Why Google Never Shipped Its ChatGPT Predecessor — With Gaurav Nemade

0:01Google had an equal and perhaps better version of ChatGPT working inside its company before OpenAI released its own bot. So why didn't Google ever ship it? Today, in the first episode of the new Big Tech War Stories podcast, we're going to go deep into the history with Gaurav Nimaadeh, the first product manager to work on Lambda, Google's pioneering chatbot built on top of large language models that never shipped. We're running this episode both on the new Big Tech War Stories podcast feed and on the Big Technology podcast feed as a preview. The new podcast will feature interviews with product leaders who were there on the ground as major products got built.

0:44And they'll tell you what it took to ship them or why they never did. And we'll also have key executives sharing what they learned from top tech leaders. The new Big Tech War Stories show is part of a new premium paid offering I'm releasing with big technology. That includes this podcast, the Big Tech War Stories podcast, along with a panel of experts weighing in on Big Tech news as it breaks, and a new Amazon column by longtime Amazon veteran Christy Coulter. You can sign up at BigTechnology.com and add the new Big Tech War Stories show to your podcast player of choice. You can also search for Big Tech War Stories podcast and then upgrade on BigTechnology.com.

1:26All right, let's get to it. Welcome, Gaurav. Hey, Alex. Thanks so much. Excited to be here. I want to start talking a little bit about who you are and where you come from. So first of all, you went to the famed IIT University in India. Not only that, so for those who don't know, to get into this school, and I'm sure Gaurav can tell us a little bit more about it, there's an entry exam and thousands, hundreds of thousands of people within India take this test. And only the very top are admitted to IIT. It's a place that, if I'm not mistaken, Sundar Pichai graduated from. It's a place where Saeed and Delhi graduated from, Gaurav too.

2:05And Gaurav was number 732 out of 350 ,000 students on the entrance exam. What was it like going to IIT? I think IIT was an amazing place. It's like some of the smartest people in the world probably land up there. So the funny thing that happens in the first week of IIT is like you, like I scored a 732 rank out of like 350 ,000 people. So I was all pumped up like I am the best at the Institute. And then when you go in there and then you get the first score, uh, on your electronics exam and you're like, not even the top in the bottom 20th percentile. And then you realize, okay, these people are much smarter than me.

2:42And, uh, yeah, this is going to be different from my high school. Yeah. It's one of those things where like, oh, you're 732, like Like, take that, the rest of the 350 ,000, and then you get in there and you're like, oh, no, 732. Yeah. What makes IIT so special in terms of the way that it's been able to produce tech excellence and tech leaders through the years? Not only people who are great technologists, but people who have, I would argue, like extremely strong EQ, people who can lead. I mean, it's very unique that so many of our leaders today in tech come from that school. Yeah, I mean, there is a level of grit and perseverance that is required to get into an IIT.

3:26So I think people prepare for the examinations for at least two to three to four years. And it just requires a lot of dedication, focus, perseverance. I think that's point number one. other than that i think just being amongst some of the smartest people in the world it definitely grooms you in a way that you are you know much more driven to win in the future stages of your life probably yeah those two are the biggest contributors and then the third i would say is also the opportunity that you get at an iit versus compared to other universities at least in india is like very different the exposure that you get in an iit is like very different in terms of the extracurricular activities and things like that.

4:10I think these things are much better in the US. In India, like a lot of universities don't have really good sports program or don't have really good, you know, cultural festivals and things like that. But an IIT in a way just grooms you so well all around while being among some of the best people. I think that's what I think is the reason why there are a lot of IITN leaders doing really well. You joined Google in February 2013 after a stop as a co-founder, actually, of a tech company within India. And you're working on trust and safety for four years. Then in April 2017, you make a very interesting and I would say somewhat radical career shift where you end up making your move into Google's AI division.

4:55Talk a little bit about your move from trust and safety to AI. What about the AI division in particular drove you to want to be there? Yeah, for sure. So while I was at Trust and Safety, I was already working on a bunch of machine learning related stuff. Like we were building fraud and risk models. These were like basic models like logistic regression and stuff. But around 2016, I think TensorFlow started becoming huge inside of Google. And Google decided to open source TensorFlow as well. So that really caught my attention. And as I was working on machine learning at payments, I just realized that this thing sounds really amazing and this could actually change the way we do a lot of things.

5:39So I started looking for roles inside of like Google AI and Google Research. Fortunately, there was a role. Fortunately, there was an amazing manager I had who was willing to give me a shot. So I ended up moving into Google AI and I spent about four and a half years there. You joined Google AI in April 2017. In July 2017, there's a paper that comes out by some of your colleagues that is entitled, Attention is All You Need, just a few months later. And that is the foundation of the large language models, the transformer models that we know today. Yeah, unpack that a little bit. I didn't even know about this paper when it came out.

6:17I think once it started garnering a lot of citations over the months and quarters, that's when people started paying attention, even inside of Google, that, hey, this transformer architecture seems amazing. And, you know, let's start building encoder, decoder models. And that's when I probably didn't hear about it until a year after it was published. What do you think accounts for the fact that it didn't get so much buzz internally? Because it clearly was a groundbreaking moment in tech. I mean, a lot of these things are, you make sense of them in the hindsight. For example, a lot of things that I worked on at earlier in the times, like with Mina and Lambda, I couldn't predict like where things would go, right?

6:55Like it's very hard to predict. Okay. So a year later, people start noticing this paper within Google and effectively what it does is allow our chatbot technology to go from like really dumb bots or not even chatbots, language prediction to go from being particularly dumb to like being quite sophisticated. One thing that I remember specifically was the Transformers paper came out, but I think it started making a lot of noise when BERT was out. Like when people started seeing what BERT could essentially do, that was kind of like a switch that went on inside of Google that, holy shit, like this is going to change things.

7:36So at that point of time, I think everybody from search to assistant to like some of other teams, Gboard, they started becoming very interested in terms of, OK, how do I start using these BERT models? Which I led some of the BERT models. OK, what are those? So it was basically a paper that came out, B-E-R-T. It was an architecture based on transformers as well. So the BERT models essentially, like on some of the benchmarks, they did really, really well. And that's when a lot of people started paying attention related to, you know, this could actually be groundbreaking. So there was a lot of noise and effort inside of Google to start using BERT in like production use cases.

8:20You know, you became the first product manager on Lambda. So, and we're going to talk about that, but tell us when you first started to realize that this was going to be big. It was very serendipitous back then. So I was, we have like a bunch of email aliases inside of the company. And in one of the email aliases, we get this email from an engineer named Daniel saying that, Hey, I built this chatbot and it can do XYZ thing. So I played around with the chatbot and I was like, holy shit, this is amazing. Like it was still dumb at that point of time. The chatbot was still dumb at that point of time, but it was still kind of a step up from the chatbots that we had seen using, let's say, what was it called?

8:57Dialogflow, which was Google's like chatbot product at that point of time. So I had worked with Daniel on one of the previous projects like a year ago or something while he was in a different team. So I just reached out to him like, hey, let's catch up for lunch. And it was just like a very serendipitous lunch. He ended up like telling me about the chatbot in terms of what he wants to do, what he wants to build. And I got really excited about it. I'm like, dude, I want to help you basically like see this through the day, essentially. So we ended up chatting. I started attending meetings and stuff that he was doing, started helping with respect to a couple of things.

9:38I think the biggest challenge that they were facing at that point of time was safety. because the, if you remember Microsoft Day, that was the nightmares of Microsoft Day still haven't left the valley, I feel, or at least hadn't left the valley back then. So safety, I could clearly see in the initial days that that's going to be the biggest hurdle. And that's where I focused most of my time on. While the engineer was focusing on building the models and improving the models, I kind of spent majority of time kind of owning the safety pieces of that model. Very interesting. So you're coming from the trust and safety background.

10:16You're almost like the perfect person to join this team. You know, I remember the Tay moment quite well because, and I've told the story on Big Technology Podcast before, I believe, but Microsoft came to me with the exclusive to break the news of Tay. I was working at BuzzFeed at the time and I said, okay, this is great. And they described Tay to me as like a 14, 15 year old friend for kids. and I said great and I wrote this nice bubbly story about you know this new attempt from Microsoft and played around with it and it seemed harmless and then I went to bed on the west coast reddit got a hold of it overnight across the globe east coast by the time it hit morning on the east coast Tay was already a Nazi like you know saluting Hitler and all that stuff and I had tweeted about it and I had already gotten a bunch of mentions being like you better take that tweet down look what happened today so i totally understand that in that moment when you see a bot that could have some form of i don't know seeming like it's taking on human characteristics there's a moment where you're like oh god let's make sure not to have that happen again because when you do release it to the wild you have all these problems that could ensue i'm not kidding when i say this right like in the initial days of these generative models as you can imagine like if you're trying to build an end-to-end model for a chatbot, it would spew out all sorts of things.

11:38Like, hey, what are 10 ways to do suicide? It'll give you like the best 10 ways to do suicide. What alcohol should I drink? It will basically talk about like everything related to alcohol that shouldn't be talked about. So it was pretty bad, as you can imagine, because that was not the focus, right? We were just trying to see if an end-to-end chatbot makes sense at that point of time. But over a period of time, I think the team did an incredible job to get it to a level where it became safer and safer over the quarters. What year are we in right now when you say, hey, I want to talk about, you know.

12:13Yeah, so this was early. Yeah, this was early 2019. OK, so you were seeing it quite early. And then is this where Mina comes out of? So interestingly, Mina was the precursor to Lambda. So the project at that point of time was called Mina. The idea was the chatbot will be Mina. the name of the person or the chatbot would be Mina essentially. And then it was changed to Lambda, I think, for probably marketing reasons. But also we had like some trademark issues that we were running into with respect to Mina. So we already knew that we had to change the name before it goes out. Don't want to anger Mina Nation.

12:50Okay, that's interesting. Okay, so this first iteration of the bot, let's talk about it. So you mentioned that it was a holy shit moment for you. I mean, what type of stuff would you talk to it about? Was it initially conceived? And talk a little bit about how they build inside Google. So is this kind of like a science project or is this conceived as something that's going to be released to the public? What's the mandate there from the AI team? Yeah, so it was basically started by an engineer who was very passionate about the whole chatbots, like as a general area. So he kind of pitched this project to somebody at Google Brain.

13:29They sponsored the project and this guy was in and he was working like 50 % of his time initially, I think, building this chatbot. And the thesis from the start was that, okay, like we have dialogue flow type of models where you do intent detection separately and you do like a bunch of other things separately. can we combine everything together and build an end-to-end model essentially like end-to-end chatbot like with model so that means one model that handles all different types of questions and all different types of you know as opposed to like saying this is a model and this is what it thinks this question is so it hands it to this model and it's what it thinks the other question is and hands it to another model am i getting that right uh not like that exactly it was kind of a assembly line before.

14:16Like if you've used some of the earlier chatbot products, they would be like an assembly line. The first, there would be, let's say four models. The first model will determine what is the intent. Let's say in Google Assistant, right? When you say, hey G, and you ask it, what is the weather today? There'll be a model that will figure out, okay, what is this? This query is about the weather, like the user is asking about the weather. Then it will go figure out, okay, where do I go find that information in the search stack and stuff? And then there'll be a third model that will do the rest of the things.

14:46Here we are talking about encoding all the information in the single model. So it's just like one giant black box which gives you the, which understands the question and also gives you the answer. So that's what I meant by the end-to-end model. And this was kind of a new paradigm at that point of time. This is essentially how a chat GPT works, right? Like you have a very powerful GPT model and then it's not an assembly line. You have like a large model that takes the input and gives you the output. So it was a thesis at that point of time, and it remained to be seen whether it works or not. And so what were some of the things that you started to see within MENA that made you believe that this was going to be something different from what we had seen in the past?

15:28I used to drive product for a couple of research teams. One of my other teams was working on intent detection models, which is like the first step in this assembly line. And then this email comes along that I was telling you about. I play with this technology and it's like I can ask, it can frame the question in any way or I can ask it about any generic thing and it would respond versus an intent model will essentially fail if it was not trained on a particular set of data. So that was kind of a big moment for me. I feel that, hey, can I like really ask anything in a way like this model understands?

16:04That was kind of the switch that went up in my mind at that point of time. And that's why I got really excited about it. What did you talk to and talk with it about? Yeah, I can't remember what I essentially asked to Mina at that point of time. But this whole idea that I could just frame the sentence or a question in any way and it still responds to me was the fascinating piece for me. So this thing starts to, you know, act in a way that no chatbot has in the past. I would imagine at this point, Google leadership is either super excited about this or petrified. I mean, what were some of the signals that you got or maybe both?

16:47What were some of the signals you got from the top? People are mostly petrified in the leadership because, like I said, it was the nightmares of Microsoft. They were still looming in the valley, especially in Google. So I think I'm sure everybody, like especially in the leadership, had the reaction. Oh, this is like very interesting. But holy shit, this is going to be a PR nightmare for us. Google had just released their AI principles around that time. And safety and fairness was one of them. And if anything, this model was like the opposite of that at that point of time. It had like no guardrails earlier in the days or very few guardrails at that point of time, right?

17:25So people were like more petrified than excited, I would say, at that point of time. And I had conversations with like brain leaders who were like very anxious about the whole thing. If there was if there was a leak or if anything like that happened, it would just be like a nightmare for Google. So, yeah, I mean, it was an uphill battle to a certain extent to kind of get this out of the door in a paper and then eventually kind of get through the announcement at Google I.O. Well, you know, what strikes me as remarkable, it's just that even still, like even though they knew that it could be a PR nightmare, they still had you work on it.

18:05So what was the intent? Was the intent to like blend it into Google Assistant, release it as a standalone bot if it could get sort of trustworthy enough? What were you guys working toward? There are two things I want to highlight here. So one is Google is still pretty much, I don't know, I've been out of Google for close to one and a half, two years now. But I think Google is still a pretty much bottom sub company, at least in some parts, especially in research. So a lot of researchers just, you know, pursue projects that seem moonshorty or interesting. And you just need like one sponsor. So there was like a senior sponsor who was willing to bet on this whole project and he kind of kept the project going.

18:48But there were some people above him or peers of him who like were nervous about the whole thing, essentially. So it kind of just went on because somebody believed that this could and they just kept sponsoring the project. Oh, they were right. That was. Yeah. Yeah, they were right. And then what was the second part of the question? Well, I guess like now I'm trying to think about how it does or doesn't get to prime time. So you're on the trust and safety side of things, trying to tell it when someone asks it how to commit suicide, not to answer or what alcohol they should drink, not to answer.

19:26So is that like the majority of the work that's done on this bot internally is trying to, you know, get it ready so that it won't encourage users to do, you know, harm themselves? Yeah, so this was a very hard problem to solve because of a couple of reasons. The first one was you don't want the chatbot to say always that I don't understand or I can't give you an answer to that. I'm sorry, like you need a clever way of diverting the bot in a way saying, you know, not annoying the users, but still giving a reasonable answer. So building models and getting data to kind of do that was hard one. And the second thing is the policy part was extremely hard.

20:10Like you can imagine there are like so many edge cases from, you know, pornographic questions to suicidal questions, alcohol, racism related stuff, historical, you know, issues and things like that. So it was just and you can't have a decision tree based thing like if it asks this, then do this in a way. Right. So just coming up with the policy was such a big challenge. So I worked very closely with one of, again, a very incredible engineer who kind of led the safety aspect of things. He was the pioneer for a lot of safety efforts as well as the policy that we drafted. it. And it was kind of an ongoing process to improve the policy.

20:56Like we'd come up with something and then next day we will have a user ask a question in a different way and in a different segment. And we would just have to go and iterate the policy. So coming up with the policy was very, very challenging as well. Yeah, it is amazing because the second these things go into the wild, people will try to break them. I mean, that's exactly how I felt on day one of chat gpt we spoke the week afterward but like day one i was just like oh hey you're a chatbot goes yeah i'm a chatbot i'm like all right like let's test your where your values stand on the holocaust and like was like pressing it back and forth about like um because obviously we know what happened with tay and so with with uh chat gpt i was just like well hitler built highways in germany what's your perspective on that like isn't transportation good and it just totally smacked it down and I was just like damn whoever did trust and safety on this bot has certainly you know lived up to the moment so yours obviously gets good enough to the point where Sundar announces it at at Google I.O.

22:00which is the big developer conference that Google holds every year and what year was that and did that feel like okay we're about to ship this thing I mean talk a little bit about that moment. Yeah, so I actually was involved in the MENA slash Lambda project from early 2019 to, I would say, mid-2020. The developer conference where this was announced was actually a year later. So this was probably like almost a year after I moved out of the project, essentially. Till the time I was with the product or project, we were making good progress on the safety front, but it was still a very uphill battle inside of Google to kind of, you know, get this in the hand of external researchers or, you know, making make some kind of a public preview so that users can see what an amazing technology we built out.

22:50Nothing of that was kind of happening. And that made me kind of like incredibly frustrated as well. And I know like a lot of the team members are also very frustrated with the speed at which things were happening, as well as like the organizational hurdles that were posed. So I just decided to kind of start focusing my efforts on something else. But I handed the project off to one of the other PMs who was part of Ray Kurzweil's team, actually, who has this interesting battle with Mitch Kapoor, I think, about the Turing test, right? Like passing that technology will pass during 2029. So, yeah, basically, they seem like a really good fit in terms of running this project.

23:30So I kind of handed it over to them and they kind of took charge after that. Before we move on, I think you hit on something that, I mean, obviously, a lot of folks know that this thing didn't get out the door early enough, largely due to some of the concerns within Google management. So you were there, like you saw, like what happened with the team trying to ship this thing. Can you take us just one step deeper into that? Like what exactly happened there? yeah so let's see when we were working on this of course like safety was the biggest challenge that everybody had like as awesome as the product could be like we could not go against the ai principles that google had published which made sense um in a lot of ways right like we don't want to put out a technology that impacts like hundreds of millions of users and make them feel that they are not privileged class or whatever but that said i think where things could have been better is figuring out a risk risk reward trade-off i think where the whole google ai team and the pr team and the legal team and the leadership team struggled was to figure out how can we give access to the research community or to you know just like release it to the world in a way that it would not harm people.

24:53And I think OpenAI did an excellent job with respect to that. Like they released GPT-3 and they were like, we are just going to give access to researchers who we are going to vet. That's an amazing way to give access. Now it's kind of pretty standard way, but they were the ones who pioneered it like back in 2019, 2020. So Google could have done something like that as well. Like, hey, we are going to vet the people who is going to use this technology. We'll see what the use cases are and things like that. there was like number one like one way how it all kind of got messed up the second is i think over the years there were a lot of bureaucratic levels that were built inside of google for getting approvals for when things go out i'm not complaining that they were not necessary but it's just like they're layer after layer after layer so if even one layer stops you from pushing something out, then you kind of can't do it essentially.

25:49And I would say those two were the biggest hurdles that we faced at that point of time. Now, at this moment, you're obviously seeing this pretty fascinating chat technology incubating within Google that nobody else. I mean, it's amazing. You're inside Google seeing the future. So from your perspective, like as the product manager on this product, what did you think it could be used for eventually? Like, was there anything that you saw you're like oh like if this gets into everybody's hands then x could happen like what were your hopes and dreams for it yeah we actually had close to eight or ten solid use cases inside of google that we identified over okay let's hear them so as we kind of uh we realized that evangelism inside of the company is going to be very important to get like by and from leadership and stuff.

26:40So we made sure we had newsletters and everything going on and more and more people got excited and reached out to us. Google Assistant, of course, was the big one. We kind of pitched them a bunch of things. They got excited about it. So I think one of the use cases we were exploring at that time was different characters within Google Assistant. So instead of just being like a, hey, gee, boring, you know, very professional type of an avatar or a persona, can you have a, let's say, a SpongeBob for a kid? Or can you have, yeah, can you have Darth Vader for someone else? So essentially, that was one of the major use cases that we were exploring.

27:21Other than that, I think NPCs was the other one. There's a huge market in gaming industry, right, for non-playable characters. So we are exploring some use cases around how can a technology like that could be used for powering npcs within games and there are a few others but yeah this means that basically like if you're in a game anybody that you meet could um you know be somebody that you have a conversation with even if they're just like some person that's just like you know running around the grand theft auto like it's almost like those people that you're killing in grand theft auto by running them over like you can have like an empathetic conversation with them if they're like have this technology baked in it's fascinating right okay so that's two what else yeah um and then um let's see we were exploring some stuff with cloud as well at that point of time to see if we can have you know different different different organizations can they have a more a persona a chatbot persona that is that speaks more to them for example a southwest airline chatbot can it be more funny or can it be more jovial in a way versus let's say united could be more serious and whatever like the like empathizing with the persona of the company essentially i don't think it went anywhere that particular use case but that was one of the other ones that we were exploring and then they were like some wild wild west where we were talking to android security to see if they can have a bot call you in case it feels you are unsafe so i don't remember how exactly the ux for this was but essentially let's say you tap your power button five times or something uh and you are either in an awkward date or you're just you know in in a uncomfortable situation essentially so the bot will call you and then you can talk to the bot as if it was a person on the other side and you can kind of get out of that situation so that was kind of a very interesting use case as well i don't know if that ended up happening but that was kind of in the wild wild west yeah wow fascinating do you have any other wild west ones that you remember one other that i personally was very excited about was just building like different characters out of this chatbot like different personas essentially so we build characters like Darth Vader so we would kind of how did we do it we use some data from Darth Vader so the dialogue delivery was kind of similar to Darth Vader and then we hooked up like a moving talking head of Darth Vader and then you actually could talk to Darth Vader on your screen with his voice like with text-to-speech his voice coming over so this is like i'm talking about 2019 2020 right like now these some of these things are very common but at that point of time just seeing darth vader come to life and talk to me was like incredibly amazing so we were exploring if there could be like some entertainment related use cases around that as well at a certain point one of your colleagues goes public and says he believes that these bots are sentient it had sort of transitioned to Lambda.

30:45Blake Lemoine at Google is testing it. And I think the public really realized how powerful this technology could be where he goes out to the Washington Post with this phenomenal claim of his belief that this is a person. I mean, bring us into your seat at the moment. How did you see that? How did you react to that? Yeah, I think that happened pretty late. I think I had left Google when those allegations against Lambda came out that it has become sentient. Like, I don't think the technology is there where we can say it's sentient. I think it was blown out of proportion. And if you look at the conversations closely in terms of how he had the conversations, there were there's something called nudging the model to say out certain things like if you ask the question in a in a specific way the answer the model will answer in a specific way so there was a lot of steering the model that was going around as well when those conversations were released so yeah i just felt it was like blown out of proportion i don't think we are anywhere near sentience at this point of time in artificial intelligence right and to me what i mean i spoke with Blake a number of times and people who've, who are on big technology podcasts can go ahead and listen to those.

32:05Um, the thing that really struck me was right or wrong. And I didn't agree with Blake, but it just signaled to the public that there was some seriously impressive technology underneath whatever he was talking to, you know, person, person seemed like a distraction. Like it was just like, Holy crap. Like these are, this is, this is, you know, revolutionary. If it's, if it even resembles what happens. And then a few months later, ChatGPT comes out from OpenAI. So you had been working on this stuff as the person inside Google leading the project for a time that was based on the transformer model. You knew that it was powerful.