not see gains the next quarter,
0:20:10 but if they actually pull through the other side,
0:20:12 they’re gonna see massive transformations.
0:20:15 I think that a lot of the frenzy around small use cases
0:20:17 and sort of the more marginal use cases,
0:20:18 I think that’s good.
0:20:19 I think it’s exciting.
0:20:20 I think they should be doing it.
0:20:23 But to me, that’s not what we’re all here to do.
0:20:24 – Yeah, it’s very much like the application layer
0:20:26 is like very much like phase one right now,
0:20:28 which is, I mean, yeah, there’s some automation,
0:20:30 but it’s largely like chatbots.
0:20:34 My hope as a startup investor is that over time,
0:20:36 there’s a window that opens for the startups
0:20:39 where product innovation will help them to win
0:20:40 and beat the incumbents.
0:20:42 Like my partner, Alex Rampel, has this phrase,
0:20:44 which is, is the startup gonna get to distribution
0:20:46 before the incumbent finds innovation.
0:20:48 And I think there’s an opportunity for it,
0:20:50 but it’s like the tech is too early right now.
0:20:51 – Yeah.
0:20:52 – I don’t know if you would agree with that, but…
0:20:55 – I think the tech is too early to imagine,
0:20:57 yeah, again, because it’s mostly cost saving.
0:20:59 I think if most of the benefit is on the cost saving side,
0:21:02 then that’s not really enough to disrupt large incumbents
0:21:04 that has already kind of like pushed their way through
0:21:07 all the costs of growing and distribution.
0:21:09 – How valuable do you think is the data inside of enterprises?
0:21:11 Like you said, JP Morgan has whatever,
0:21:14 15 petabytes of data or something like that.
0:21:17 I never wrote the numbers, but like, is that overrated?
0:21:19 How much of it is actually useful?
0:21:22 Because to date, most of that data has not given them
0:21:24 some meaningful competitive advantage.
0:21:26 So do you think that actually changes?
0:21:28 – I think AI is the first time you could see
0:21:30 that potentially change because basically,
0:21:32 obviously there’s the whole big data wave.
0:21:34 Big data boils down to better analytics,
0:21:36 which is helpful, like marginally helpful
0:21:39 for business decision making, but not deeply transformation.
0:21:41 – It does massively change the way the products work.
0:21:42 – Yeah, exactly.
0:21:45 Whereas now you actually can imagine
0:21:47 some massive transformation in the way the products work.
0:21:49 So let’s take any big bank.
0:21:54 A lot of the valuable interactions between a user
0:21:57 and a large bank like a JP Morgan or Morgan Stanley or whatnot
0:21:59 are human driven, are people driven.
0:22:02 And they try their best to ensure that the quality
0:22:03 of experience is very high across the board,
0:22:06 but obviously with any large process,
0:22:07 there’s only so much you can do to assure that.
0:22:10 But all of your prior customer interactions
0:22:12 and all the ways in which your business
0:22:16 has worked historically is the only available data
0:22:18 to be able to train models to do well
0:22:19 at this particular task.
0:22:21 And if you think about like wealth management,
0:22:23 there’s very little indistribution data of that
0:22:25 on the internet that you could trade a model off of.
0:22:26 – So there’s behind the walls,
0:22:28 there’s actually quite a bit, it’s very rich.
0:22:29 – Yeah, huge amounts of data.
0:22:31 So I think that a lot of the data is probably
0:22:33 not super relevant to actually transforming your business,
0:22:35 but some of the data is hyper valuable.
0:22:38 So I think enterprises have a lot of trouble
0:22:41 and challenge around actually utilizing
0:22:43 any amount of data that they have.
0:22:46 It’s poorly organized, it’s sort of all over the place,
0:22:49 they pay consulting firms tens of millions of dollars,
0:22:51 hundreds of millions of dollars to do these data migrations.
0:22:53 And it’s even after that no change in results.
0:22:54 Yeah, no change results.
0:22:57 So I think it’s historically very difficult place
0:22:59 for enterprises to really drive transformation.
0:23:01 And so in some ways this is the race,
0:23:04 are they gonna be able to figure out how to utilize
0:23:08 and leverage their data faster than some startup
0:23:10 figures out how to somehow get access to data?
0:23:11 – Create a massively different product
0:23:13 with a little bit subset of the data.
0:23:14 – Yeah, exactly.
0:23:17 – Shifting gears to how you run your company
0:23:19 and how you built your company.
0:23:21 One of the things that you’ve talked about
0:23:25 is a mistake that you made during the go times
0:23:29 of 2020 and 2021 around hiring.
0:23:31 And this notion that in order to scale,
0:23:32 you had to hire a ton.
0:23:35 And it’s something we saw with all of our portfolio companies
0:23:37 was like, hey, there’s war for talent.
0:23:39 And it meant that we gotta go higher,
0:23:40 we gotta go higher, we gotta go higher.
0:23:42 So what were the lessons that you learned
0:23:44 through that process and then how have you changed
0:23:45 how you’ve done things afterwards?
0:23:48 – So over the past few years,
0:23:51 we’ve basically kept our headcount flat.
0:23:54 I mean, we’ve grown it very slightly as the business grown,
0:23:57 but the business itself is five X, well, six X.
0:23:58 You know, the business has grown dramatically.
0:24:02 And the takeaway from this entire process is
0:24:07 it feels very logical that more people equals better results
0:24:10 and more people equals more stuff being done,
0:24:12 but rather paradoxically, I think,
0:24:14 if you have a very high performing team
0:24:15 and a very high performing org,
0:24:18 it’s nearly impossible to grow it dramatically
0:24:21 without losing all of that high performance
0:24:22 and all of the winning culture.
0:24:24 – Yeah, reducing the communication and coordination overhead
0:24:26 actually increases productivity.
0:24:27 – That’s definitely true.
0:24:29 And I think it’s actually something even deeper,
0:24:31 which is that a very high performing team
0:24:33 of a certain size is almost like
0:24:36 this very intricate sculpture
0:24:38 and this interplay between all the people in the team.
0:24:40 And if you just add a bunch of people into that,
0:24:42 even if the people are great,
0:24:43 like it just screws the whole thing up.
0:24:46 And no matter what, as you add people,
0:24:47 you’re gonna have regression to the mean.
0:24:49 You know, if you kind of observe companies
0:24:50 that do scale healthcare a lot
0:24:52 and that’s pretty core to their financial results,
0:24:54 I think they acknowledge that regression,
0:24:55 that mean regression.
0:24:57 So if you think about like the scaling
0:24:58 of large sales teams, for example.
0:24:59 – Yeah, sure.
0:25:00 – You acknowledge that you’re going to have
0:25:02 that mean regression, but you just operationalize
0:25:04 so that you’re like a little bit above the mean.
0:25:05 And if you’re able to do that,
0:25:07 then the whole equation still works financially.
0:25:09 – Yeah, I’d say sales is different than product.
0:25:11 – Yeah, totally, of course.
0:25:13 But our observation is just startups work
0:25:15 because you have very high performing teams
0:25:17 and you wanna keep those high performing teams
0:25:19 intact as long as you possibly can.
0:25:22 You know, I think a common startup failure mode
0:25:24 is that you have something that works,
0:25:26 but everybody in the company is really junior.
0:25:28 So then things are scaling,
0:25:30 but all the wheels are kind of falling off.
0:25:32 Your investors tell you how you should hire
0:25:34 some executives, you go through these searches
0:25:37 that are somehow uniquely soul crushing every time.
0:25:40 But you go through, if you’re graded in,
0:25:41 it works half the time.
0:25:43 – Yeah, yeah.
0:25:44 – So you go through the exact searches,
0:25:47 you’re bringing exact and then you give the exact
0:25:49 a lot of rope and your exact say,
0:25:51 hey, we need to hire a massive team
0:25:52 for us to hit our results.
0:25:55 And you’re like, yeah, I mean, I’m pretty experienced,
0:25:57 you seem really experienced, let’s do what you say.
0:25:59 And you let these big teams sort of be built.
0:26:02 And the reality is I think this almost always
0:26:03 results in ruin.
0:26:05 I think that this isn’t to say that you can’t hire executives
0:26:07 from the outside, but I think what you need to do
0:26:09 when you hire executives from the outside
0:26:11 is they really get steeped in how the company works.
0:26:15 And before they make any major sweeping suggestions,
0:26:18 they get into the rhythm and the operations of the company.
0:26:20 And they understand why does the whole thing
0:26:21 work in the first place?
0:26:23 Why are the things that are working working?
0:26:25 And then they make thoughtful suggestions.
0:26:27 Initially, they take small steps
0:26:29 and you sort of like you trust and verify
0:26:30 each of these small steps.
0:26:33 And eventually maybe they can make more sweeping suggestions,
0:26:35 but it should be at a point where they have
0:26:38 a clear track record of making small steps
0:26:38 that have been really beneficial.
0:26:39 – Oh, that’s interesting.
0:26:41 That’s interesting and very tangible, right?
0:26:43 It starts small when you hire a big executive
0:26:44 and it’s a little bit counterintuitive.
0:26:46 And it’s not the way that any of those executives
0:26:48 want to go.
0:26:50 – Yeah, I think that there’s kind of an exec fantasy
0:26:52 that I’ve noticed, which is, and by the way,
0:26:54 I think executives are great people
0:26:55 and they’re like, they’re incredible.
0:26:58 But there is a tendency for an executive fantasy,
0:27:00 particularly for like Silicon Valley companies
0:27:02 with young founders and whatnot, which is,
0:27:05 oh, I’m gonna come in and I’m gonna fix this whole thing.
0:27:08 I’m gonna make this a professional operation.
0:27:10 You’re recruiting teammates at the end of the day.
0:27:12 You’re not recruiting like some magic wand.
0:27:14 You’re recruiting a teammate who you believe
0:27:17 over an extended period of time is gonna have great judgment
0:27:20 in making repeated decisions about the business.
0:27:21 But, and this is where we’ve made mistakes.
0:27:24 It’s like, you’re not buying some magical bag of goods
0:27:27 that is gonna bring this magic formula into your business
0:27:29 that will all of a sudden make the whole thing work.
0:27:31 On the flip side, there’s a founder fantasy.
0:27:34 The founder fantasy, or the founder CEO fantasy,
0:27:37 which is, oh, I’m gonna just hire a bunch
0:27:39 of incredible execs, throw over me fucking pros.
0:27:41 And then I’m gonna go–
0:27:43 – They’ll do the stuff I don’t wanna do.
0:27:47 They’ll do all the stuff I don’t wanna do.
0:27:48 And I’m gonna be able to sit back
0:27:50 and watch the machine work.
0:27:52 And that’s also extremely unrealistic
0:27:53 because the flip side is also true.
0:27:55 The reason that you are a good founder CEO
0:27:57 is because you make very good decisions
0:27:59 over and over again over an extended period of time.
0:28:02 And to pull yourself out of those decision-making loops
0:28:04 would be kind of crazy.
0:28:05 – That’s a pattern we’ve seen a lot,
0:28:07 which is I’m gonna hire executives.
0:28:08 I’m gonna step back a bit.
0:28:10 And then it’s, oh, shit, realization
0:28:11 that like, hey, some big decisions go wrong
0:28:13 and wait, this is the point of me being here.
0:28:15 – Yeah, I think it can work
0:28:19 if your industry is very stable, potentially.
0:28:20 – Well, look at any public company
0:28:24 when they change CEOs and the stock price moves like 2%.
0:28:26 And it’s like, oh, okay, well, actually,
0:28:27 it doesn’t really matter.
0:28:29 That is a cog, but that is very different
0:28:32 from a high-growth startup that’s run by a founder.
0:28:33 – Exactly, yeah, yeah.
0:28:35 And I think that a lot of startups
0:28:36 and a lot of companies are valuable
0:28:38 because of an innovation premium.
0:28:39 – 100%.
0:28:41 – Investors believe that founder-led companies
0:28:43 are going to out-innovate the market.
0:28:47 And so your job is to out-innovate the market, so.
0:28:49 – You better be in the strategic decisions.
0:28:50 Yeah, for sure.
0:28:51 How about MEI?
0:28:55 So you recently rolled out this concept.
0:28:59 I think like half of my ex-feed was praising you
0:29:01 and that’s probably more than half.
0:29:03 Some portion of my ex-feed was yelling at you,
0:29:04 talking about the concept
0:29:06 and what are your observations of rolling it out so far?
0:29:10 – Yeah, so MEI, we basically rolled out this idea
0:29:12 of merit, excellence, and intelligence.
0:29:15 And the basic idea is in every role
0:29:18 we’re gonna hire the best possible person
0:29:20 regardless of their demographics.
0:29:23 And we’re not going to do any sort of quota-based
0:29:25 optimization of our workforce
0:29:28 to meet certain demographic targets.
0:29:29 That doesn’t mean we don’t care about diversity.
0:29:33 We actually care about having diverse pipelines
0:29:35 and diverse top-of-funnel for all of our roles.
0:29:37 But at the end of the day,
0:29:39 the best, most capable person for every job
0:29:41 is going to be the one that we hire.
0:29:45 It’s one of these things that was mildly controversial,
0:29:50 but I think it’s also, if we were to just take a big step back
0:29:53 as to who should companies be hiring,
0:29:54 I think it’s kind of an obvious-
0:29:56 – Sort of common sense, yeah.
0:29:57 – Yeah, it feels kind of obvious that-
0:29:58 – Lost the plot, yeah.
0:30:00 – Companies should hire the most talented people.
0:30:02 And I think there’s obviously this became this big question
0:30:05 of like how much social responsibility do companies have
0:30:06 in what they do?
0:30:09 My take is I operate in a very competitive industry.
0:30:13 Scale’s role is to help fuel artificial intelligence.
0:30:14 It’s very important technology.
0:30:17 We need incredibly smart people to be able to do this,
0:30:20 and we need the best people to be able to accomplish this.
0:30:22 I think that this is something that,
0:30:24 I think most people at scale would say
0:30:26 was sort of like implicitly true,
0:30:27 or sort of it wasn’t like a departure
0:30:31 from how many of us thought of what we do at scale.
0:30:33 But it was really valuable for us to codify it
0:30:34 because it gives everybody confidence
0:30:37 that even if this is how we operate today,
0:30:38 companies change over time,
0:30:40 we’re not gonna change this quality.
0:30:41 – Well, this has been awesome.
0:30:45 I wanna close with an optimistic question and forecast,
0:30:49 which is what is your sort of own view of,
0:30:50 or definition of AGI?
0:30:54 And what is your expected timeline to when we reach that?
0:30:56 – Yeah, I like the definition of this that’s sort of like,
0:31:00 let’s say 80 plus percent of jobs
0:31:02 that people can do purely at computers,
0:31:03 so digital focused jobs.
0:31:05 AI can accomplish those jobs.
0:31:07 It’s not like imminent.
0:31:09 It’s not like immediately on the horizon,
0:31:12 so on the order of four plus years,
0:31:14 but you can see the glimmers,
0:31:17 and depending on the algorithmic innovation cycle
0:31:18 that we talked about before,
0:31:20 could make that much sooner.
0:31:21 Yeah, that’s awesome.
0:31:22 – Very exciting.
0:31:23 Well, Alex, thanks for being here.
0:31:24 Great to chat with you as always.
0:31:26 Learned a ton.
0:31:27 Really appreciate it.
0:31:28 Yeah, thanks for having me.
0:31:32 – All right, that is all for today.
0:31:34 If you did make it this far, first of all, thank you.
0:31:36 We put a lot of thought into each of these episodes,
0:31:38 whether it’s guests, the calendar Tetris,
0:31:40 the cycles with our amazing editor, Tommy,
0:31:42 until the music is just right.
0:31:44 So if you like what we put together,
0:31:49 consider dropping us a line at ratethespottcast.com/a16z,
0:31:51 and let us know what your favorite episode is.
0:31:54 It’ll make my day, and I’m sure Tommy’s too.
0:31:56 We’ll catch you on the flip side.
0:31:59 (gentle music)
0:32:01 (soft music)
0:32:11 [BLANK_AUDIO]
What if the key to unlocking AI’s full potential lies not just in algorithms or compute, but in data?
In this episode, a16z General Partner David George sits down with Alex Wang, founder and CEO of Scale AI, to discuss the crucial role of “frontier data” in advancing artificial intelligence. From fueling breakthroughs with complex datasets to navigating the challenges of scaling AI models, Alex shares his insights on the current state of the industry and his forecast on the road to AGI.
Resources:
Find Alex on Twitter: https://x.com/alexandr_wang
Find David on Twitter : https://x.com/DavidGeorge83
Stay Updated:
Let us know what you think: https://ratethispodcast.com/a16z
Find a16z on Twitter: https://twitter.com/a16z
Find a16z on LinkedIn: https://www.linkedin.com/company/a16z
Subscribe on your favorite podcast app: https://a16z.simplecast.com/
Follow our host: https://twitter.com/stephsmithio
Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.