No Priors Ep. 44 | With Former Square CEO Alyssa Henry

1. 0:00 – 2:30

   Alyssa’s experience and career trajectory

   1. SGSarah Guo0:00

      (music plays) Hi, listeners, and welcome to another episode of No Priors. This week we're joined by Elissa Henry, who recently retired from being long-time CEO of Square within Block. Before that, she was the vice president of AWS, running, amongst other things, the storage products or the digital storage bucket for the world. And before AWS, she ran order management software at Amazon, and started her tech career at Microsoft. She remains on the boards of Intel and Confluent, and was previously on the board of Unity. I'm a huge admirer of Elissa's leadership. Welcome, Elissa.

   2. AHAlyssa Henry0:35

      Great to be here.

   3. SGSarah Guo0:36

      Let us start by talking about Square. You led there for almost a decade through enormous growth. Was there a single moment that most defined your experience?

   4. AHAlyssa Henry0:47

      I don't know if there is a single moment that most defined, although there were, there were a ton of different moments. Obviously, we were known as the little white reader company and the farmer markets payments company when I joined. Cash App wasn't really a thing. You know, Tidal wasn't even a thought. Crypto wasn't a thing. Um, the Square business transformed from that little white reader into, um, you know, a much larger business serving, um, businesses of many, many different sizes, from still the smallest to also, you know, large stadiums and multinational companies. So, the moment that really was... I don't know if it was most defining, but one of the, the... I still tear up a little bit about it, is actually the day we IPO'd, and what was so exciting about that was, you know, if you looked at the sign that was outside the New York Stock Exchange, it was covered with just the logos of all these small businesses, right? And, and the, the note was, you know, "The neighborhood is going public." And I think that really kinda sums up, you know, just the mission-driven, purpose-driven organization that Square and Block is, and was, and, um, and just the impact that the company has had on just countless, um, small businesses, helping them, give them the tools and the technology that, in many ways, had only been previously available to, you know, the likes of Amazon and Walmarts of the world.

   5. EGElad Gil2:05

      It's funny, I was a early investor in Square, a small one, and then I, um, worked at Twitter, so I worked with Jack in two very different contexts. The way that he ran the, the two businesses was radically different, right? I think, um, it's everything from the org structure, where Square was always more sort of GM/business unit almost driven, versus Twitter, which was always sort of functionally oriented or, you know, often functionally oriented.

2. 2:30 – 4:09

   Transition from engineer to manager

   1. EGElad Gil2:30

      Um, how did, how did your career shift over time? Because I think you started off more on sort of platform and technology, and then you took over a big business area, and I'm a little bit curious about that evolution and how you went from being somebody who's done a variety of things, they're more product and engineering driven, to somebody who's really running a whole business area.

   2. AHAlyssa Henry2:47

      Well, my career's kind of gone back and forth between product and engineering, between, um, functional leadership and general management leadership a couple of times over, you know, my several decades in, in tech. You know, I started as an engineer, and then moved into product manage, uh, well, and engineering management, then back into IC as a product manager, and then into a general management type role at Microsoft, then back into a functional role when I moved to Amazon, then back into general manager role with P&L, and then back into a functional role when I joined Square, then back... (laughs) You know, so it's, I've gone back and forth. I do like the general manager, the end-to-end, you know, multi-discipline, multifunctional leadership roles. Um, it just uses more parts of the brain, and as you get more senior and senior in leadership, you know, the job becomes more and more about how you both instigate and resolve conflict (laughs) um, in order to kind of keep things on the edge, like, on that creativity and execution threshold. And, um, the, kind of those, those problems in terms of either generating or resolving conflict are just more interesting when they're multifunctional, multidimensional in nature, it, for me.

   3. EGElad Gil4:02

      Yeah, that makes sense. Um, I guess, like, in the context of, um, Square and its foray into AI,

3. 4:09 – 7:46

   AI implementation at Square

   1. EGElad Gil4:09

      there's a set of areas that traditionally have been areas where payments companies have sort of applied ML. You know, so that'd be things like fraud detection or other areas like that. Um, could you tell us a little bit about how that evolution occurred at Square, and what areas you, you find most intriguing going forward?

   2. AHAlyssa Henry4:24

      Well, as you state, in financial services, um, application of machine learning has, um, you know, has been key for a long period of time. And particularly if you look at the, you know, a business like Square, where you've got millions of small customers and it's, it's not a one, you're, you don't really have a one-to-one relationship with the vast majority of your customers, just because of the scale and the size of them. Um, you really have to bring technology to bear in terms of understanding, like, a whole range of things from, you know, who's a good actor and who's a bad actor, um, to, you know, who do you target for a specific product, who, who's gonna be most likely to find, you know, a marketing product useful or least likely to find it useful, or that sort of thing. So, there's lots of internal applications and have been for years, um, in terms of machine learning, manage risk, you know, manage fraud, as well as to cross-sell and grow the business. And then, more recently, even kind of prior to this kind of latest, um, big shift in the AI landscape, we, we were using GPT-2, um, as part of, uh, the Square Messages product, um, being, uh, you know, a virtual assistant to help, um, customers, you know, answer responses to customer inquiries or a variety of things. But what, what's so exciting to me about, um, kind of really how the landscape has changed and the, the technology advances in the last year are how much better the tools have gotten, and how much more broadly applicable they are in terms of bringing kind of expert assistance to a much larger audience, right? But...It effectively unlocked the consumer and started to then show what this technology could do, um, when then, you know, further integrated into domain-specific areas. You know, you go talk to small business owners, most of them will tell you, "Gosh, I know I should be doing marketing," right? Like, "I, I know I, I, if I was more effective in doing that and reaching out with my customers, you know, I could drive more business. But I gotta tell you, you know, I work all day and then I come home at night and I gotta take care, you know, take care of my family. And then it's 8:00 PM and I'm starting to think about, gosh, you know, do I just, uh, chill for a minute or, you know, or am I gonna spend the next three hours trying to, you know, create an image and write text for the campaign and everything like that?" And what they'll tell you is like, "I know I should be doing this stuff, but it's just too hard and it takes too much time." And I'm not an expert. Like, I got into doing this 'cause I love cupcakes, not because I like writing email marketing, right? (laughs) Um, and so what's exciting about all this technology, you know, that's one example, but there's so many of these kind of different things where, um, just the, the ease of use and the accessibility opens up what previously was effectively just massive white space, right? It was customers or people that if it was easy enough to use, if it was accessible enough, if it was cheap enough, they go, "Yeah, that would be, that would be huge for me." But it was- wasn't accessible. It was too expensive. It was too hard to go find and hire a marketing consultant to do it for me, and the ROI wasn't there and blah, blah, blah. So I think this, this, you know, the evolution that's occurring right now is, is exciting in part just because of really the h- you know, previously unaddressed demand, um, that it's unlocking.

   3. EGElad Gil7:45

      Y- you've mentioned

4. 7:46 – 12:14

   Small business AI applications

   1. EGElad Gil7:46

      some really compelling ways that, um, different SMBs can really use generative AI, and I think one of the things that is a little bit underdiscussed in the AI world is the impact of this technology, um, particularly generative AI, to e-commerce or other forms of commerce and fintech in other areas. Are there other areas that you think nobody's really addressed yet or that are big opportunities in this space? Because, I mean, adopting GPT-2 was super early, right? You all used this technology before most people were aware that this was a big deal. And then to your point, there's some really interesting things that you've been doing in terms of merchant coaching and other areas that, you know, I think are, are really fascinating. Are, are there big areas of e-commerce that you just think are gonna be swept up in this technology that people aren't talking about enough?

   2. AHAlyssa Henry8:31

      Well, I think almost every aspect of kind of a small business, you can find applications and, you know, some of it, the technology's not quite there yet, um, but it's rapidly getting there in terms of some of the finance and numbers and quant pieces. Um, you know, some of the, the quantitative hallucinations have been a little bit more than, um, than the others. But- but it's all rapidly going and I think it's... You know, if you look at the largest e-commerce players, you know, the Amazons and the Walmarts, right? Like, they've been investing heavily in this area. So, you know, the, the larger e-commerce companies, um, have definitely embraced it. And frankly, in e-commerce in general, it's been for these retailers or for e-commerce retail platforms because one of the things is, um, if you're in e-commerce, basically everything's already digitized. So, um, one of the difference between, um, in-store commerce and particularly local commerce and, um, and e-commerce, is the fact that, you know, just to basically to operate in e-commerce, you have to d- you have to digitize. You know, you have to have images to, you know, show what your product is. You have to show... You have to have, you know, compelling description of it. You have to track inventory, you know. So there's a bunch of stuff that you have to have which m- this is w- again, some of the white space for small businesses, local business in particular, is, um, the rate of digitization in-store significantly lags the rate of digitization online. And that goes back to, it goes back to the fundamental problem of it's too either hard or expensive to, to effectively digitize. And again, this is where, you know, really in kind of all aspects I think the, um, Gen AI is lowering the... You know, making it 10X faster, 10X cheaper kind of thing, right? You know, we'd launched a... Squared launched a product couple years ago called, um, Photo Studio, uh, because we heard from businesses that they wanted to sell online, but they didn't have product photos, right? And so the first iteration of it is we actually had a photo st- like, like a physical photo studio in Brooklyn (laughs) -

   3. EGElad Gil10:32

      (laughs)

   4. AHAlyssa Henry10:33

      ... and, uh, and we had a 360 camera and people would ship us their products, right? You know (laughs) ? And, and, and we'd take what would look like professional-grade photos for them, which was because we were addressing this blocker that so many of them had to getting online, right? Um, but you fast-forward and we then evolved it into iPhone app that was using, you know, again, less mature versions of, you know, image detection and generation to re- use AI to remove backgrounds and things like that. Again, making it easier. You don't have to ship, reducing the cost. And now if you look at, um, what you can achieve with some of the latest stuff, like, the, the barriers come down even further. I, like, so it, it's incredible kind of all of these different things. You know, and then, if I'm talking about, you know, uh, the kind of selling size or the revenue generation side, um, but I think there's a l- there's a lot of back office as well, too. Um, a, you know, from employee, um, employee management and tools and communication to finances. You know, predicting and understanding what your cashflow actually looks like and, and what is your top-selling product and all these sorts of things, where a lot of the data was, is accessible, but again, you know, most of these owners, they're not MBAs, right? They didn't... The, the line is they got into business and they like working in their business, but not on their business. And so the business side of it is not the interesting part. It's the craft or the, you know, or the customer interaction or the hospitality of all those things. And so I think the, all of these advances that we're seeing are making it easier, um, and will make it easier, um, to operate with better expertise and less time and effort, um, on the business side of the business.

   5. EGElad Gil12:10

      Yeah.

   6. SGSarah Guo12:12

      There's so much there. I, uh, I feel like one

5. 12:14 – 15:04

   Latent demand for content generation

   1. SGSarah Guo12:14

      thing that's been a surprise to many tech people, business people has been, um, how much latent demand there is for content generation of, of different forms, right? In, like, every business context. I remember, um, y- like, a year ago, with things like Midjourney or with Stable Diffusion, like, you know, progressively better diffusion models, in general, there was a vein of like, "Oh, like, that's cute and it's, like, uh, a novelty, but really, how many artists are there in the world and how many people are really interested in art for the sake of art?" And you see it apply to, um, everything from, uh, like, product photography to being able to generate f- like, short form or spokesperson video. And, you know, the number of people, as you said, that want to avoid the camera, like, you know, the professional studio in Brooklyn, or don't wanna be on camera at all, or, um, uh, just want to do, um, uh, like, really attractive product photography or marketing and sales videos at 1/100th and 1/1000th of cost is, is quite large, right? And so I, I think it's, like, really interesting to think about the demand categories here for these new capabilities, which, like, a lot of them don't feel like traditional software businesses.

   2. AHAlyssa Henry13:34

      Yeah, and I think that's always been one of the amazing things about technology and one of the things that, um, you know, I find just so compelling about our industry is that when you can increase ease of use and accessibility, you just unlock all this latent demand, this white space that exists, right? Um, and you see it over and over again, right? You know, originally, what... who was the market for a word processor? Well, it was all the secretaries, right? Like, uh, they're the only ones that are gonna... Now ev- everyone uses a word processor, right? (laughs) You know, the TAM just exploded. You know, um, and even, you know, taking the Square example, how many people could accept a credit card and take a payment, right? Like, but there were all these really small businesses that were completely underserved because it was too expensive and too hard, right? And, you know, you make something better, cheaper, faster, and all of a sudden, you unleash, you know, all of this latent demand. And I think that's... we're in the process of that with this technology in, in sort of multiple vectors. Um, and so I think it is really gonna reshape a lot of things. And it's not just a... people like to talk about, "Oh, well, AI is gonna take away jobs." It's like, yeah, well, maybe, but, like, but a lot of it is actually work that's not getting done that could be ge- that someone could get done, right? Um, you know, it's like sending that marketing campaign or it's, um, actually, you know, putting together, you know, a, a real logo or it's, you know, writing better copy that actually is compelling and descriptive, the things that just were never gonna get done otherwise. And, and now they're getting done.

   3. SGSarah Guo15:03

      Yeah.

6. 15:04 – 16:54

   The origin story of Square’s GPT-2 products

   1. SGSarah Guo15:04

      Can I, uh, ask you to tell us the story of how you guys ended up, um, playing with GPT-2 for, uh... I think it was like a customer, like merchant-facing responses to begin with? Because that was... we have other, um... Elad and I, I think, each have other portfolio companies that were experimenting, but it was quite early. Like, it didn't really work or it took a lot of work to get something useful out from a-

   2. AHAlyssa Henry15:26

      Yeah.

   3. SGSarah Guo15:26

      ... just messaging perspective.

   4. AHAlyssa Henry15:28

      I would like to say it was maybe more strategic than it was, but Vinod Khosla was, uh, on our board for years, and, um, he had a portfolio company, um, that was being courted by another company to, to acquire them, and Vinod called me up and said, "Alyssa, you, you should go talk to these guys and see if maybe, like, Square might be, uh, an interesting place for them." So, I met the two founders, Stanford machine learning PhD folks, and they walked me through kind of what they were doing, um, with slightly different context, but just got super excited about the potential there as well as the, the two people on their team. And so we acquired that company in, I wanna say, 2018 or something, and, um, and put them to work on kind of stitching together some of these customer-facing experiences, um, leveraging some of the early work that they'd done. And because again, we just knew that, you know, there was a real customer problem for Square merchants, we... well, you know, goal was to apply technology to make our, um, merchant jobs easier and give them time back to, to focus on things that matter. So, that team's still going strong and continuing to, um, y- you know, expand the capabilities and obviously move further down the line in terms of models and whatnot.

   5. SGSarah Guo16:45

      So, one more thing on Square. I feel like I'd be remiss in not asking you after the almost decade you were working there, like, what else, even AI aside, do people not understand

7. 16:54 – 18:46

   Consolidating ecommerce workflows

   1. SGSarah Guo16:54

      is changing in e-commerce right now? More of the same in digitization? Anything else you, you think, like, trends people should understand?

   2. AHAlyssa Henry17:01

      Digitization is a big one, and I would say integration. It's even more so true in in-person commerce versus e-commerce, but it's true in both places in that you go watch a business owner or their team kind of work, and you're like, you, you watch their workflows, right? And, um, even the stuff that's digitized in many cases, what you see is they've got, you know, they've got multiple browser windows open and they're cutting and pasting from one tool into another tool, or they're downloading from this, you know, and then they're emailing it to that. And it's like the, um... I think the, the workflows, even, even the ones that are digitized are not integrated, and then, of course, the ones that are, that are manual and not o- not digital, you know, the integration i- is worse or nonexistent. Um, and so I think there's just a huge opportunity and I think we'll see the next phase kind of evolve, you know, in the same way that, you know, many industries and many parts of tech, what you see is kind of best of breed early on. Uh-... where different parts of the landscape are built out. But then ultimately, what you see i- you know, it's the classic bundling and unbundling. Right now, a lot of the stuff is unbundled, and I think we're gonna go into a bundling phase, um, because it addresses a number of things. It a- it addresses integration, um, and it also ... You can typically offer a bundle, um, you know, for less than the sum of the individual parts, and that kinda thing. So, I think we're, I think we're gonna see, you know, from an Indige perspective, kind of more aggregation and more bundling, um, because we've been through ... When we went to, when we went from in-person to e-commerce, a bunch of categories kinda got created, and we're gonna see the, the consolidation and the bundling of categories, and the blurring of lines between them.

   3. SGSarah Guo18:43

      I wanna go back a little bit, uh,

8. 18:46 – 23:07

   How will AI change cloud services

   1. SGSarah Guo18:46

      in your history. So, you were previously at the forefront of the cloud revolution for a long time as the first GM for AWS storage. And, um, you know, just beyond storage too, responsible for a huge number of innovations in computing that we all use now, S3, Glacier, Lambda, EBS. I'm missing some. Um, how do you think about ... Like, there, there are a lot of, uh ... There's a lot of discussion of AI changing the cloud services landscape. Like, do you see this as a new wave of computing?

   2. AHAlyssa Henry19:17

      It w- I mean, certainly there's a bunch of new aspects to it. Um, you know, uh, cloud computing historically, ev- you know, very, very CPU-intensive. Some GPU as well too. Um, various use cases. But obviously, you know, AI. You've just seen an explosion in GPU-based compute. Obviously, there's tons of demand right now just for compute capacity, um, for training. It's not replacing existing workloads. It's adding new workloads, um, as people are figuring out how to then expand, you know, companies like SquareBlock, figuring out then how to, you know ... How are we gonna apply these technologies? And then where do, where are we gonna go do, you know, go do our training and, and whatnot? So, I, I think it's an exciting time. Um, you know, one of the fun parts about being in technology is like, you just ... You get these big shifts that happen, um, you know? So it's n- never (laughs) a dull moment. And, uh, I, I think the race is definitely heating up, and, you know, it's in many ways the ... I think it's a land grab, and, you know, all, y- lots of different players are figuring out how they go grab land. (laughs)

   3. SGSarah Guo20:27

      One thing that's interesting about this wave is how, uh, monolithic the services are today. Uh, like, and you ... I think you can think of this as essentially year one or year two of, um, having access to, uh, these large foundation model services, right? But the interfaces are really simple. It's not even ... You were just talking about bundling. It's like, a single natural language call, versus if you contrast that to, um, like what ... You know. Th- the joke is you can't even keep track of, like, the Amazon services released. Right? (laughs) Um, like, do you think we get a wealth of services over time, the way many services have emerged in cloud? Or it's just you, you lob in, you know, more and more complicated prompts to a single model?

   4. AHAlyssa Henry21:06

      Um, it's probably both. Um, the ... You know, if you go back AWS at the beginning, right? Um, you know, S3 was sort of the, you know, the first ... SQS was actually technically the first, but S3 was really kinda the first service, and incredible simp- in- incredibly simple API. Right? Um, like, you know, four REST operators or something. (laughs) What was compelling about it is it was so easy to use, right? Um, and then, oh, obviously a whole bunch of stuff sprang o- up around it. Uh, you know, S3 became not just, you know, a, a first ... Well, was a first-class service done right with, you know, direct customer relationships, but it became foundational as well for many of the other services that were built on top of it. Right? So you go, you go trace kinda the, you know, the, the call stack, if you will, within most AWS services. Probably ... I would argue probably all of them. And then, you know, you can go find S3 somewhere as a component of it. Take OpenAI. Started, you know, relatively simple, but ... And, and adding to stuff. In fact, making them, you know, ease of use actually. Even though, uh, adding a ... new capabilities. In some ways, adding functionality that makes call patterns even simpler, right? With threads and messages and some of these other things. And so, I suspect we'll se- continue to see an evolution where we're gonna get some more capabilities that extend some of the core foundational, um, services. We're gonna see work that makes using them continue t- to simplify things that can be simplified! And then I do think we're gonna see, you know, um, some specialization as well, some additional model. I mean, you're already seeing some of this today, right? If you look at, um, you know, like the aggre- er, call it, you know, Amazon Bedrock. Right? You know, the aggregate service at some level, right? Where, you know, you can host and run all of these other different models. You know, it's a single service, but it's kind of a, it's kind of a bundle, if you will, of a variety of models. So, you know, we'll see. Uh, it's usually some combination of the both. Um, and, uh, I, I agree with you. We're, we're super early on right now.

   5. EGElad Gil23:04

      Uh, if you look at the major cloud providers today,

9. 23:07 – 25:16

   hyperscaler foundation models and the AI land grab

   1. EGElad Gil23:07

      um, two of the three have major alignment with a underlying foundation model, or a foundation model company. So for example, Google's, you know, uh, very publicly building out Gemini as sort of its next generation foundation model. Um, OpenAI has close alignment with Microsoft. Do you think it matters whether or not Amazon has its own close paired foundation model? Or do you think it's just, there'll be a lot of open source models with the integration of multiple third-party vendors? Like, does that, does that alignment at all matter as we think about the future oligopoly world of the cloud providers?

   2. AHAlyssa Henry23:41

      (laughs) I, like I said, I think it's a land grab, and lots of people are trying to figure out what's going on. But yeah, Amazon has the Anthropic, um, alignment and investment as well too, and so, you know, they're pretty close, although I think Google just annou- recently announced, uh, investment in Anthropic as well, um, I believe. Microsoft is also ... You know, they've got the OpenAI, but they've also said, "Hey, we're doing some of our own stuff." So I think ...I- it's unclear where- how this is all gonna shake out and who the winners are gonna be, so I think everyone seems to be placing multiple bets from a combination of part- you know, I- I think it's gonna be built by partner. It's probably all- do all three.

   3. EGElad Gil24:16

      From an end customer perspective, that makes a lot of sense because if you have an enterprise that's using your compute, there may be a variety of different models and approaches they wanna take. And so it does seem like the monolithic world seems reasonably unlikely, unless you're a model company worried about some competitive dynamic with the- with the underlying, um, cloud provider. But the flip side of it is all the cloud providers are also funding a lot of the different model companies, so... Um, so that makes sense.

   4. AHAlyssa Henry24:40

      It's not just the cloud providers are fund- are funding them, but it's also if you look at the marketplaces on the cloud providers, so like they have partnerships with all these people where, you know, to- to get down your Azure/AWS bill, you say, "I'm gonna spend this many millions of dollars on your cloud." Some of it's gonna be, "I'm using- I- yeah, it's- I'm using raw compute or storage or whatever," but a good chunk of it is also like, "I'm gonna buy this third-party partner through your app marketplace and use that to satisfy my quota of, you know, how much I pledged." So, I think all of these models, um, business models all intermesh.

   5. SGSarah Guo25:15

      Alyssa, I don't think you

10. 25:16 – 28:08

    Enterprise demand for open source models

    1. SGSarah Guo25:16

       will remember this conversation, or I'd be surprised if you did. But I asked you, I came to like ask you a question maybe in your first year at Square about, uh, I think it was like some Hadoop-related thing, so this really dates, um, this conversation now. But I was asking you about it, but it was some sort of like data infrastructure open source thing. And you just looked at me and you said, "Sarah, Amazon loves open source. We make more money off open source than any of the open source companies do." Right? And I think it's a- I think- first of all, that was like the- one of the most terrifying business conversations I've ever had. You were perfectly nice about it, but I was just like, "Oh my God, what am I doing doing these open source companies? She's right. This is terrible." Um, what do you- what do you think, um, happens in that landscape of like the open source models?

    2. AHAlyssa Henry26:03

       Um, well, there's certainly demand. Like th- there's strong customer demand for open source models, right? Um, for- enterprise demand for it, right? 'Cause it's, uh, it's not, quote, "black box." Um, you theoretically could in-house it all if you wanted. Um, so I think, you know, any time there's- there's demand, you know, the products will find a way into the marketplace. And any time there's a passionate developer community who, you know, is interested both in sort of giving to the community, but also it's a way to make your name as a, you know- as an engineer too by participating in these projects and, you know, being a core committer or whatnot. You know, I think open source is gonna continue to evolve. The question is, is, yeah, where- where and how do you make money off of it? Obviously, you know, there- there are some companies that, you know, have done well taking open source, or some of the founders or, you know, of the- of the project or whatever, and then start- you know, launching companies around it. I put Confluent, um, with Kafka in- in that bucket. Um, but yeah, there have been some others that have struggled. You know, Hadoop was there for a while, but kinda never- never quite got the commercial piece working well. Um, and I think just timing, not- not enough then differentiation, um, relative to what the cloud providers could do, um, pick up and- and go o- And, you know, I think one of the things with the cloud providers too is because you have sort of default customer demand, you launch one of these services using the open source. In many cases, you'll ha- you'll have existing customers that wanna use it, and so then you're immediately getting customer feedback, you know, to help improve and twe- and then help tweak it on your infrastructure. And so, you know, I think that's the- you know, that's the cat and mouse. But, you know, there's certainly customer demand, um, for open source in general, and certainly, um, uh, open source AI models, um, right now. And so, I think we'll see both.

    3. EGElad Gil28:02

       So the- the one other, um, you know, really, uh, impressive association you have is being on the board of

11. 28:08 – 31:02

    Startups in the AI semiconductor space

    1. EGElad Gil28:08

       Intel. And there's obviously been a variety of different computation waves that have occurred over time, you know? And it feels like every wave of, um, technology has an underlying, uh, different sort of massive, um, semiconductor company that emerges, right? And so we had, uh, Intel and AMD for the microcomputer revolution. We had ARM and Qualcomm as part of the mobile revolution. Uh, NVIDIA I think has really emerged as a big driver on the GPU side and sort of the AI revolution. Um, how do you think about, uh, startups in the AI/semiconductor space, and, um, you know, are there specific areas or paths that you think are most interesting or intriguing relative to this?

    2. AHAlyssa Henry28:48

       I think your observation is right about each kind of wave there being, you know, kind of a clear one and t- and two. Each wave there is also a three and a four and a- you know, you can kinda go down there. But I think there are- you know, in semis as in many, many industries, I do think there's a- you know, kind of a standard, you know, number one, number two, um, kind of market position, and it's really hard to be number three, and you're dead if you're number four. And so, I do think right now, you know, there's still... I- obviously NVIDIA is selling the most GPUs. That's fun. You know, that's driving, you know, a lot of AI. But, um, but I think there's still room. I don't think it's gonna be a monopoly on the area. Um, and, you know, I- I do think there will be a clear number two. Um, and right now, there's opportunities like for, you know, different companies that are- that are cr- I think running towards it and trying to, um, trying to take that position and then perhaps over time challenge number one. We're kinda going through the standard thing where you're- as- as you progress in each of these, the tooling goes up the stack, so you start to see some things where maybe...Things were more coded to the metal. You start to then, you know, the, it shifts to, to tools, and that's part of what then creates an opportunity, you know, for one and two and that kind of thing as well too.

    3. EGElad Gil30:06

       That makes sense, yeah. A lot of people talk about the defensibility of, um, NVIDIA in part being due to CUDA and then some aspects of Interconnect and things like that. Um, a, the other thing that I hear sometimes people talk about is, just as there was s- sort of this very positive dynamic in the Wintel world, you know, Windows and Intel reinforcing each other, maybe today that's kind of the GPU transformer world where GPUs are in part optimized now more and more for transformer-based workloads. And transformers are obviously, have been optimized by large armies of people relative to GPU. And so, you may also have a virtuous cycle through that, um, on a relative basis as s- sort of a second driver on top of what you're saying in terms of that software stack being really relevant or important.

    4. AHAlyssa Henry30:44

       Yeah.

    5. EGElad Gil30:44

       Um, so it'll be interesting for sure to watch that.

    6. AHAlyssa Henry30:47

       The two pieces have always kind of gone together, right? And, um, you know, and it's a push and a pull on both sides. Certainly, you know, having worked at Microsoft, you know, in the, in the '90s, right, you've been very close with, uh, with Intel at the time.

    7. SGSarah Guo31:00

       On the, uh,

12. 31:02 – 34:32

    Scale up architectures vs scaling out

    1. SGSarah Guo31:02

       AI accelerator, um, topic, it, it's really interesting. Like, I think the, the view of many of the AI semi-startups is that actually, uh, GPUs, they're really good at matrix multiplication, but they're not specifically tuned to transformers' architectures. Um, but it's very hard to make long-term bets in AI right now with how quickly everything is changing. Like, even from an architectural perspective, I think people are talking about the limitations of, um, uh, the attention mechanisms that we have and experimenting on the research side with different architectures in a way that they were not three months ago. And if you had asked me, I'd be curious if there's, like, interesting survey data around this. We can look for it or, or ask researchers. But if you had asked me, um, how committed are people to transformers as the dominant architecture, uh, at the end of 2023, I'd say very committed, and I feel less confident about that today.

    2. AHAlyssa Henry32:04

       Yeah, I mean, any time you're in a s- uh, in a phase of kind of rapid change, right? You know, they love the Jeff Bezos quote, you know, focus on things that, you know, you know will not change, um, because (laughs) you know, there's gonna be so many of you, like, trying to make a bet on things that, you know, are not just fundamentally true. You're gambling at some level. Um, what I, what I think is interesting in a lot of this stuff though is, um, again, technology often goes through these kind of cycles where what you see is, you see scale up, th- you know, scale-up architectures to a pla- point, and then you reach some sort of a s- tipping point where you just c- you're not making as a much, as much process on sort of a, a scale-up architecture. And so, then you start to break it down, and you kind of just scale out, and then you kind of rego... Anyway, so we've been through these curves multiple times. I think what's sort of interesting in the, in the semi space, uh, in many ways it's been kind of a scale-up model for years. And I think part of what's happening, and this is already underway, right? But you're starting to see, you know, more, yeah, more of these chiplet designs, more use of advanced packaging, and like, and it's really starting to look more like in some ways almost like a, you know, a microservices architecture, whatever (laughs) you draw the software analogy. But, but I think one of the things about... One of the reasons, you know, kind of a, a system design architecture is interesting is because, um, in some ways it allows you to predict at smaller, and like, you, you, you can tune and predict smaller components, and you can rip out and replace components rather than having to kind of change the whole thing. And so, I think we're, we're, we're definitely moving towards more and more modular architectures, and I think that's gonna give more and more flexibility, which then I think can help accelerate the innovation cycle as well.

    3. EGElad Gil33:42

       Yeah, it also feels like at this point of any innovation cycle, there's always, um, a, a ton of experimentation that happens, and, you know, that, that happened, uh, I think on everything from social products to more recently with crypto, where there was all these different L1s that were invented to be s- sort of scalability alternatives to Ethereum, and then it moved into L2. And that, it just feels like every way of computing you suddenly have this burst of, "Well, this thing is really working, let's try five other things that could work potentially better." And then, you know, I, it feels like 90% of the time-

    4. AHAlyssa Henry34:14

       (laughs)

    5. EGElad Gil34:14

       ... it collapses back to the original thing that you just keep scaling it or whatever, but we'll see. Um, I guess, you know, you had this-

    6. AHAlyssa Henry34:20

       Darwinism (laughs) .

    7. EGElad Gil34:20

       Yeah, Dar- (laughs) yeah. Yeah, Darwin selects for the thing that keeps going, so-

    8. SGSarah Guo34:24

       A lot of wants to bring back the monolith. (laughs)

    9. AHAlyssa Henry34:27

       (laughs)

    10. EGElad Gil34:27

        Yeah, I love good monoliths, and, uh, I really love, uh... No, I'm just kidding.

13. 34:32 – 36:08

    What’s next for Alyssa

    1. EGElad Gil34:32

       Um, you know, you've had, I think, one of the most impressive careers in technology in all sorts of different ways. You've a- been involved with some of the most important companies in the world. And, and literally every decade that you've operated, you've been at-

    2. AHAlyssa Henry34:43

       Thank you.

    3. EGElad Gil34:44

       ... um, the most or one of the most important companies. What's next?

    4. AHAlyssa Henry34:49

       (laughs) I don't know.

    5. EGElad Gil34:49

       Like how do you think about the next couple years, the next decade?

    6. AHAlyssa Henry34:53

       Um, I don't know. Uh, it's a good question, and it's one I'm working on figuring out. Um, my husband retired from work two years ago, and for the last two years he's been like, "Alyssa, come out and play with me. Come play with me."

    7. EGElad Gil35:06

       (laughs)

    8. AHAlyssa Henry35:06

       You know, I'm like, "Ha, I'm working on it." Like, and and we working 78 hours a week, right? (laughs) Yeah, it's like, "What are you doing, girl," you know?

    9. EGElad Gil35:13

       (laughs)

    10. AHAlyssa Henry35:13

        So, um-

    11. SGSarah Guo35:15

        Aw.

    12. AHAlyssa Henry35:15

        (laughs) And I'm like, "Well..." (laughs) Um, so, you know, I think I, obviously I like, I, I love technology. I love deep technology. Square was super interesting. It was the highest up the stack I'd ever really worked. It was the first time ever working on, you know, financial stuff, Intel and Confluent and whatnot where you kind of s- help scratch the itch of I still like... Fundamental technology, like, it's fundamental, right? I mean, there's something, there's something to it. And so, still reading stuff, still, like, you know, I watched the, you know, watch the OpenAI Dev Day, right, you know, tinkering around, but he's, my husband trying to keep me as busy as possible (laughs) you know, and, and running around with him. So we'll see, I don't know. Um, I don't know if this is gonna be a permanent retirement or a, you know, or a sabbati- you know, a, a sabbatical kind of thing. Don't know. We'll see. So I'd love to hear, like, what are you guys most excited about in this space? If you could pick kind of one thing, what are you, you know,

14. 36:08 – 39:35

    What Elad and Sarah are excited about in 2024

    1. AHAlyssa Henry36:08

       for the next year ahead, what do you, what do you hope to see or, you know, what do you hope to, you know, be involved with?

    2. SGSarah Guo36:15

       I'm gonna make a lot go. I actually invested in all of the good ideas this year, so I'm gonna-

    3. EGElad Gil36:20

       (laughs)

    4. SGSarah Guo36:20

       ... take the next year off.

    5. EGElad Gil36:21

       She's also announcing her retirement at the same time. (laughs)

    6. SGSarah Guo36:24

       Half decade's in, I'm done.

    7. EGElad Gil36:25

       Yeah, she's just over with it. You know, if you basically look at the last year... And it's only been a year since ChatGPT came out, right? And it's been a year and change since Midjourney and Stable Diffusion came out. And so I think really the last year has just been everybody waking up to the opportunity of what could happen with generative AI. I think there's been an enormous amount of investment in the foundation model side, and I still think there's lots of open questions in terms of where does that all go. But I think we know at least a handful of who the incumbents, at least in the next couple years will be. Maybe not forever, but at least for the next three or four years. You're starting to see the infrastructure side start to get filled out in different ways. And for me, the area that is still wide open is just all the various applications, both on the B2B side as well as the consumer side. And there's an enormous amount of white space there and a lot of open things to do. And so I think that's just a huge transition that's coming, both in terms of incumbents adopting AI to their existing workflows, as well as a huge chunk of the services economy being converted into, into code. And in this case, instead of traditional software being converted into, into AI. And so I think that's probably the story of the next two, three years.

    8. SGSarah Guo37:25

       I'm, I'm excited about it. I'm not actually gonna take the next year off. I think, (laughs) I think we're very early in the exploit cycle. Like, one of the things that has been even, um, surprising over the last six months is when you are doing a lot of very new things with technology, um, uh, there's a rush to try them once it has been proven. My favorite example would be like eight by eight, like diffusion model image generation and how far we have come in the last two years from when people said like, "Oh, look, this is possible, but the quality, the controllability, the usefulness of this is, like, not even close." And even a few months ago, like y- you'd have leading researchers that say like, "Ah, video, like who knows if..." You know, that, that's a, that's a huge number of technical problems that feel unsolvable. And, um, I think you increasingly see on, um, the creative fronts, but many different applications, like how quickly you can get over some, um, some minimum quality that is useful, right? And it takes people who are playing with GPT-2 or like an eight by eight image to picture what quality and scale improvement can, um, can happen. But I think a lot of really smart engineers are paying attention to that now, and I think that'll accelerate the exploit cycle a lot. So, um, it, you know, it's, it's new UI. Um, it's a lot of, as you were talking about, uh, um, latent demand, so it's not as obvious where you're not like, "Ah, like I'm just gonna replace this existing software category." But I, I'm really excited that a lot of that experimentation is gonna happen in a more sophisticated way at the application layer next year, and you get to, like, ride the capability curve too. So, um, no, no, no retirement this next year. Alyssa, thank you so much for joining us.

    9. EGElad Gil39:11

       Thanks for having me.

    10. SGSarah Guo39:12

        Find us on Twitter @nopriorspod. Subscribe to our YouTube channel if you wanna see our faces. Follow the show on Apple Podcasts, Spotify, or wherever you listen. That way you get a new episode every week. And sign up for emails or find transcripts for every episode at no-priors.com. (instrumental music)

Episode duration: 39:35