John Locke (00:23):

Hey everybody, this is John Locke with Spotlight On, and I'm joined with Jack Krawczky, who is a former entrepreneur in residence at Accel and current product lead for Google Bard. Welcome to Spotlight On.

‍

Jack Krawczyk (00:36):

Thanks for having me, John. It's great to be here.

‍

Jack’s upbringing and background

‍

John Locke (00:37):

It's nice to have you here. I wanted to start out with your background because I think you've got a really interesting journey from growing up in Pittsburgh to Google. Tell us about that experience, how you got from Pittsburgh, dude to product lead for Google Bard.

‍

Jack Krawczyk (00:56):

Well, I did grow up outside of Pittsburgh to an immigrant family. We actually immigrated to the US in the eighties from Poland and had this real kind of cultural divide of being the immigrant family in a town that grew up from natural gas mining into the steel industry in the 1900s. And just sort of watching this transformation that's taken place in Pittsburgh as an immigrant into the US.

‍

John Locke (01:27):

And were you born in Poland or were you born in Pittsburgh?

‍

Jack Krawczyk (01:29):

I was born in Poland.

‍

John Locke (01:30):

You were born in Poland? Yeah. And how old were you when you came over?

‍

Jack Krawczyk (01:33):

I was three years old.

‍

John Locke (01:34):

Okay. Yeah. And how did your family end up in Pittsburgh choosing Pittsburgh?

‍

Jack Krawczyk (01:39):

Engineering has been at the heart of my family. My father now retired, was an electrical engineer and there was a joint venture between General Electric and the Mitsubishi company called Power X that made semiconductors in Western Pennsylvania. And those sort of industrial roots, being interested in math, how the world works, how the world can work better was just deeply ingrained in me in both my upbringing and the town that I grew up. And seeing the power that industries have to change when technological really revolution takes place is something that is really fueling me today. Working on AI, we've seen what automation has done in the manufacturing industry. This world that we're entering in with generative AI specifically allowing you to create better ideas in less time and higher volume, we're already seeing the impacts that that has on businesses like consulting, advertising agencies, even software development. People are able to create better ideas and less time that's really about to change the service industry that as we know it.

‍

John Locke (02:53):

I think it's so cool that you have engineering roots in your family and that's what brought you to Pittsburgh. So you come to Pittsburgh when you're three, you grew up in Pittsburgh, and did you always have some inkling that you wanted to be an engineer or you wanted to be in technology or when did it become clear to you that that's what career that you wanted to have?

‍

Jack Krawczyk (03:18):

I remember playing this video game called Jill of the Jungle.

‍

John Locke (03:23):

It's not the one I thought you were going to go with. I thought it was going to be like Zelda or Tech Mobile.

‍

Jack Krawczyk (03:27):

Well, growing up in an immigrant family that didn't have too many means we couldn't afford a Nintendo. And so at Marshall's, I purchased a 99 cent PC game called Jill the Jungle. That was a scrolling platform game. And I just remember having this experience of like, you can make computers do that.

‍

John Locke (03:50):

You sell a lot of yourself in Jill of the Jungle.

‍

Jack Krawczyk (03:53):

I mean, it's a classic Jack and Jill story. Just that feeling of, wow, something can just render these pixels on a screen and do something interesting and sort of captivate you in an interesting way. And so I wanted from that point on to learn how it all works. And I'm sure my story is not unique to many people of falling in love with computing and video games, but I wanted to learn how it worked. That's when I started really getting into computer programming and as many people who were on the internet around those days can attest to dialing into AOL was a little bit tricky. So we used to pull together the probes of the world and create automated dialers with visual basic, scripting early on was a deep set passion just to get on the internet to who knows, I don't want to talk about all the things that people would do on AOL in the mid nineties.

‍

How Jack ended up at Google (the first time)

‍

John Locke (04:52):

So you go through high school and you go to Carnegie Mellon, stay in Pittsburgh. And then how did you get from Carnegie Mellon to Google?

‍

Jack Krawczyk (05:05):

So my first job out of college was working in high frequency trading systems. So I had this passion for math over the span of my education, and I thought really high volume number problems were really fascinating. And this was around the time, I guess then it was called algorithmic trading. You started to go into a world where it wasn't just people trading on the New York Stock Exchange, it was computer speaking to computers to make trades of not only stocks, but foreign exchange and fixed income. And I didn't have the words for it then, but I realized that working in technology, being around mission orientation was really key to my personal passion. And so we had this mission around that time in 2006 as I'm ramping up in my analyst years of algorithmic trading is really emerging. We can find inefficiencies in markets between credit and foreign exchange and equities and allow people to invest their money in efficient ways, sort of minimize arbitrage in trading environments and really create fundamental based investing.

‍

(06:23): And I thought that mission was so amazing and I thought technology would be able to enable it. Then February of 2007 happens and the whole efficient market project gets canceled because we know the world is going to go upside down as the mortgage crisis starts to bubble. It's about a year before it really hits and project gets canceled. And I remember sitting in this all hands meeting and our managing directors in there telling us what we're going to do. And I raised my hand. I'm like, if we know the world's going to go haywire, shouldn't we maybe try to build something to stop that from happening? And I'll never forget that moment where in front of a large room, I think to embarrass me, he responds with, do you have any idea how we make money in this business? And the reality was they made money on volatility, on volatility and trading.

‍

(07:17) And I just remember feeling so defeated at that time that I'm like, wait, I'm just building something to extract value from the world, not create it. And so just on a whim, I get home that night, I polish up my resume of a year and a half working in banking, and I just randomly apply to a job at Google called Financial Services Industry Specialist. Sure, I'd probably check those four words. And it was this amazing time at Google where I think they were getting like a million resumes a month to work there.

‍

John Locke (07:58):

And this is 2007, 2008?

‍

Jack Krawczyk (08:02):

February of 2007. And I get the interview and I pass whatever filters came through. I go to the interview.

‍

John Locke (08:10):

You're living in New York at the time, but this is before Google had the New York office.

‍

Jack Krawczyk (08:15):

Google had the New York office. There was something about California that was really interesting and compelling, and I knew the East Coast but didn't know California. So I fly out here and think like, oh, Google got to get ready for all these classic brain teasers, these amazing things. And I get asked this question in the interview by one of the interviewers, which is, can you explain how a mortgage works? And I think to myself, wow, what a classic brain teaser question. We've got to get into the fundamentals of how collateralization works, etc. And I explained it. I hope I did a good job. I ended up getting the job, spoiler alert, and then I joined Google in September of 2007. I joined them for the first time. And I find this person during the interview and I said, I thought that was such a classic question. It was so amazing. And this person goes, oh no, I was buying a house and I genuinely hadn't done the Google search yet. It was a great lesson to learn early on in my life. No one has all the answers. People

‍

John Locke (09:30):

People have other things going on.

‍

Jack Krawczyk (09:31):

So much of what we're trying to do is disambiguate the world. And so I don't think there's any one person in the world that has all the answers. But I came to Google and I just saw this amazing thing, which was the value of curiosity. The curiosity of how might the world work, what are some of the questions we might ask to get deep questions around? And early on, I had a great opportunity to work with our chief economist on a project where we saw ad-click behavior during the financial crisis. So I started September of 2007. Things started going haywire in 2008, late 2008. Hey, it seems like there's a lot of searches for brokerages around the time of markets going crazy. And so we run this analysis to see what happens during the vix, the volatility index being extremely high, and we learn well, people are more likely to open a brokerage account when markets are volatile because they think they have things that are happening.

‍

(10:38): We're taught buy low, sell high. People were buying high markets, having a good day, people are opening more brokerage accounts, market's having a bad day, people opening fewer brokerage accounts. And we started to learn the sort of curiosity of wait, how is people's behavior working? And we were able to translate that into creating helpful bidding algorithms for our customers of the day. And I think back to that a lot now, 15 years later, being in this generative AI moment, that same curiosity from 15 years ago is being applied again. What happens in this world where we've gone from computers do things for you, to now doing things with you, helping give you the words on how to have a challenging conversation, helping you clarify your ideas for that pitch that you're about to come and give Accel to get the funding that you need for your business. It's kind of remarkable and profound in that we're giving people a new capability that's never existed. And so that curiosity, as much as it existed 15 years ago when I started at Google, I guess 16 years ago, it's still there. And that's part of the magic. 

‍

The origins of Bard, and the balance between helpfulness and limitations

‍

John Locke (12:01):

I want to go back to the beginnings of at least the Bard team that came together at Google. And I think the story of how you got engaged on that team is an interesting one because when you came back to Google for the second time, you didn't necessarily come back to Google to be project lead for Bard. You came back to work on other projects at Google. So going back to the end of last year, at least from my view, from the outside, it seemed like Bard stuff was starting to come together. Walk us through how you got involved and how the Bard team came to coalesce.

‍

Jack Krawczyk (12:44):

Yeah, so when I came back to Google in 2020, I had the opportunity to work on Google Assistant. So digital assistants across, Hey Google, Siri, Alexa, had a couple of years of really being entrenched into our daily lives, not just on our phones, but in smart speakers, TVs, smart displays, et cetera. And it was really interesting to see this form of conversational computing emerge, but a lot of these things were one-off questions, what's the score of the sports game? How do I make brownies, et cetera. It wasn't a deep conversation that we wanted computers to have in sort of the same way that has been imagined in things like Star Trek. And so early on I had the opportunity to play with what was then called mena, then became Lambda language model for dialogue applications. It was sort of the first chat like interface where you could ask a question or prompt these language models and get a very compelling sounding response.

‍

(13:55): And there was this deep desire to put it into a product like Google Assistant. But the challenge was and still is, these things are very confident even when they're very wrong. They're designed to create compelling-sounding language and that can come at the expense of things like making things up. It's called hallucination now that it has a fancy term. And we were looking, what are ways that you can bring it into something like Google Assistant? And as we interacted with it, also around that time, there's a public narrative that starts to emerge. The sort of fear of AI, AI is going to lead to extinction, etc. Now, I don't want to trivialize that conversation because I think there's a lot of important safety measures that we need to think about with this technology, but the opportunity was profound in there and it started to become clear that building this directly into Google Assistant as a first step wasn't going to be the right way to set the proper expectations of, Hey, this thing's a possibility generator. This thing's a creative collaborator. How are you going to get people to think of it that way? And so our bet became create Bard right around the same time chat GPT enters into the world in lexicon, and the conversation around the world goes away from AI is going to make us extinct to AI is going to make us extinct, but oh my goodness, look at this amazing compelling thing too. And that excitement, the imagination that triggered, and it's hard to believe that was a year ago.

‍

(15:41): We are barely 12 months into this consumer adoption cycle and narratives want to be built that the laws of physics don't apply. Consumer adoption is completely fundamentally changed. Like no, we're still in the hacker days of this, and we have to think about how we get this technology to be something that's truly helpful. It's why we launched Bard as an experiment. We want to say, Hey, we think this is helpful, but we also know it has a ton of limitations. The balance between helpful and limitations, I will admit is moving more toward helpful, faster than I think anyone around the world has expected, but it's still in its infancy as a concept.

‍

Measuring success, sourcing user feedback and competing with ChatGPT 

‍

John Locke (16:24):

And so given that, and it is also much in its startup phase still, when you're looking at Bard from the perspective of the team managing it, what are the things that you do as the Bard team to measure success given this is still so early?

‍

Jack Krawczyk (16:46):

Feedback from people using it is the most important thing. What are the things that people are finding helpful? Where are the areas where it's still challenged? And this is where standard product development tactics come in, you try to understand a problem that you're solving, solve it. What can you learn and move on? So we know hallucination is this challenge.

‍

John Locke (17:11):

I want to ask you one thing about that - because of the massive scale of it, what do you use to get a good feel for user research? Because there's so many different software applications for it, there's old fashioned ways to do it. When you're trying to get into the head of your user, how do you do it?

‍

Jack Krawczyk (17:38):

Will say nothing beats just talking to people that use it.

‍

Jack Krawczyk (17:44):

People that use it frequently, people that use it sparingly. And there are different ways to do that. I found, I guess it's called X now.

‍

John Locke (17:55):

Indeed.

‍

Jack Krawczyk (17:58):

Can I still call it Twitter?

‍

John Locke (17:59):

Let's call it Twitter. 

‍

Jack Krawczyk (18:01):

Going to Twitter and talking about what we're doing and also being very transparent. So we have an updates page on the Bard experiment where every time we release something new, we don't only say what we're doing, we explain why we're doing it, to invite a conversation, is this helpful? Is this not helpful? Going on there, talking to people, whether it's through direct messages or replies, it gives you a very amplified view. I wouldn't say it's the most indicative of how people are using it, but it sort of gives you the extremes, the memes that start to emerge of what's helpful and what's not helpful. We have a Discord server that we've been working through developing a healthy community through people sharing their stories, being a fly on the wall, participating in some of those conversations, and you start to hear these amazing stories. We had a person go on there who shared their story. They are neurodivergent diagnosed with autism, and they have been using Bard to help create responses to emails that they fear may frustrate the person on the other end.

‍

John Locke (19:13):

Amazing.

‍

Jack Krawczyk (19:15):

And starts to fill this gap. Then there are the idea developers. There was this local business entrepreneur that comes in there and says, I've been uploading screenshots of my pitch deck to Bard and I'm going to my bank for a loan. How can I use language to better articulate what I'm trying to do? And getting ideas on how to make a stronger pitch deck and that sense of research of what people are using it for I think is really key. But then there's also, we have great user research teams inside of Google that are looking at not just understanding our current user, but anchoring on, Hey, three quarters of the world still has not used this technology. And it's not a lack of awareness, it's a lack of awareness of how to make it helpful for them. So understanding the concerns, channeling that into not just the products we build, but influencing the go-to-market strategies that we build as well.

‍

John Locke (20:10):

At this stage, how much at all are you focused on Bard versus ChatGPT? Because this is one of the few areas where there's been a really strong Google challenger, at least on something that is you all have identified as core to your business. So in this stage where you have two of you doing it at real user scale that are head to head, but it's still 12 months into the journey, how do you think about the competitive dynamics there? 

‍

Jack Krawczyk (20:54):

So I talked about the three respects at Google, respecting the opportunity and respecting the user, are two of those three, when you look at the opportunity that's in front of us, it would be irresponsible to not acknowledge people have choice when they're coming to clarify their thoughts, to develop their ideas, to brainstorm into bringing something from idea to reality. And so of course there's an awareness of the other alternative products that people have choice to when it comes to selecting Bard, where we genuinely spend the majority of our focus is of the people that are using it, what are the things that are working really well? How can we get them to feel like it is maximizing that helpfulness in their life? There's also the world of people that haven't used this technology are still the vast majority of the world. There's a depth of understanding that we need to develop why they haven't made that leap. And so where we're really focused on is can you understand that problem deeply?

‍

How Jack uses Bard in his everyday work and life‍

John Locke (21:56 ):

And do you find, because you have the most recent versions of this at your fingertips, when you look at your day-to-day work, how often are you incorporating those kinds of prompts and questions and interaction with Bard into what your day-to-day work is at Google?

‍

Jack Krawczyk (22:17  ):

Bard has just become a part of my life in everything. It's not the AI producing the idea, it's the human that's trying to find the right words to make a pitch, refine a concept, explore an avenue of solving a problem that they hadn't considered before. So what you're going to see is people in their jobs that they have today, being able to produce more in less time, and I believe that creates amazing economic opportunity because in the beginning, it'll just be, Hey, our jobs are going to be a little bit easier, but then, oh my goodness, it's profound what we can create. 

‍

John Locke (23:01):

Yeah. I think having heard you talk about this before, some of the user research that you've done suggests that actually people that are getting the most out of it, very early days and we're under 12 months in are the small business owner that needs help on a marketing campaign that doesn't necessarily have a marketing team. It's those kinds of examples that you seem like you're hearing more and more as common refrains from people that are using Bard to the fullest at least so far. 

‍

Jack Krawczyk (23:33):

It gives you this feeling of a superpower of productivity because it can fill in gaps that you know you have. For example, I grew up in a household of immigrant parents, English as a second language. They would actually turn to my sister and me for help proofreading correspondence that they would have to want to make it sound more professional. The fact that they would now have these tools to be able to say, what do you think will be interpreted by this? How do you think I could make this sound more professional? How could I make it sound more stern or less stern? These sorts of things are applicable to everyone. I see a lot of adoption in the small and medium business world because of there are wide factors. How do you make sure all the data silo and security is meeting the needs of what IT departments are? That's why we have a Google Cloud business. I think there's a lot of interesting opportunities that are going to emerge around data siloing and the creation and cultivation of fine tuning data to get the maximum horsepower out of these models, but it's really remarkable to see this is a generally applicable technology, the first one that we've seen, I think, since cloud computing.

‍

The inputs, outputs and transparency of AI models

‍

John Locke (24:54):

Yeah, really interesting. How important do you think it is to have some transparency around the inputs that go into the model itself? What is the best way for you all to articulate that to the public so that everyone believes that this idea of the models being based on reputable sources, as you're saying, is as transparent as possible?

‍

Jack Krawczyk (25:26):

I think a lot of scrutiny is put around the input mechanisms for these models when the true magnitude of impact is on the outputs that they produce. I think on the input side, it's an important conversation to have. We've taken steps for publishers across the web to be able to opt out of the training of future models and not only opt out of that, opt out of some of these DoubleCheck features, like what I mentioned earlier, that choice and control is absolutely key for publishers, and so we want to provide that transparency. But again, these things are programmed to produce compelling sounding text in a given language, and so that's why focusing on the output, those measures that we talked about, labeling the limitations, providing tools like fact checking as we do with the Google IT button on Bard, that's really where the importance of the assertion of limitation is key. A tool that's going to help you come up with hypothetical interview questions has a much different risk profile than tool that's going to be using a language model to produce medically oriented content. And we've got to focus on what those risks are based at the application layer than necessarily at the input layer. The input layer of course requires control and choice on behalf of the creators, but it's really the output that we believe needs the most,

‍

Operating as a founder inside an existing company, and one of Jack’s favorite Bard use cases

‍

John Locke (27:03):

Yeah, makes sense. I want to move to just talking about how you have set up the Bard effort at Google because we have a number of companies in the Accel family that get to some scale, maybe not the scale of Google, but get to some scale and for whatever reason want to launch effectively like a startup team or organization within a big company to tackle a new problem. And there's so many different examples of this happening over the years, but this is the Bard and what you're doing in generative AI within the broader Google family is really playing out in the public eye and you've got the added layer of complexity of a company and open AI that has raised a stunning amount of capital and has a really incredible set of people around it. How are you thinking of, walk us through what it has been like to launch this as effectively a startup within Google, and what are some of the lessons that you've learned from that process that our founders could apply anytime something like that comes up within their business? 

‍

Jack Krawczyk (28:10):

The first thing I'll say is I'm very remiss to use the term of launching a startup inside a large company because I feel extremely fortunate that I get all the fun parts of the startup life, which is looking at the opportunity ahead, being able to focus on the product and the go-to-market pieces. And as a personal leader, I know there's an amazing team that's focusing on all of the HR aspects of what we're doing, all the finances, et cetera. Of course I'm aware of those things, but I get to be able to squarely focus. And that to me is if you are one of these companies trying to start an effort like this, creating that ability to apply a founder like obsession with generating insight and translating it into value, that's how you set it up for success. And so we really value, as much as the world talks about the speed of delivery on Bard, which I'm extremely proud of, what we're driven by is the speed of insight.

‍

(29:12): This is a fundamentally new technology. Internet technology has always done things for us now it's helping do things with us, the insights we need to generate of, hey, let's start with a hypothesis of a creative collaborator with AI Bard, okay, standalone web experience. How quickly can we learn whether or not that's a true application? You start to learn, you start to see those things. You start to say, Hey, this is really cool, but people are suggesting that they want visual input. We know the models have that capability. You launch image input as we did in July, and you start to see things manifest, like people uploading a picture of their business pitch to figure out how to increase the probability of getting a loan at their local bank. It is remarkable to see those insights and creating that culture of translating an insight into delivery. That's what you need to focus on as a founder inside your own company.

‍

John Locke (30:11):

What are some of the things that you've been most surprised by? Again, we're what, 10 months in or so to this journey? It feels like it's been a lot longer than that. What are some of the things, as you've put this out and you've spent all this time on user feedback and research and you've seen it spread around the world, what are some things that you've been surprised by in terms of how people are using the application?

‍

Jack Krawczyk (30:37):

It can bring out some of the best in humanity, I was speaking with a leader of a foundation in Eastern Europe that was working with another foundation in Taiwan about how to create a refugee program for Ukrainians in Eastern Europe. And this person told me, Taiwanese culture is very different from the eastern European culture that I'm used to. And so we created this MOUA memorandum of understanding and there's a cultural divide. And I was using Bard to help me understand and anticipate what some of that cultural divide was going to be so that we could focus more on delivering the value of this refugee program.

‍

(31:24): And I'm just sitting there having this conversation being like those crazy insane hours that my teammates and I put in, and people are using this to help bridge cultural divides. People have inherent goodness, people have remarkable imagination, and the fact that they want to be able to apply it to make it better, to create good in this world, I feel like a lot of the conversation happens around the risks and what could go wrong. And of course it's important to focus on that, but we can't forget about the inherent good that people have and want to do in the world. And that just, I mean, that's what gets me inspired every single day to work on this thing.

‍

A mental framework for how businesses can think about incorporating Generative AI 

‍

John Locke (32:10):

Yeah, it's fantastic. It's fantastic. For any organization or business that is wanting to figure out ways that they can use generative AI to be better at whatever they do. What are your suggestions on just how you jump in? Because I do think we've gone through, we went through a period this spring where almost all of the businesses in the Accel family in one way or another, we're trying to figure out ways that you could use the technology through making software faster and SEO or whatever it may be. And now I think people, I think for the most part, companies are realizing maybe it's not this just panacea that's going to make the business revolutionize every business, but there's still all these ways that it can make companies better. What would your advice be if you were on the board of any just general business or organization that wanted to get involved but wasn't sure exactly what the best way to apply the technology was to their respective business?

‍

Jack Krawczyk (33:19):

I'm eager to learn more from some of the founders that I've heard talking about it as we treat it not as adopting software, but hiring programs. So if we were to hire a helping team to help our existing teams, whether it be in customer service, product management, marketing, et cetera, come up with deeper analysis, come up with more creative ideas, what are some of the ways that we could do that? And then conversely, how could we take that thinking and apply it into how we position our products? I think that level of thinking is still nascent. The concern it may evoke, especially if you're putting it into part of your product is, oh my goodness is are you saying that we should replace humans? And so that's where I think the go-to-market effort. If you're thinking about this not as how to apply it in your business, but in your product offering, you need to spend a ton of time being very, very clear about the value that it provides.

‍

John Locke (34:24):

Give us a little sneak peek as we head into the end of the year. What are you most excited about in 2024 in this area?

‍

What Jack’s excited about for generative AI in 2024, and building a product for a global audience

‍

Jack Krawczyk (34:34):

One of the things I'm most excited about is combining the world of generative AI, doing things with you, with established paradigms of tools that do things for you. And so we announced assistant with Bard, which is taking, it's kind of near and dear in my heart. I had the chance to work on assistant for three years before Bard taking this product that has really provided value to hundreds of millions of people around the world for helping them get things done. Setting that timer while you're busy in the kitchen, getting your brownies made, making sure that they don't burn, thinking through like turning the lights on and off in your house as you're leaving and you want to save and conserve energy in your home. When we start to combine generative AI with the ability to take actions, there's an added layer of responsibility. How do you give a human in the loop the authority to grant access to taking those actions? But when I think about next year, that's going to be a pretty exciting vector.

‍

John Locke (35:41):

You obviously sit in a very unique place at Google, where anything that you launch is global. How do you think about building a product for a global audience from day one?

‍

Jack Krawczyk (35:52):

Well, we designed for a global audience, but we didn't deliver to a global audience to start. So when we launched Bard in March of 2023, it was available as a wait list only product in the US and UK only in English. The foundational model had capabilities of speaking more languages, but we hadn't yet figured out what's the right way to design a multilingual reward model? What's the right way to present the information around brevity versus breadth of content that you want to cover? And so we actually narrowed the hypothesis that we wanted to learn. So we focused on two countries with one language, and then very quickly we said, well, we know we want to expand globally. Let's not expand globally immediately. Let's learn how to tune the model for one or two different languages. So we picked Japanese and Korean, we launched in Japan and Korea. 

‍

(36:50): The variance in the language was very high. So different alphabet, different cultures. Can we build the right mechanism? So it was as much about process as it was technology. Can we build the right trust and safety mechanisms and teams to be able to evaluate these things? And so as we generate those playbooks, then we go expand to the rest of the world in 40 languages. And the most intimidating thing about all of it that even talking about it right now feels crazy. That was a span of March 21st to July 13th. Right. And the pace at which this technology is capable of moving is very humbling. 

‍

(37:52):

I find the work of Andrew Eng in Stanford to continue to be among the best. He was one of the originators of Google Brain. Professors at Stanford have done a lot of remarkable work. Reading the research that emerges out of universities like MIT, like Stanford, there's amazing work that comes out of Paris that comes out of Switzerland. The research to me is much more interesting than reading about it in the news, and that's how I've immersed myself into this world as well. I highly recommend everybody reads the paper from 2017 that created the transformer, “Attention is All You Need”. It's a remarkable way of understanding the origins of this technology, how it evolves, and there's just countless free resources and they're getting published, and it's part of what makes this moment feel like how mobile felt in 2007, how cloud computing felt in 2011, 2012, open discussion, people trying things and writing about it. It's pretty remarkable.

‍