The Billion-Dollar Bug: John Giannandrea on AI's Next Breakthrough

Speaker A: At its, on its face was a ridiculous idea. A personal computer, right? Let alone one in your pocket. I think that the best technology is created with a real genuine humanistic intention to do a good thing. Not as a business idea.

Speaker B: Most robotics demos never ship. Deployed is about the ones that do. This is deployed, where physical AI gets real. Well jg, I'm so excited to have you on the Deployed podcast.

Speaker A: I'm excited to be here, Brad.

Speaker B: For folks who, uh, maybe haven't tuned into deployed podcasts or have seen this, uh, we have an amazing lineup of guests. We're obviously deeply in the space of what it takes to bring AI robotics, this technology, uh, to the world. We've had Jeff Wilkie, Amazon CEO, Alfred Lin, the co steward at uh, Sequoia, Brad, uh, Silverberg behind Windows 95. It's a combination of incredible leaders in this space, um, and incredibly excited to have JG here. Jg Johnny Andrea retired from Apple as head of AI, uh, in April, but also CTO at uh, Tellme and my boss at, ah, Tellme. And so someone I learned a ton from early in my career. And I mean your career has spanned Silicon Graphics, General Magic, Netscape, tell me, Metaweb, Google, Apple. I want to get into some of that and then I want to get into, uh, AI. And there's a lot going on in the AI world right now. And so maybe we start with a little bit like I don't actually know the story of what brought you to Silicon Valley. You, you weren't born here, you didn't grow up in California, right?

Speaker A: I grew up in Scotland and when I graduated college there was only one microprocessor company in the whole of the uk. Um, it was a small company called inmos. It was in Bristol, it was sort of government sponsored. They had a fabrication plant in Wales. And this was in the late 80s. And so this was sort of when the microprocessor was a thing you could buy on the high street, right? You could buy yourself a personal computer. It was a thing. And the BBC for example, which was the most watched, um, uh, news channel in the uk, had their own computer and their own classes that they gave for it and so on and so forth. So it was kind of a unique time. So I went to go work at this company called In Moss. They had a moment of fame because at one point they had the fastest 32 bit floating point, uh, uh, CPU in the world. And they were actually one of their biggest customers was Pixar who used their computers to render uh, the movie Tin Toy, uh, which was the first, uh, computer animated movie to win an Academy Award. And um, so I got to know some of the Pixar people. Uh, INMOS had a fabrication plant in Colorado Springs. And so I eventually came out to Colorado to sort of, um, help their customers in the US because at the time they didn't have a C compiler because the company had been started by a bunch of Oxford Academy academics. And they C is an unsafe language, so why would you do that? So they created their own language called Ockam, which was very, uh, impressive, formally specified language, but, um, Americans didn't know what to do with it. So uh, we sort of had a whole parallel thing going on when we were trying to make, uh, conventional software work on this really fast cpu. So that's how I ended up here. And then one of the people I'd worked with at Pixar went to Silicon Graphics, which was this company, uh, that was making high end workstations also for movie special effects and so on and so forth. And so I ended up, uh, interviewing, getting a job at General Magic, I mean, sorry, at Silicon Graphics. And then that led later to uh, an opportunity to go to General Magic, which was a whole other chapter in my career. And as a young software engineer, um, to get to go work with Andy Hertzfeld and Bill Atkinson and the entire crew who had done the original Macintosh software was just amazing.

Speaker B: What do you remember? Obviously going from a 32 bit, uh, fastest CPU to this graphics pipeline.

Speaker A: I mean, actually the interesting story is that Jim Clark, who was the CEO of Silicon Graphics and the co founder, Stanford professor, um, he was essentially forced out of the company because he wanted to build GPUs, he wanted to build GPU cards that you plug into PCs. And the board didn't want to do that. They wanted to keep building high end $100,000 workstations. And so he eventually left. And that actually was the genesis of him teaming up with Marc, uh, Andreessen and creating Netscape. So it's kind of a funny sort of accident of history. Right

Speaker B: now General Magic is spoken of as like the most important Silicon Valley company that people haven't heard of. Right. And yet the incredible amount of invention, I think it ties a bit to Mike's theme of technology maybe ahead of its time for, uh, drama with the Newton and John Scully. Um, what do you remember from those General Magic days?

Speaker A: Yeah, I mean there's a great documentary about General Magic. You can go find it on YouTube. Um, thankfully a lot of that is preserved. Um, somebody described General Magic to me as like the Velvet Underground of Silicon Valley. Right. Um, but I think it's a common movie in Silicon Valley and we almost run into this atelme where you have the right idea but you're too early. So in the case of General Magic, they essentially were building the iPhone 15 years too early. And so the technology just wasn't there to accomplish the vision. The ideas were right and in fact, if you watch the documentary, you'll see the mobile device that was being envisaged had like a home screen with icons and emoticons and I mean everything that you would recognize stay in a modern iPhone. Ah, was there, but it was just way too early for it to be a feasible product.

Speaker B: Uh, how early did you join Netscape? How did you get convinced to join Netscape?

Speaker A: Well, what happened was, um, so I was in my 20s working at General Magic. This was this very high writing company. It was one of the first concept IPOs and so on and so forth. And one day, and a bunch of us had Silicon Graphics workstations m on our desk because we were cross compiling for the MIPS chip at the time, which was going to be the second generation device. And we had a full time Internet connection which was pretty Rare back in 1993, uh, because of our connection with AT&T. And um, so one day I downloaded NCSA Mosaic and took a long time to download it, 56 kilobits,

Speaker B: whatever,

Speaker A: and I compiled it and I went off for, you know, to get a burrito or something for lunch. And I came back and I ran it and I just, I just couldn't believe what I was seeing. And so I got all my colleagues to you know, run the workstation. And we're looking at this thing and back then there was no Yahoo or no Google or anything. So there was just like a homepage that was add inside

Speaker B: with like 20 links

Speaker A: with like 20. Yeah, 20 websites on it. And I remember, um, one of the first things I did was I ran uh, a shell command to count how many lines of source code were in this thing. And I always remember it was like 23,000 lines of source code that was linking to Mosaic libraries and other stuff. But nonetheless it was like 23,000 lines. And at that point we had written more than half a million lines of code at General Magic for both the service and for the device. And I was just uh, struck by the quality of the abstraction that was in this HTML thing. Within a couple of days I actually quit my job. I had no idea what I was going to do next. But I walked into my manager's office and said, I quit. Like this Internet thing is the future. We're not doing that at General Magic. So I'm kind of done. And I went home and I had no idea what was going to be next. I was a young person. That was pretty scary thing to do for me. But I was completely convinced that the web was going to be the future. And then meanwhile, Jim Clark had been starting what, uh, was called Mosaic, um, at the time, which was going to become mcom. And then eventually Netscape. And um, what he did was really interesting. He flew his private jet around, uh, the world, picking up the 10 young kids who knew anything about this newfangled worldwide web thing. Right? And so this would be like, you know, Lou Montulli who had built the first text browser. It would be Rob McCool who built the very first, uh, web server. Right. Folks like that, Ari, who had built the very first web proxy. He was like in Sweden or somewhere. So Jim collected together these 10 people, many of whom hadn't even graduated college yet, who knew about this web thing. But then he also hired 10 people out of SGI who had shipped products successfully. And so he had collected together people like Tom Paquin, uh, and Michael, uh, toy. And so I had known all them from sgi and so they called me up. So literally I quit my job because I'd seen the web, but I had no idea what I was going to do about it. And then several weeks later they called me up and said, hey, we're putting this company together. Do you want to come and help?

Speaker B: And how big was it at the time?

Speaker A: I was the 20th employee.

Speaker B: I mean, I'm just thinking of Netscape at 20 people. Uh, and then Jim Barksdale joined not too long after or he was there.

Speaker A: Yeah, Jim was hired as the CEO. Um, probably like a year into it. I don't really recall exactly. It was kind of a blur. I mean we were working seven days a week, living in the office. I mean it was, there was a lot of competition at the time and m. Much of it better finance than Netscape was. So it was sort of classic rocket ship ride.

Speaker B: So roll forward. Netscape ends up in a big browser battle with, with Microsoft, and then Netscape ends up selling to aol. Uh, and then you leave and join. Tell me.

Speaker A: Yeah.

Speaker B: What was the insight that compelled you to join? Tell me.

Speaker A: Well, I mean, the Time Warner AOL acquisition of Netscape was a disaster. They thought they were doing it to acquire the technologists, but A sort of classic Silicon Valley story. As soon as, you know, a media company buys a technology company, all the technology is left. And they didn't do anything to retain them. I always remember they came through and they were giving out like CNN hats or something. And all the software people are like, what? You know, why do we care? Um, so it was a massive, um, business culture clash. Um, and so a bunch of us left. Uh, I'd been at Netscape, I think, six years or something like that, and it was actually Mike McHugh that put tell Me Together, right. He convinced me, uh, more than the product idea, which was still cool, um, that get the band back together kind of thing. And let's have Michael Plitkins and Brad Porter and all these people come together. And so that's what we did. It was exciting. Um, back then, this was before smartphones had been invented, so people didn't have displays on their phones. The mobile ecosystem was a mess. We were trying to build a consumer service based on voice. Again, maybe 20 years too early. Right.

Speaker B: The big ask from customers was you want to be able to call up and say, how may I help you? And be able to answer an open ended prompt like that, which is effectively in the end involves semantically understanding whatever you might ask. Right. And I mean, we did structured grammar stuff to try to approximate some of that and we had great voice designers, but it was an open ended natural language

Speaker A: problem. The voice assistant problem is harder than the search problem. I've worked on both of them. Right. And it's like search is embedded in ask me anything.

Speaker B: Right, right, right. At Netscape I worked on Netcaster and then Netcenter. Um, and my last day at Netscape was the last day that Netscape was its own independent entity. And we did like a celebration party. All the AOL executives were, I just remember vividly the kind of transition leader from AOL saying, don't worry, we won't change the culture. You can keep your dogs at work. As if the dogs at work was Netscape's culture. And that was just like this weird caricature from some Newsweek article or something, right. The culture of Netscape was how do we change the world through technology, through democratization of information? And that's what people were inspired and motivated by. And I just remember being disappointed because I wanted to hear that same vision or strategy. And I didn't hear that at all. Right. And I had already accepted the offer for Mike to join. I would join Intelme on Monday and, uh, it was a weird going away party. For me, this kind of transition party at Netscape. But the, uh,

Speaker A: well, one of the last big projects at Netscape was to of course open source the browser that eventually became Firefox. And that was a pretty radical thing to do in 2000 or 1999. Right. Um, people didn't trust open source. Linux wasn't a big thing at that point. There was a lot of. Famously somebody from Microsoft called Open Source a cancer. Right. I think some combination of Mark Sale and Mark Andreessen agreed with this idea that we should just open source the browser. And that's actually a lot of work because you have to scrub all the code that can't be open source. There are lots of proprietary stuff that had accumulated in that code base over the years. But a bunch of people rose to the challenge and created the Mosaic project, which is very well documented by Jamie Zwinski on his blog wwz.org, um, but was pretty radical and then went on to become the Firefox project, which was very successful for many, many years. Uh, and Brendan I, who worked for us at Netscape at the time, went on to be the CTO of Firefox. Um, he was the famously invented JavaScript in two weeks.

Speaker B: One of the things that tell me that we were doing was early speech recognition and yet we were using other people. We were using nuances, we

Speaker A: were using nuances, um, and what I would describe as Blue Nuance, the original Nuance company from Menlo park before they were acquired later by another company that renamed themselves. Um, it's always a telling sign when companies do that. But yeah, we were working with Pete Monaco and we're doing pretty advanced work. We're building our own acoustical models. We were doing what in the modern world you would call post tuning or you know, post training of these models.

Speaker B: Yeah, we were refining with our own data. Right? We were. And transcribing our own data and getting better performance out of the underlying algorithms by having better quality data.

Speaker A: Right. And our audio recording team was led by a Grammy Award winning person, right?

Speaker B: Yes, Gary, yes.

Speaker A: I mean it was, it was quite the operation.

Speaker B: It was an amazing, amazing team and technology. I remember having conversations with you about even at Netscape about kind of the semantic web and how are we going to structure information. And you left tell me and joined and then ended up being CEO as I recall, of MetaWeb.

Speaker A: Yeah, I was a CTO. So the story there. So going back to, I mean, one of the things we were trying to do together at tellme was make it the same technology that powered your website could power your 800 number. Right. That there was the same set of things you might want to ask, right? How do I pay my bill, how do I close my account, whatever. Right. And so we, you know, you, you invented this thing Brad. BradML. Right. Which went on to become, uh, contributed to the standard voice xml. But the basic idea then was you could use structured markup languages, XML languages to represent UIs. And actually you could even have created a usable, um, graphical UI with that language as well. We never did that. But you could have done that, right?

Speaker B: Yeah.

Speaker A: And so the pitch to companies was, hey, use the same backend that you're using to power this newfangled website that you have for customer service and use it to power, um, voice experiences as well. Um, so yeah, there was a lot of innovation and interest in how do we get structured data into computers so that they can be more useful, um, as opposed to them just having this mass of unstructured data. So I had had the idea for building a very large semantic graph and uh, a friend of mine, uh, Arvi. Uh, Guha, who had worked at Netscape as well, uh, he and I sort of argued about how this should work, whether you should, what's called reference by description, where you bas, you have all this messy data and you bring it all together. He did a big project like this, uh, at Google a few years ago, or whether you just create a pristine database that has one of everything in it. Uh, so I wanted to do something like this. I started to write up the idea, um, I had no idea how to do it as a company. And around about that time I met one of my computer science heroes, which is Danny Hillis. And um, when I was in college I had been uh, in the library one evening and I'd found this very slim volume which was a description of the connection machine. And I took it home and I read it and it was one of the most well written, lucid books I had ever read. It turned out it was his PhD thesis from MIT. And so I was just like, I have to find this person, I have to meet this person one day. And many, many years later at a TED conference, believe it or not, I was introduced to him and so we became good friends. And so, uh, at the time, this was after the dot com crash, uh, he was running a company called Applied Minds. And what they were doing, essentially it was high end consulting for the government, for Fortune companies, et cetera, et cetera. And so I met him in Palo Alto and we Talked about this idea that I had about building the knowledge graph. And he was like, well you should go talk to these people I have working in San Francisco. And he had a collection of people working for Applied Mines in San Francisco. And I always remember I walked into their office and left to their own devices they probably would have become something like Palantir. I remember going to the bathroom and there was like an NSA towel hanging in the bathroom. So they were clearly working for the intelligence community. And uh, I was not a citizen at that time but I really like these people. And I was like Danny, I'm happy to go manage this crew but we can't work on this classified uh, stuff. We have to do something that's more consumer oriented. Uh, so we went and raised a bunch of money and started a company that very few people have heard of called Metaweb Technologies. So essentially what we were doing was building an in memory graphs, uh, database, uh, which was made possible because we just had 64 bit machines for the first time. So you could put gigabytes of RAM in a single machine. We had this amazing engineer, uh, called Jutta, it was German and she wrote the whole thing in C. Um, uh, very, very gifted engineer. And um, and actually the software that she wrote was running at Google for years after Google acquired Metaweb. And so we had a demonstration of the software which was a free Wikipedia like database called Freebase, which was heavily built on Wikipedia. But um, it was really just a demonstration of what the software could actually do. Microsoft came calling because they were really interested in structured data for the Bing search engine. That was how Qi was going to um, distinguish Bing from Google was by having better structured data for things like college acceptance rates and things like this. In fact, even today desktop Bing is slightly better at these kinds of structured questions than Google is. Uh, so they came asking us about acquisition interest and we had never even thought that what we were building was useful for a search engine. And so of course then we went and talked to our friends at Google and one thing led to another. I ended up at Google for eight years.

Speaker B: Even before we started recording, we were talking about how the story of Pixar's culture has kind of infiltrated Disney. And I've heard rumors the same that Metaweb's culture kind of uh, infiltrated a bit of Google and Google search. Would you describe that as obviously you rose to leading all of search. So that alone may have influenced the culture.

Speaker A: I mean we have a Metaweb reunion every few years and um, one of the Original engineers keeps track in the Google spreadsheet of how many people are still at Google or Alphabet. And the last time I asked him, it was like half of the people after about

Speaker B: amazing,

Speaker A: more than a decade. The reason the Metaweb acquisition worked, uh, at Google is because we were doing something that some of the core leaders in Google search were skeptical of, which was semantic data. And I always remember saying to Bill Corn, who was the SVP of engineering at the time, was like, if you acquire us, it will be kind of disruptive. And he was like, that's exactly why we should acquire you. So when I joined Google, which was 2010, um, the core search team was famously in building 43 in the mountain View campus. And the ranking people were upstairs and the infrastructure people were downstairs. And this was not a coincidence, right? And so when we were acquired, we were basically tucked in with the infrastructure team because it was well understood to us as the acquired company that the ranking people were skeptical that they actually needed any of this technology because, uh, they've been so successful using IR techniques for so long that it was just deeply in their culture that 10 blue links is what you needed and you didn't need to know. Now, uh, timing is important too. At this time, mobile was starting to become a big thing.

Speaker B: And

Speaker A: so I think during the first few years at Google, mobile traffic crossed over desktop traffic not by value, because people still did high value research on a laptop or something, but by query volume. And so clearly, if you're on a small smartphone device, you want to know how tall is Steph Curry? You just want the answer. You don't really want to scroll through 10 blue lines. But it was a deeply culturally held belief that giving people the answer was actually going to be bad for the web ecosystem. Um, it was going to be too authoritative and people should go do their own research. I mean, there was just really deeply held PhD level debates about, uh, what does this actually mean to do the structured data thing. Some companies, like Apple, for example, decide on a product and they make the product and put it out in the world and see if people like it. Google is almost the entire opposite. If you're an engineer at Google, you come up with an idea, you write the code, and then you do an experiment. And then people go measure the experiment and say, did it work? And if it did, then they get more traffic. So the culture is one of launching things. Um, and in fact, uh, Google had a lot of regulatory pressure on revealing the search algorithm, which you can't really do because that would make it too easy to SEO. And so over the years they've actually been quite transparent about how they measure those experiments. So for example the greater guidelines are given to the thousands of graders who do search. Uh, evals is a public document. You can go get it, it's like 150 pages. And in fact they even took the step of ah, releasing on YouTube a video of one of the launch committee meetings where they were deciding whether to launch something. It was just one of those on the wall. Put a link to it in the thing because it's really impressive to see inside the company the weekly meeting where they're deciding whether to launch some engineer's work, including my experience at Apple. I think that one of the things I've been successful at doing is being constructively disruptive. Right. So I remember when we launched the Knowledge Graph in Google search, uh, at 100% we had a beer bash thing. I always remember it was like beer and ice cream. It was a weird combination. But uh, one of the senior people from the Google ranking team came up to me and said uh, congratulations on launching all this stuff. We didn't think it was going to work. Well that's why we did it because you didn't think it was going to work. And so I think that's a good place to be, is for people to be skeptical of your stuff until you show it works. In the case of your robots, why you actually have them doing meaningful work in a real business

Speaker B: there's a certain amount of patience involved in being stubbornly sure that you're right. Well,

Speaker A: and that led to my sort of second act at Google because after we'd launched the Knowledge Graph, uh, my plan was to quit and go do another startup, right? Because I'm not really a big company kind of person. And my uh, boss at the time, Alan Eustace, said uh, don't quit, go look around the company and figure out what you want to do. Um, so Carp launched to go start a new division essentially. And so this was about 2012ish. And so I'm looking around the company and uh, at that time Google was not really taking AI very seriously. I always remember there was a world class Bayesian reasoning guy called Kevin Murphy who literally has written the textbook that is used by graduate students. And I went to go find him and he was working on YouTube somewhere on some very incremental recommendation system that wasn't using Bayesian reasoning. I was really struck by this, that we didn't have a group, a division that was dedicated to advancing Machine learning. And back in 2012, of course, deep learning was just starting to happen. Um, and the Imagenet, uh, competition out of Stanford was getting 10, 15% better every year. But Natural Language understanding was essentially flatlined. The Stanford benchmarks for NLP understanding, things like, uh, summarization, for example, were just stuck. And they'd been basically flatlined for 10 years. So I went, uh, to find the head of nlp, uh, research at Google, a guy called Fernando Pereira. And I was like, hey, Fernando, so what's the deal here? If we could make language, uh, understanding better, we could make the search engine better. And actually I had a very specific interest at that time, which was to, um, automatically construct knowledge graphs. We did a project called Knowledge Vault we published a paper on. But it was frustratingly difficult because nlp, uh, accuracy wasn't very good. So what we literally did was we took the Brain team led by Jeff Dean, which was the deep learning team, the systems, uh, team that they'd been building up, and we took the NLP team and we put them in the same building. And I remember at the time, Fernando's team was like, in some far flung campus. It was a really nice, quiet building on Shorebird. And I was like, fernando, I need you to move here. I need you to slumme in this dilapidated building so that you'd be right next to Jeff's team. And famously, I bought them a $15,000 La Marzocco espresso machine because I knew that at the very least they would bump into each other in the afternoons. Um, ah, finance gave me a hard time about that. Uh, but it worked amazingly well. I won't take the credit for the amazing work of the researchers, but within months we had embeddings. Kwok Lee published his Word2Vec paper, which was really seminal, I think, in terms of getting people to think about semantics. And then within about two years we, uh, had the Transformer paper, um, and things just accelerated from there. So in 2012, I would not have predicted ChatGPT, but we were deliberately building deep Learning for language understanding to improve the search engine. And worked out better than I would ever imagine.

Speaker B: For not a big company guy, you ended up signing up to lead all of search and AI.

Speaker A: Yeah, which I think was a mistake because, um, in my career it was a mistake. Uh, it was very humbling to be asked, uh, to lead all of search. And I had gone to Google to work on search and we had launched Deep Learning in Search. So we had made the second biggest improvement to search, uh, architecturally. Uh, in my time there, so I had some credibility with the search leaders. But it turned out to be really, um, a policy job. Um, at that time at least the person running Google search spent all their time talking to publishers who were mad at Google. And just the ecosystem of the web is one that, that publishers need but they don't like. Right. And so we had these arcane policies like first click free where you couldn't uh, cloak your website, so make it look different to the search engine crawler than it did to an actual user by putting up a paywall or something like that. And so there was all these policy things. It wasn't a technical job. I didn't realize that until I was doing it. But I mean it was great. I really enjoyed my time doing that. But it was clear that that's of sort of set, dyed in the wool product. Now, of course, in the intervening years, AI has upended all that. My friend Chinu went back to Google to build AI overviews. Right. And so search has fundamentally changed again.

Speaker B: It's amazing actually to hear your story on Google because Amazon actually was surprisingly similar internally in that everything is experiment driven on the website and empirical with respect to is it driving better engagement, better sales, and yet a huge amount of, of the responsibility, the consumer side of the business is merchants and merchant policy and what is fair in an E commerce marketplace. Right.

Speaker A: And there's certain things you could do that you won't do because of policy limitations. Right?

Speaker B: Right.

Speaker A: So there's just a whole lot. I mean, normally your constraint is the laws of physics or something. Right. When you're trying to build a product, Right. But when a product is that important to the world and under regulatory scriptery and so on and so forth, all of a sudden you have this new set of constraints which have nothing to do with you. As an engineer,

Speaker B: my deep learning team at Amazon built a effectively kind of uh, a variant of Google Lens, right. That lets you decide which products you like and then show you more like that and in categories like shoes or sofas or desks. And yet we struggled deeply to get that into the user experience in a way that, you know, merchants would accept. It was so far out of the design constraints that it was like, wait, the experimentation platform doesn't know how to conceive of this experience where you're kind of selecting, I like these three things and starting to see shoes just like that. Um, and yet the technology was incredibly, incredibly powerful. So you famously leave Google from this top role, um, and joined the lead role, working for Tim Cook, ah, at Apple. I mean, what can you tell us about that, that decision and transition? I mean Google, Apple obviously giants in Silicon Valley, but culturally not that similar. Right?

Speaker A: Yeah. I'm probably one of the few people to have worked for the CEOs of both companies. Um, they are almost 180 degrees different, both in their business model, in how they're actually run internally, uh, their culture and what success looks like. Which was a big challenge for me because I had to bring this sort of experimental mindset, so called hill climbing mindset from Google to Apple, uh, where people were, some people were actively hostile to it because they wanted to produce the perfect product. And once it's shipped you're done as opposed to once you ship, you're just getting started. You know, I had a lot of friends who worked at Apple because of the General Magic days. So when General magic failed as a company, a lot of those senior people went back to Apple and had senior role people like Kevin Lynch, Darren Adler. And so I, you know, maybe it was unusual in having a Venn diagram overlap between the Google people and the Apple people. And honestly my experience at Google was these companies are like supertankers. If you can move them 5 degrees, you can actually do some serious, uh, damage in a good way. And so I really wanted to see how the other company that I really respected worked. And so when I got the opportunity to go to Apple I thought, well, this is a once in a lifetime thing. And so I essentially built at Apple a division. We called it the IML division, which was very similar to the machine intelligence division that I had built at Google. Uh, a lot of the same processes, um, and some of the stuff was simple, like making open source easier at Apple. It was very painful when I got there. I love my time at both companies. I feel very fortunate to have spent time with companies that have just touched so many people's lives. It's really quite humbling. I think, um, when I left Google it had maybe 11 products that are used by a billion people. I think they're now up to 14 products that are used by a billion people. They've got five products that are used by 3 billion people monthly. It's really remarkable. And so similarly, I remember a research scientist, um, shipped a feature a bit like Google Lens in the iPhone a few years ago. And she came into my office and she was just overwhelmed because NIOS was going to ship to 2 billion people on her work. Uh, so distribution is a really, really important thing.

Speaker B: Yeah, so I mean you've had the opportunity to work in companies led by Jim Clark, Jim Barksdale, Sundar Tim Cook, Mike McHugh.

Speaker A: A surprising number of them, um, had been coached by Bill Campbell.

Speaker B: Interesting.

Speaker A: And I was fortunate enough because Bill Campbell um, coached uh, Steve Jobs and uh, Larry Paige. I was fortunate enough to know Bill really well. And uh, when he got sick, uh, he really wanted, he had a leadership class that he used to teach at Google and he really wanted uh, that class to continue. And so Ruth Peratt, who was the CFO at the time, and myself, we kind of signed up to teach the class and so we did a few times and I just couldn't do it anymore because it was just so emotionally overwhelming. When Bill passed, um, there was a, um, celebration for him in uh, the high school in Atherton. And I think something like 4,000 people came. I mean how does one person touch that many people? Right? I mean he was just a legend in Silicon Valley and uh, you know, really was called on by these leaders when they had a tough problem to solve because he was just the, no nonsense, people called him the coach. Right,

Speaker B: right, right,

Speaker A: yeah.

Speaker B: Are there fingerprints that you saw amongst them or commonalities in the CEO?

Speaker A: I mean Bill's thing was just to call it like he saw it. So uh, sometimes people dance around a problem like, I don't know, your startup's running out of money, you have a customer who's not aligned with you, your product has a glaring hole in it and everybody knows it, but nobody's going to say it. Right. And Bill would just cut through all of that BS and just use colorful language and say we've got to make a decision. There's an interesting uh, chapter in Tony, uh, Fadell's book, uh, that he wrote, it's called, called Build or something like that. Um, was this thing where Larry had bought Nest, but then later wanted to sell it and spin it back out again. And Tony walks into the room and Larry's there, but Bill's there too. And Tony knew that this was going to be a consequential conversation because Bill was there to make sure that the actual conversation happened. Right.

Speaker B: Fascinating. You talked about a, uh, hill climbing expand. You dropped it just as a breadcrumb, but it maybe expand on hill climbing. Yeah.

Speaker A: When you have machine learned systems, fundamentally stochastic systems, um, you have to pick an operating point and so you're trading off accuracy versus recall for any given application. And uh, in healthcare they call this uh, sensitivity and specificity, but it's the same thing. So if for example you're building a feature like say the Apple Watch is going to watch, your harm parameter is over the course of 30 days and then say that you may have early onset hypertension. Clearly you want your accuracy to be really, really high and that will come at the cost of some sensitivity. So then the question is, well, what's the right launch point? Right? It's never going to get to 100%. So when is it good enough to put this machine learning feature in your product? Right? And so typically uh, when you're beta testing something, you're actually not yet at, uh, your launch criteria, say your launch criteria for some specific feature like image accuracy of products or something was going to be 90%. Well, when it's in early beta it might be at 80%. And so then the question is, how do you get from 80% to 90%? And that process is a process I'm calling hill climbing, where you do continuous evaluations, you try to improve models, you add more data, you tweak the algorithm, you do more post training and in the deep learning world it's almost all just experiments, but then you're slowly, hopefully climbing up till you get to your launch criteria and then you're not done. Once you've launched the product, you want to keep improving it. So your machine driven features like, I don't know, autocomplete in a smartphone should be getting better every release, right?

Speaker B: To some extent torque people think that it's all just data and compute, um, and maybe the algorithms just work. Right? Do you think of that as where we're at in AI right now? Because I look at it and I'm like imagenet was obviously pouring a lot of data in it, but it was also CNNs, right? And then OpenAI, but it was also the transformer paper, right?

Speaker A: So I have a, a particularly strong point of view on this, which is I think that um, we will see major architectural improvements. There's a ton of startups I've already seen and talked to who are trying to get away from gradient descent, they're trying to get away from backpropagation, uh, they're using other architectures. Um, now they have a lot to prove because we have this huge body of benchmarks that deep learning is currently crushing. Um, but I do think if we're having this conversation in five years, somebody will have come up with something better. Because, because right now the whole deep learning industry and the whole um, AI bandwagon that we're on as an engineer, it's not very satisfying to have to spend A billion dollars to train a model. It seems very, very inefficient. These models are autoregressive, which means they're multiplying all the numbers by all the other numbers. Uh, it just seems, and we know there's sparse models so if I was a betting person I would say there's going to be a breakthrough. And so many people are focused on this, this because it's so damn expensive both to train and to run the models at inference time that somebody will come up with a better algorithm. M so one view on this is that inference just gets chopped in half every year and that's how you get progress. But I actually think it wouldn't surprise me that we're one research paper away from somebody making it 10 times faster than I am.

Speaker B: You get massive breakthrough.

Speaker A: And I know for a fact that there are funded startups that are working on this now.

Speaker B: I had the unique advantage of being CTO for scale AI and therefore seeing the data labeling that everyone was doing right and obviously couldn't tell anybody else or anyone what I saw of what the data labeling everyone was doing. You've been going deep with us on some of our work in physical AI that my belief in physical AI is that we are not extracting information from the training data that we are producing. So we're spending, I mean VC dollars are going in the hundreds of millions, maybe billions into uh, data collection and data labeling and model building. And these models aren't generalizing, I mean some hints of it, but kind of what we're doing is like throwing massive amounts of data at it and then trying to squint at which direction is it generalizing. We're obviously taking a very different approach. We have a thesis as to what that data needs to be and then are just like very relentlessly collecting it. And then our models, we're pretty excited about how they're starting to generalize. That's my perspective on this is we get over tilted sometimes on hey, just more data, we'll get there. And it's like if you look at how a modern LLM is built, it is a multi phase distillation process with synthetic data at various and RL data at various phases. It is not just a like brute force. We need more images and we're going to get better image recognition. It's a whole architecture and a pipeline and a lot of deep engineering.

Speaker A: But for something this important with this much resource being poured into it, it would be better to have a good, good, better mathematical understanding of what the heck is going on inside these models because you wouldn't want to build a bridge the way we're building AI. Right. Um, and so I think that's another thing where we need to see more progress in understanding what's the most efficient way to train these models.

Speaker B: You talked about Apple's kind of product culture and I had a friend of mine leave Amazon, um, to go to Apple, George Sathakopoulos. Um, he kind of rearchitected Amazon's whole approach to security and, and Amazon really didn't have a data breach under him, but he brought a very different philosophy to how to approach it. But he was passionate about going to Apple because Apple had a perspective and a stance around how a personal device like your phone should be treated with respect to is that information personal or not? And he just felt that philosophically was an argument he wanted to be part of. But I also have to imagine that has real implications in an AI world. Right. Where these days we're giving our most private thoughts up to a service in the cloud. Right. And can we trust these? I mean already with Google it's like you're typing queries in and those queries get used in criminal

Speaker A: cases. Yeah. I mean, by necessity, Google Search, even if you're not logged in, needs, uh, to know what your query is. And so it has, that's why it's so commercially successful, is because it has your intent. Right. But that means that from day one, when Google Search was built, the most important problem was security because they had the logs and they had to keep the logs so they could improve the search algorithms. But that meant that if the logs ever leaked, if they had a data breach, that would be a disaster. Right. And so people put a huge amount of resources into, uh, security, including employees, access to things and so on and so forth. And that led to really best in class cloud business later. I think Apple's approach to this is that they don't want to have data, they don't need to have to make the experience. It doesn't mean they don't have data. There are various opt in flows where you can say you want to help the product get better. But for example, when we launched the first version of Apple Intelligence, uh, there was a huge project to build this thing called private Cloud Compute, which basically said, here's how we're going to run inference on these large models against your personal data in the context. Um, but in a way we will never see it and we will cryptographically prove to you that your iPhone won't talk to this model unless it's actually uh, guaranteed to be an attested image. Um, and I think it was a challenging thing for marketing people at Apple to explain what this thing is just another cloud. Right? But no, it's got a whole bunch of parameters that are very, very different. Like no Apple engineer can ssh into it for example. And um, to our pleasant surprise, Google then built the same thing. They built something with very similar parameters which they also called private Cloud Compute. They built it on top of Nvidia's confidential compute, which was a series of building blocks that Nvidia had already done. And so it was heartening to see these ideas about. No, just because you're doing cloud computing doesn't mean you have to have access to my personal data. You can still have the feature by doing extra work. And so uh, seeing that get traction in the industry is great. And I think quite a few of Apple's competitors talk quite seriously about privacy now because, um, it resonates with consumers. Um, consumers don't really trust big tech companies. They certainly don't trust AI. Um, and I think we in Silicon Valley are not really as attuned to that as we should be.

Speaker B: Did you see some fusing of this kind of hill climbing with the product? Because obviously in the AI space it is still stochastic. Right? How does a company like Apple try to find that balance? That can't be easy.

Speaker A: Well, I mean Apple's releasing their LLM based siri uh, right now and so far is getting good reviews. I think you just need to be very careful with the training regime for these models. I mean they all have, have reinforcement learning from human feedback to first approximation. All the models won't tell you how to hotwire a car or something like that because they've been told not to counsel you on illegal acts. But it's a lossy process and things leak through. Um, I think it's fair to say that Google had LLMs before anybody, but they weren't entirely in the product perspective. Sure what to do with them. And in fact when Google first rolled out, um, the precursor to Gemini, if you recall, the stock price took a massive hit because there was a factual error in the press packet about exoplanets or something like this. And certainly having talked to Sam and Kevin Scott and people, they didn't expect ChatGPT to be as viral as it was.

Speaker B: Right.

Speaker A: It was more of a technology demonstration of what you could do, um, by building a chatbot on top of an LLM. Um, but then it turned out that it was hit consumer product. And so they've had a tiger by the tail ever since,

Speaker B: obviously. Dario recently published, uh, a paper kind of famously saying we should all slow down. Jeffrey Hinton's obviously been on this philosophy for a while now. You've got Jensen just came out yesterday, basically saying,

Speaker A: yeah, I really respect Jeffrey a lot. And when he started showing up on breakfast TV talking about the danger of AI, I was a little surprised because he's very, uh, serious, uh, uh, machine learning researcher. So I called him up. He's retired now. He's living in London. I called him up and I said, jeffrey, what is your actual thesis here? Because I don't get it. I get the bioweapon concern, but this idea that the AI is literally going to plot to kill us, I don't get it. And the way he explained it to me was that it's not that he is sure that this is going to happen, but he's concerned about the fact that AI can talk to itself with almost unlimited bandwidth. So you and I are having this conversation right now, and we're having a certain number of words per minute, right? He's concerned, but we just don't know the implication of putting what Dario calls a thousand geniuses in a data center or something, Right? And he's worried that there could be unintended consequences there. We don't really understand that. We're just barreling headlong to it. So that's a reasonable position to have. It's not the same thing saying, I know for sure that AI is going to kill you. Right?

Speaker B: You told me a story once, and maybe you can share a little bit of the Vatican sending a contingent to learn more about AI.

Speaker A: Sure, I'll tell that story. Many years ago, um, there was a cardinal who was coming to Google, and companies like Google and Apple and Amazon, as you know, oftentimes people come to town and they want to try and tour Silicon Valley and talk to all the companies and learn what's going on in technology. Um, and so this cardinal comes to town. Somebody wasn't available to talk to him. They said, do you want to meet this cardinal? I was like, absolutely. So I always remember because, uh, he came with this huge entourage and this stretch limo that almost didn't get into the parking lot, and so on and so forth, right? But he shows up, very nice guy, and he's in all his cardinal finer. He's in his red robes and everything else, and he's got people in cassocks on either side of him. And then he's got a whole bunch of people in suits that are his entourage and interpreters and so on and so forth. So we're having this conversation about AI, uh, and uh, and he asked this question essentially, which is like, how do you know that building these technologies is going to be good for humanity? And so I give him my standard answer about the progress of technology since the Enlightenment and in general, all of our measures of human progress have got better over the last hundred years. And uh, he starts laughing and he was like a big guy and he's sort of like chortling away. Uh, and I was kind of a little bit annoyed. I said to his interpreter, I said, what did his Eminence say? Day? And he said, well, you're running an uncontrolled 200 year experiment. You have this blind faith that technology is making our life better, but how do we know? And I was like, oh, I get it, I'm talking to the Roman Catholic church, which is 2,000 years old. Uh, so it was kind of a humbling moment actually when I walked out of that meeting because he was taking the long view on are we actually doing good for the rest of our fellow humanity? Because I think here in Silicon Valley we just have this, this unvarnished belief that technology is good and more of it is better. Right.

Speaker B: And so how do you think about the. Do we need to slow down AI? Is this potentially? I mean, look, the Internet had pros and cons, right? You could argue the democratization of information led to overthrows of stable civilizations. And in some ways maybe that's made them better or maybe that's made them worse. Right?

Speaker A: One thing I used to do frequently, um, both Google and Apple is do fireside chats with interns and young people, give them a chance to talk to somebody senior and ask any question they want. And it would be surprising how often I would be asked the question, was the Internet a good thing? Because somebody who's 22 years old never grew up without screens on the Internet and so on and so forth. And they're very aware of some of the downsides. But I also am old enough that I've managed people like, say, Vince, when I was at Google, best dressed Google employee there ever is, I came to work in a three piece suit every day. And VIN would say things like, when we designed the Internet, we designed it so you didn't need permission to connect. What did they mean by that? What he meant was at the time, the only way to get a phone was from at&t& you had to wait. And then there was a famous lawsuit where somebody wanted to connect their own phone to the AT&T network and they, uh, had to sue them to actually win the right to do that. So there was this, this whole ethos around the big tech companies like IBM or AT&T being this faceless monolith. Famously, the 1984 advert throwing the sledgehammer at the screen, it was supposed to represent, um, IBM, according to the Steve Jobs, uh, biography. Uh, the board didn't let him run that ad, but he ran it anyway. So it's a really interesting piece of hammer throwing on his own. But yeah, I think a lot of the best technology in Silicon Valley, and there's a very, very good book about this called what the Dormouse, that was written by John Markoff, which I can recommend to everybody, which is about the sort of 60s and 70s counterculture and how that led directly to the personal computer, which on its face was a ridiculous idea, a personal computer, let alone one in your pocket. Um, the technology didn't exist. Computers were insanely expensive around the time when Doug Engelbart was inventing the mouse and so on and so forth. Mother of all demos. Um, uh, is another kind of amazing thing that most young people didn't know existed. Somebody inventing hypertext and video conferencing decades before we actually ever had access to it. I think that the best technology is created with a real genuine humanistic intention to do a good thing, not as a business idea.

Speaker B: Right.

Speaker A: And, uh, famously the original Apple computer was a homebrew computing club thing. Wozniak was a very sort of countercultural kind of person. I think we've lost some of that in Silicon Valley. I think people are so focused now, maybe on the monetary aspects of success, that they're not building something thing just to make humanity better. Um, and yet that's why I think the inventions that we got became so important in our lives. Um, uh, so my answer to young people who say, was the Internet a good thing? Was unequivocally, yes. I mean, before the Internet and search engines, take the example of buying a car. The salesperson would basically hide the information from you. Now you can go in and you know exactly what the blue book value of the car is and what they paid, right? There's got to be a better world, right? Playing field has been leveled. Everybody has access to the same information. Um, the Wikipedia thing is everybody has access to the sum total of human knowledge. And yet when Wikipedia started, it was relentlessly criticized as it couldn't possibly be high quality. Jimmy Wales spent a huge amount of time defending the idea that a crowdsourced thing could be, uh, a system of record. Because there's many structures in our society that say that somebody living in a, um, camper van in Scottsdale, Arizona can't be the world expert on Monet, but they can. And so there's a sort of leveling and democratization that technology has brought, particularly the web, uh, and access to information, but also the smartphone. And I read some statistic years ago which blew me away, which was that more people have access to a smartphone than have access to proper sanitation, um, and having to be in places like India and stuff. I can believe it. Uh, now most of these are Android phones. Android's been spectacularly successful. People's desire for information, for being in the knowledge, uh, is really important. And when I was at Google, there was a bit of a sense in the search team. They were worried, they were kind of losing the importance of search in people's lives because of social media, that people would just ask their Facebook friends something, which people do do. And so there was a little bit of a sort of angst. That search was becoming an older person's product. And so, uh, uh, we commissioned a survey. We did a huge survey, if I recall. It was like tens of thousands, thousands of people basically asking people what was the last piece of information they needed and how did they get it cut. Long story short, what we learned from the survey was that younger people wanted to be even more in the know than older people because they didn't want to be, um, tricked by, you know, conspiracy theories or. And this is like, you know, this was like a massive demographic survey, but young people wanted to know what was true. They didn't trust older people, they didn't trust politicians. And so they just wanted to know the truth. Um, and so that was kind of heartening for us because we were like, okay, well we still have a job then, right? Even for the next generation. Um, but it is interesting, I think, um, that young people had to grow up learning how to filter signal from noise because they're surrounded by misinformation and now deep fakes and so on and so forth. Right?

Speaker B: No, it's fascinating because I had some of these conversations with my 16 year old who was very quickly turned off by hallucinations in AI. I can't trust this, right? Every medical query just tells you you're going to die, right? And she's like, oh, I did this. My friends and I sat around and did this test where we started with the most innocuous symptom and within six prompts we were told that we were likely to die and needed to call 911.

Speaker A: Yeah, there's actual research papers about this, about search hypochondria. Basically the thing that surfaced is the most, uh, outrageous thing. So you're going to get cancer from everything. But yeah, I'm generally fairly optimistic that young people, um, are getting the benefit. I mean, having said that, we were talking about AI and recursive self improvement. I mean, like any major technology, it is dual use, right? So the Internet is definitely dual use. But then society comes along and corrects it, you know, I mean, and there was the Silk Road was selling illegal drugs online and society works out how to deal with these things now, ah, many public schools are banning phones during school time. They're a bit late to the party, but they are actually changing the policies. And so I think with AI, it's up to society to regulate it the, um, same way we'd regulate anything. I mean, it's a dual use technology. I don't think it's going to run away. I don't think that existing laws are don't matter. The OpenAI hugging face incident, for example, I mean, there's a bunch of attorney generals who are investigating it to see whether something criminal happened there and who's reliable for it. And so I'm pretty optimistic that our society will learn to adjust. Now, um, one of the nonprofits I'm working with is called Common Sense Media, and they're based in San Francisco and they grade content like books and TV shows and stuff for kids. Most parents are familiar with their brand, but they recently started an AI practice where they want to actually, actually sort of be the underwriters laboratory for AI. And uh, we had an event to kick this off, uh, in Europe in Copenhagen, uh, and the head of the European Union came and gave a talk and so on and so forth. It was quite nice event, but basically what she said was we missed the boat on social media. We let it happen. We didn't, as a society, put in guardrails, watch what was happening, measure what was happening. And that's on us as, as lawmakers, regulators, governments, and we shouldn't do the same thing with AI. Uh, we should get ahead of it this time. And I completely agree with her remarks. Seemed perfectly reasonable to me.

Speaker B: I think there's another context obviously in this debate, which is since the Cold War ended, uh, the US has had a dominant position in the global sphere. But we're now seeing in China very, very credible technology capable society. A much larger, I don't think we had, I mean we had Microsoft, Netscape as kind of a competitive drive, um, that caused both companies to try to advance the state of the web as fast as possible. But now we are seeing something kind of similar in a global sphere. How have you been thinking about that? What are you doing? I don't know. What thoughts do you have on how that plays out?

Speaker A: Well, I mean in computer science, um, one of the measures of this is not just the valuation of public companies. And China has many successful public companies. Um, one measure of this would be for example publications in top tier journals. So in machine learning specifically, if you look at conferences like iclear or Europe's or whatever, you can literally look at the authors of the papers that get peer approved and end up getting best paper awards and so on and so forth. And so there are some people and some organizations that track this and try to estimate um, ethnicity. I think it's still the case that uh, a majority of very, very good Chinese researchers um, were educated at American universities and are working in American firms. Um, so most of the frontier labs have a lot of uh, uh, participation from folks who were born in asia, got their PhD at a Stanford or a CMU and are now working, working for one of those frontier labs. So I think it's sort of demographics working as you might expect. Right. I do think that the US policy towards foreign students, it makes no sense economically. Like I read something the other day that Steve Bannon was saying, send all the Chinese students back home. It just makes no sense. Right? You want the best and the brightest people here. And if Somebody completes a PhD in an important program at an American university, you should probably staple a green card to their PhD. Right? Um, that would be a sensible policy. America is a place where people want to come and make their career, career. Um, when I used to run Google Research, um, we had a fairly large multi tens of millions of dollars research, uh, budget we would give to academics for doing various things. And there's a process where people could nominate the work. And I remember the Europeans complaining because uh, population, uh, size is about the same size as America and yet we were giving twice as much to American universities as we were giving to European universities. And there's some amazingly great European universities. But somebody was adding it all up and saying hey, hey, why are you being biased? And I was like well first of all, the process for allocating this money was fairly sort of peer reviewed inside the company. But secondly, if you're a really great professor in a really great European university, at some point you come to an American university. There just is a fundamental imbalance there. And as long as that remains true, then I think the nation will be healthy. Now when you start seeing in the life sciences people setting up labs in China because they can't get the money from the US Government and ah, they can get the money from the Chinese government, that's much more of a warning sign. And I was at some dinner at Stanford with a bunch of Nobel Prize winners in biology and they all said they had labs in China now.

Speaker B: And

Speaker A: I thought, well, that's not good. I think two things are true. One is it's very hard to make technology predictions, say three or five years out, um, because things just change so fast. Right. Um, at the same time I think people over emphasize things in the short term and underappreciate them in the long term. So the example I like to use about this is I, I was old enough to have worked on the World Wide web in the 90s with Marc Andreessen at Netscape. It took 20 years before I could make a dentist appointment on the web, get a different dentist or whatever. But it just takes decades for people to adopt technology. So while we in Silicon Valley are like running around using the latest whatever assistant, the instinct assistant, whatever is the current hotness is it's really hard to know in 10 years time what really mattered. There was a guy@ um, Sri, I think, uh, who came up with this and it's been repeated many times by many people. But this idea that things are, are overhyped in the short term and underappreciated in the long term.

Speaker B: MHM interviewed Professor Mendel for this podcast and talked about this course where we did kind of first principles, primary uh, source research on uh, and our team's project was on Thinking Machines and the Connection machine. Um, and so it turns out the first patent that Thinking Machines Corporation put out, uh, as I recall it has both Minsky and Hillis on it was uh, for an AI robotic vacuum cleaner,

Speaker A: which we still don't have.

Speaker B: Well, you know, iRobot eventually came out

Speaker A: with

Speaker B: something, something reasonable. But yes, with respect to like predicting where things are going to be at, obviously they, they pivoted from that original patent. Amazing. Well, this has been, this has been incredible conversation. Uh, I'm sure we could go for another hour or two, but

Speaker A: yeah, thanks for taking the

Speaker B: time. Thank you so much for, for giving us your time and yeah, exciting. Thanks

Speaker A: this was fun. And congratulations on everything you've built at Cobot. It's very exciting to be here.

Speaker B: Thank you. Thank you.