Derrick Connell: 0:04

I still to this day think that is one of the single best I was ever asked. What did you learn? Unexpected things will happen unless you plan for it. So mistakes will happen, not might happen. So be careful.

Hannah Clayton-Langton: 0:19

Hello world and welcome to the Tech Overflow podcast. I'm Hannah Clayton Langton, and we are the podcast that explains technology to curious minds. I am without my co-host Hugh Williams, but I have got an special guest for us today, Derek Connell. Derek started his career at the London Stock Exchange and then spent 28 years in Microsoft in a number of senior roles, most recently as the corporate vice president for Search and AI. Derek now advises technology and product leaders, and to know he is mad about mountain sports and photography. Derek's first book is coming out this week, 21 Summers. I had so much fun interviewing Derek. We had an awesome conversation about innovation, and when it goes wrong, he's got some awesome insights. So let's get right into it. Derek Connell, welcome to Tech Overflow.

Derrick Connell: 1:07

Thank you so much. I'm an active listener. person.

Hannah Clayton-Langton: 1:11

That's what we like to hear. Okay, so let's start at the very beginning. You spent nearly 30 years at Microsoft, spanning a period of change. And on Tech Overflow, we've talked before about the different tech eras, starting from personal computing to the internet, to social media, and now AI. How did that feel for you on the inside at Microsoft?

Derrick Connell: 1:31

Well, I started in 1992, so I did live through most of the of big change. So the internet came around mid to late 90s, mad period And then you know Google came late 1999, so search started to come in ads. But I guess the next big one was cloud. You know, companies like Amazon were born in the cloud. So previous companies were offline companies, real estate, cetera. But they were born in the cloud. And then the phone came, 2008, and around the same time as media. And then later in the 2010s, I guess that's when the early of what is now ChatGPT LLMs, which you've talked about in previous episodes. But yeah, there's these waves of innovation. It's great, crazy, because you've got to hold on, ride but it was good.

Hannah Clayton-Langton: 2:17

And Microsoft as a business was sort of born out of Windows and Office and then moved into all sorts of things. So when did Microsoft decide to move out of Windows and and into what I presume was Bing and beyond?

Derrick Connell: 2:28

Well, that would have I mean it started with the wave of So, you know, before that you had the PC, then the graphical user interface came along with Windows, and then one of the first applications, big applications on Windows was Office, Word, Excel. They were individual products, but then they packaged Office. But then the internet came along and you know, you had AOL back in the day. Yeah. So you're going down memory lane.

Hannah Clayton-Langton: 2:51

Yeah, my first email address was AOL.

Derrick Connell: 2:53

Yeah, I see.

Hannah Clayton-Langton: 2:54

Yeah.

Derrick Connell: 2:54

A lot of people still have Yahoo addresses or Hotmail

Hannah Clayton-Langton: 2:58

I must still have mine. I don't know what happens if you just leave them.

Derrick Connell: 3:00

Uh it's probably gone. Okay. Retired. Um but that that was the you know, Internet Explorer, that the first, and Microsoft decided to invest in sort of to do with what it meant to be on the internet, cloud search, different way of working, etc. So that's how we got into search, which is where I worked from 2003 onwards.

Hannah Clayton-Langton: 3:22

Speaking of search, we've just done an episode on search, and Google now hold over 90% market share. How was it back then? Was it similar? And if so, sort of how did that feel for you guys over at

Derrick Connell: 3:33

Well, it was um it was a daunting moment because in 2003, you know, Bill Gates and Steve at the time said we we want to in this technology. So it was more of a technology bet. And Bill is, you know, if you know anything about him, he's a long-term thinker. So he thinks about what is this investment going to do for world, for for the company. So search was was it. Google were out there, they took all that share away from

Hannah Clayton-Langton: 3:57

Yes, Yahoo.

Derrick Connell: 3:58

Yeah, which you mentioned in your episode, which I I really enjoyed. It was a that they were really competing, Google versus and Google just had a much better product, and they just a much better way of searching the internet, which you about. But that technology, machine learning, big index, big crawling the web, that's the technology that Bill wanted to have, and the software development that went along with Like, how do you build a big cloud service like that? Because it's different to how the rest of the company was out.

Hannah Clayton-Langton: 4:28

Well, and I was gonna say Windows and Excel and Word, that was a CD-ROM-based business, I presume, back at the time. And that's quite different to cloud-based search engines.

Derrick Connell: 4:38

Yeah, and there were probably there was two major between the products that Microsoft had at the time and then what was happening on the internet, which then you want to you want to sort of figure out what happens if the world one way or the other. So we were on the new way, but Windows, yeah, Windows and Head, the way they were built was this old waterfall which is you figure out up front what the product is, you a year sort of figuring out the specifications, then you maybe three, four years building it, testing it, and then at the end you have a product, you burn it on a CD, or you it on a PC when you buy the PC. And with CDs, you'd send it out to companies, and then the IT department would install it on all the PCs in a company. And that's how software was distributed in the old world the cloud. But on the internet, different. It's on the cloud, accessible through a browser. And then the way you develop it is what we call agile, you're you figure out what you're trying to do. And you know, for us it was develop a high-quality search and you measure it, you're shipping every day, and putting things out in front of customers and you're experiments, and that's the modern way of modern in 2004. Now it's probably you know old compared to what we're with LLMs and ChatGPT and things like that.

Hannah Clayton-Langton: 5:54

Well, the four-year horizon is quite jarring because four years ago had ChatGPT sort of picked up the public consciousness, not even. So you wouldn't want to be planning now for four years out. If anything, I think our time horizons get shorter.

Derrick Connell: 6:06

So um exactly. I mean, back at the time, think about this. If you started working on a version of Windows in 2004 and you had a four-year cycle, when you ship it, the world is now a phone. Which, when you were specifying that product, the phone even exist, the iPhone wasn't there, Facebook wasn't there. So, how could you possibly predict the future? And that's one of the problems with Waterfall, is that really hard to predict the future.

Hannah Clayton-Langton: 6:29

So, how did Microsoft make that shift of mindset from burning onto CD ROMs? I bet there's some listeners who won't even know what That's ChatGPD. And something much different. How does that cultural shift happen in practice?

Derrick Connell: 6:44

Well, in practice, it it didn't. Um, in practice, we were almost like a separate company. So the way Microsoft operated, even at now it still similarly. You had the Windows group building Windows, shipping Windows, Office, Building Office, Shipping Office, and then the Bing team, we were in a different building.

Hannah Clayton-Langton: 7:00

Wow.

Derrick Connell: 7:00

We were building our own product, we were operating in the way, the sort of cloud-based agile development. And then, yeah, leaders, like I would go talk to my peers in Windows and Office, and we'd share information. But yeah, it was very separate.

Hannah Clayton-Langton: 7:14

So you were like an in in an incubator, especially.

Derrick Connell: 7:16

We were definitely an incubator at the time. Yeah.

Hannah Clayton-Langton: 7:18

Okay. And in that time period of change that we've mentioned, you you mentioned that the phone came out, the iPhone or smartphones out, and that moved us into the world of apps. But what was the innovation that you were building from like a product perspective?

Derrick Connell: 7:29

Well, innovation, when we started out in 2004, you know, was out there. They had 95% market share, 96% market share on the PC. The phone wasn't even there yet, so phone wasn't a big But they were the they dominated in all the countries the world. So our job was first just build what Google had. So it took us a few years to do that. And then in 2009, we released Bing to the world, and we 2% market share, 3% market share. So we were really nothing. And then the goal was over the next few years, prove that could gain market share. And we were gaining market share every single month in And you know, once we passed 10% market share, we were happy because it's like, okay, this is this thing is real. But we we knew would we ever really change the behaviors people go to Google and search on Google? Very, very difficult. So innovation meant where can you find new opportunities where you can disrupt the market? Um, because if you try to compete head-to-head, you might through, you might get a deal on mobile that you're the for the Apple phone, which we tried to do, and that was part of our strategy. But we were out there scanning the market, trying to out are there new ways to give users access to information, which is essentially what a search engine is. And so that's we were always out there scanning and trying whole bunch of different things to get access. Like we had a deal with Facebook where we were integrating search results into the Facebook search experience. I don't know if you even saw that or noticed it.

Hannah Clayton-Langton: 8:58

I didn't at the time.

Derrick Connell: 8:59

Yeah, but that was an example. We were like down in Silicon Valley doing hackathons with Mark Zuckerberg, and he was there for two days with us, and we're trying to find ways to innovate and get web search into Facebook because you know it's a huge business search. If Google has 90% share of the search market, $200 billion 10% is still a $20 billion revenue business, which is a anybody. Trevor Burrus, Jr.

Hannah Clayton-Langton: 9:23

Yeah, plenty to go around. What I'm really fascinated to talk to you about today, spoiler alert, is we talk a lot about innovation and failing fast risk taking, which I think when we think about it or when a thinks about it or a leadership team thinks about it, we think about the happy path, which is you take a risk and it pans out. Or that's at least what you hope for. But sometimes it does fail. And I I know that you've got some experience in that, uh, that's where where I'll take us next. But um you worked on an early conversational chatbot, also as Tay, which did turn out to be fairly controversial.

Derrick Connell: 9:57

Yeah.

Hannah Clayton-Langton: 9:58

Can we talk about that a little bit more?

Derrick Connell: 9:59

Yeah, I'm happy to talk about it. In fact, as a listener, I have loved the episodes like the you know, swipe on Tinder and features that were huge successes. But if you're managing a team of innovation people, you have to look at the spectrum, which is there will be hopefully a hit. And you could do a thousand experiments, try lots of like VCs do, and you may never get a hit. But if you get a hit, like wow, hit, you'll have a lot of in the middle, which no one ever really sees. They don't notice it because they were just experiments did, they came, they went, and then every now and again have a failure or something that people notice is a and and Tay is an example of that. It has its own Wikipedia page, you know, it's famous. I have visited the day with me.

Hannah Clayton-Langton: 10:43

Okay, and so let's start um with the basic. So, what was Tay?

Derrick Connell: 10:49

time. Tay was released in 2016. So imagine 10 years ago, people think Chat GPT conversing a product that gives you answers is new, two years old. But in 2016, we um in fact 2015 we had a product in China, was the the earliest version of what we call a product. And what we saw was in China, I visited China quite a bit I had a team, big research and development team. And when I would arrive in the airport, most people in Europe, when you're on the phone, you hold the phone at the side of your head and you're talking to it. So it's very clear what you're doing. There was no FaceTime back then. And so you arrive in Beijing and you would see people around with their phone in front of their face. And straight away you might think now you'd say, Well, FaceTiming. That's not what they were doing. What they were doing was they were having conversational voice input to WeChat. And so WeChat, for those of you that don't know, is a huge in China. So think about if you combined Amazon, Twitter, or X as known now, Facebook, Zoom, all in one product. Everything in one app. Everything in that one app. You can book a car, etc. That that was that's WeChat. And the way you interacted with WeChat was voice. Because in China, typing, because it's a very long language, it's really hard to type. So it's just much easier to talk to it. So we saw people talking to their phone. And when you're innovating, you're always looking things that might change the way people interact with a because that gives you an opportunity to do something new.

Hannah Clayton-Langton: 12:22

And WeChat has always been ahead of the game for us. I think I've mentioned on a previous episode I have a brother that lives in China. I've been out there and I've got WeChat because that's how talk. And 2019 I went out to visit him and it was mind-blowing to that not only could he pay, but he was authenticating with his face to pay, which now is normal to us. But they've always been at somewhat ahead of us in that respect with regards to innovation.

Derrick Connell: 12:43

Yeah, because innovation comes from at the simplest form. You look for areas where users have problems and you solve problems. And so you're in China, typing is a problem. So voice interaction was good enough at that time. It's got better over time as it gets used. So where there's areas of innovation in China is input models. One of those companies introduced WeChat as an innovation. So we're going to try this one app that does everything, it worked.

Hannah Clayton-Langton: 13:10

Okay, so you go to China, everyone's walking with a phone in of their face, and you realize that there's a interaction there, which we've not had on our side of the So what does that mean for Microsoft?

Derrick Connell: 13:23

So what we thought was let's do an experiment. Can we build an app that sits inside WeChat that has UI? Now, if you're doing a conversational UI, SearchBox, you to Hugh about this, you type two or three words and you get 10 blue links. But if you're talking to a device to get information, you're not expecting to get 10 blue links that you type on. You're expecting a conversational answer. So we we just stretched ourselves and said, I wonder if can we build a product that interacts with you and gives you answers and can also ask you questions.

Hannah Clayton-Langton: 13:55

Okay.

Derrick Connell: 13:55

So it's conversational.

Hannah Clayton-Langton: 13:56

So it sounds quite a lot like ChatGPT or LLMs generally.

Derrick Connell: 13:59

It's uh well it was the two parts of ChatGPT are it's UI. Yeah. And then it's got this incredible technology for giving you very relevant answers to the thing you asked for. So at the time there was no LLMs, but we tried the Chat GPT UI. We called it conversation as a platform. So we had the UI working, and then underneath it, we had some machine-learned answers to questions. We combined it with search, so we could answer questions what's the weather in Beijing? What's the sports results for the local team? But also, because in Beijing at the time, there was a lot people who'd moved, young people who'd moved from the to work, and they lived in apartments on their own. And there was a lot of loneliness. People were lonely. So they would talk to this product, Chow Ice, and ask like, you know, I'm lonely tonight. You know, do you have any recommendations for how I could meet someone or um where I could go to join a club? And then the product was answering. And then what we decided to do was we didn't want it to be one way. We wanted it to be almost like a human conversation, which is every now and again the product would ask the user a you know, how are you feeling tonight? And backwards and forwards. And so that proved to be a huge success.

Hannah Clayton-Langton: 15:02

And this wasn't Tay, this this was experimenting this was the first China first. China first, okay.

Derrick Connell: 15:08

And so it worked incredibly well. We ended up with over a hundred million daily users in China using this product. Yeah. A human conversation has seven turns. So I ask you something, you ask me something.

Speaker 2: 15:20

On average, you mean okay.

Derrick Connell: 15:21

And so we were just trying to see would it work? And the average in for Chao Ice was 23.4 on average. So it was sustaining this conversation backwards and forwards. And the data they were using was, you know, we sourced, you talk about training data with LLMs. It was similar. We had training data, which was questions and answers on sites and all kinds of things. We used that to generate Q and A.

Hannah Clayton-Langton: 15:45

And that's fixed training data that you would load in.

Derrick Connell: 15:48

We loaded that to bootstrap it. Yeah. And then once the model started working, we were getting all of this new training data, which is the questions that was asking us, the answer that was giving it back. We would ask the user a question and get the answer. So we were getting more and more training data all the time, and then we were using machine learning to make the models So huge success. So then we thought, why don't we try it in the US? Take our experiment. What could go wrong? And that was Tay.

Hannah Clayton-Langton: 16:14

That was Tay.

Derrick Connell: 16:15

Yeah. Same model. So we didn't have WeChat and we didn't have voice. So what we thought was we have Twitter, you know, it's a place where people talk to one another, and we'll do text input to start with. You know, maybe some people will use voice, but we'll an equivalent in the US. And we released Tay, I think it was March 2016, and Tay was a product for 18 hours.

Hannah Clayton-Langton: 16:39

18 hours. Okay. So how quickly did it go wrong?

Derrick Connell: 16:43

It went wrong in almost the first hour.

Hannah Clayton-Langton: 16:45

Wow.

Derrick Connell: 16:46

Um, because we had a we made a mistake. We made a few mistakes, but we made one big mistake, which the in order to get training data, we had this feature, was repeat after me. You know, the kids thing, repeat after me. No, as kids we learned, we should have known, we learned repeat after me has some side effects, which is, you know, get people to repeat after me.

Hannah Clayton-Langton: 17:05

Come on, you game it even as a child without the tech.

Derrick Connell: 17:08

And so we had a bot attack in the first hour, where the version of bots, we had a number of servers that started to teach Tay very bad things, racist things, xenophobic anti-Semitic things. So Tay was learning all of these things and taking it in, it became the training data. And so within an hour, it became a really bad um piece of

Hannah Clayton-Langton: 17:31

Okay, so Tay launches big launch or no, not quite.

Derrick Connell: 17:35

It was a it was an incubation. It was just branded Tay, separate story on the name, because we had some other issues with the name.

Hannah Clayton-Langton: 17:40

With Taylor Swift.

Derrick Connell: 17:41

With Taylor Swift, yeah. We didn't have to deal with that because we closed it down quickly. Um but yeah, it was a it was a big mistake, relatively But it was yeah, it was historical in the history of Chat and conversational UIs, it was one of the moments where big moment in in its history. Um it was a failure. Trevor Burrus, Jr.

Hannah Clayton-Langton: 18:02

Okay, so when we say fail fast, within an hour, as the product team behind it, you know it's failed, and then it takes 17 hours to close down. And in practice, what's happening? This thing goes out on Twitter, and then 42 minutes in it's things that I won't say.

Derrick Connell: 18:20

Yeah, yeah. You can look it up for yourself on Wikipedia.

Hannah Clayton-Langton: 18:23

Okay. And that is when you say a bot attack, just for the the lay what does that mean in practice? It was a not a nefarious actor or no, it was a nefarious

Derrick Connell: 18:32

So uh even today you'll have uh bots out there, which are server farms that are pretending to be real people. And so you know, you see them today. There's bots on Twitter, bots on Facebook that pretend to be real people and they'll they're machines.

Hannah Clayton-Langton: 18:46

But someone's put them up to it.

Derrick Connell: 18:47

Someone's put them up to it, and then you can because it's a machine, you'd have thousands of accounts on the server teaching or or writing things on the internet. So that's what happened. Within an hour or two, Tay is learning, takes a few hours it to learn, and then of course the same people who've done the bot attack go in and ask Tay the questions that it's to give certain answers to, screenshot it, and then start it out on the internet.

Hannah Clayton-Langton: 19:14

And is Tay known to be linked to Microsoft at this point?

Derrick Connell: 19:17

Uh no, well, it's it's not branded a Microsoft product it's an incubation. But if you dig around, you can see that it's a project from Microsoft.

Hannah Clayton-Langton: 19:24

Okay, so someone's tweeting out Microsoft.

Derrick Connell: 19:26

Yeah, Microsoft just released this racing.

Hannah Clayton-Langton: 19:28

Maybe C News or something. Okay, okay. And how bad was your day? Like how how does that play out for the teams back at home base?

Derrick Connell: 19:35

Well, I mean, I I would say the day, the 24 hours from the you realize, you go through three stages. Stage one is get this thing stopped, like stop. Um, deal with this problem. You know, it's a fire, put the fire out, so you're in mode. Then you go, okay, the fire is put out, and then you you of, as you would, you say, What happened? Like, how did this happen? You're diagnosing the problem and trying to figure out, okay, we made a mistake, but how did this happen? You figure out it was a bot attack, and you try to figure who was it, how do they know it was coming? Because it was a it was a solid release, and normally you put things out on the internet, nobody notices, so knew. How did that happen? So you know, you're doing the post-mortem and trying to figure out can we fix this and re-release it? Should we re-release this? I very quickly realized no, it's an incubation. We're not re-releasing this, but that's stage two. And then stage three is you have to face the consequences of you've taken a risk. Sometimes they hit, most of the time they flop, sometimes fail publicly in a big way. So, you know, then you have to deal with that. Explain to the rest of the company what went wrong and go the CEO and explain what happened because you know, big PR I guess you could call it.

Hannah Clayton-Langton: 20:47

Were you that sort of the point person?

Derrick Connell: 20:49

Yeah, we came, there's a few of us. 7 a.m. go see the CEO.

Hannah Clayton-Langton: 20:52

7 a.m.

Derrick Connell: 20:53

Yeah, you know, first thing. You know it's a problem when you're the first thing in to the headmaster. Yeah, yeah. But you know, uh, one of the things about innovation is, and this is a leadership opportunity for anyone out there who running a team doing incubation, things can go wrong. And the question as the leader of that organization or CEO a big company like Microsoft is what do you do? Because you can react in a number of ways. And I was so happy with the reaction of our CEO because we in, it's obviously a big problem. And the first question he asked was, What did we learn?

Hannah Clayton-Langton: 21:28

Not even no swearing.

Derrick Connell: 21:30

No swearing. What did we learn? And that is someone who knows how to manage innovation, how to motivate a team. You know, obviously hold us responsible because something's gone wrong. But first question, what did we learn?

Hannah Clayton-Langton: 21:44

And he's he but he's also, I guess, thinking back to he's expecting you to have that answer.

Derrick Connell: 21:49

Of course.

Hannah Clayton-Langton: 21:50

It would almost be worse if it had happened and you didn't have the answer than if it hadn't happened at all.

Derrick Connell: 21:54

Yes. Now, was he happy with what we learned? He was happy with some parts of it, and he was unhappy. With some parts, but you know, because we made a mistake, mistake, repeat after me. Come on, rookie mistake. But he was very happy that we tried, he was very happy the learnings about conversational UI, influenced a lot decisions that Microsoft have made over the subsequent years, investing in LLMs to make better technology, investing in UI.

Hannah Clayton-Langton: 22:19

And when we think about the learning there, there's different levels, right? There's Derek Connell and Piers. What did you learn? And I presume you've taken something specific away from that on a very personal level. And then there's what did the company learn with regards to And I think what did the industry learn with regards to the So can you just talk us through those different levels?

Derrick Connell: 22:39

Yeah. So what did I learn? I mean, A, I learned keep trying. The overall thing was a hit. I mean, there was interest in this product.

Hannah Clayton-Langton: 22:47

Aaron Powell And did you work out how someone found out it being was there like a leak?

Derrick Connell: 22:53

Yeah, that would that would be my suspicion. We never found out exactly who. Um but we did I I have an idea where it came from, which part of the product it came from.

Hannah Clayton-Langton: 23:03

You can tell me off mic.

Derrick Connell: 23:04

Yeah.

Hannah Clayton-Langton: 23:05

But that's I suppose positive because a leak means it's

Derrick Connell: 23:08

So I also learned that there's differences between China the US. So context matters. Yeah. Um, because in in China, you know, for good or for bad, the country is very regulated uh on the internet usage. So therefore users have learned to become self-regulated. Whereas in the US there is no um self-regulation when it to what people do on the internet. But the part that was very positive was we learned that a UI was a good direction to go.

Hannah Clayton-Langton: 23:36

Okay. But that that point is so interesting. Your test group essentially, which was the the users, were self-regulating for cultural reasons that you guys just didn't anticipate.

Derrick Connell: 23:49

Yeah.

Hannah Clayton-Langton: 23:50

Wow. Okay.

Derrick Connell: 23:51

In hindsight, hindsight's great, 2020. You look back and go, of course we should have known that. Um but that's why we did the test, right? And you you have to test these things and put them out there. And to be honest, without the repeat after me feature, you it could have been interesting. But yeah, we were a little gun shy after that, you know, do it. But certainly pursued the conversation UI.

Hannah Clayton-Langton: 24:11

Aaron Powell, and did you have Repeat After Me on in China?

Derrick Connell: 24:14

No, because we had enough training data. So this is where when you're sometimes when you're you have a problem to solve, and then you have to find a to solve that problem. And then that's where that mistake came in. In China, there was just a lot of publicly accessible training data with questions and answers on shopping comparison sites, versus the US, it was a not as good So we thought, oh, we'll put that feature in because it'll QA data that we could use for training. Trevor Burrus, Jr.

Hannah Clayton-Langton: 24:42

And do you think the the takeaway that it was an interesting that was an industry-wide takeaway, I presume?

Derrick Connell: 24:47

Aaron Powell Yeah, I think there was a few things that Because I know for sure, because we talked to people at and other uh companies, that the Tay is known. So people at OpenAI know about Tay. Conversational UI, we did a lot of work over the following three years out there just talking about conversational weirdly talking to people in the valley about China and them to WeChat. I was amazed that people didn't know about WeChat in China in 2019.

Hannah Clayton-Langton: 25:14

Yeah.

Derrick Connell: 25:14

But just talking about that UI model. Um, so that has absolutely taken off. And like you look at Copilot, um, Microsoft Copilot, it's conversational UI and it's it's taken off big time. And then we also realized the tech underlying technology scale.

Hannah Clayton-Langton: 25:29

Okay.

Derrick Connell: 25:30

So you could only do so much with QA data and machine before it just reached its limits of capability. And then the transformer technology was the big breakthrough. Transformers leading to LLMs, and now you have technology you can combine with that conversational UI, and then you get ChatGPT. It's great.

Hannah Clayton-Langton: 25:49

And why is, to state the obvious, why isn't ChatGPT telling racist things? Like what's the correction that we've made there?

Derrick Connell: 25:56

Yeah. Well, A, they learned that you you know you need to put on, which is normal. It's like if you go do a Google search, it has filters that you will, you know, remove certain terms, not answer questions, and so the same technology is used. And then the training data is it's the LLMs, and those are phenomenal. So you have the data, you don't have to do these tricks. So if I test my technical knowledge, yeah, go for it, because you you should know now, because you've got all these where you've been live.

Hannah Clayton-Langton: 26:23

Listen up. The training data is a whole process that happens before a new model. So it's fixed and therefore filtered. Although interesting debate around where the line with filtering sits. But that's fixed, so it can't be disrupted in the same way your TAI live training data was essentially like manipulated by these bots.

Derrick Connell: 26:46

Yes. So there's two parts. There's the training data set. So with LMs, it's everything that's public. And Mez talked about it uh two episodes ago. The entire web of human knowledge that's publicly available as its training data. But then the other part, which you talked about in the search session, is there's two parts. There's the signal we get from the questions that people are asking. Search is pretty clear. Search for something, you scroll the page, you click the third link, you go to the page, and you don't come back. That's a strong signal that number three should have been one. So when you retrain the ranking model, you'll get number and the number one position. You haven't changed the underlying training data, but you've changed the ranking model. So ChatGPT can use that signal that here's the questions are asking. They asked a follow-on question, they asked the third they say thank you, and that signal that this conversation useful and the answers the model gave were good, so we can the ranking model for the Which will be more of a live Yeah, they can they can do that pretty quickly.

Hannah Clayton-Langton: 27:45

Okay.

Derrick Connell: 27:45

The underlying model gets updated less frequently, but the question that the user's asking and then reaching in, mean, you're reaching into billions of documents to pull the relevant answer. So if you answer the question incorrectly the first time, got that wrong. You got it right the second time, you can retrain. And then when you reach in, you pull out the right answer.

Hannah Clayton-Langton: 28:05

Aaron Powell And if there was a a lesson for the industry on and AI, let's call it, and you're training a model, an LLM, all the data from the internet, do they filter out the bad from the internet, or are they filtering at the point of the interaction?

Derrick Connell: 28:21

Aaron Powell Yeah, there's a few places that filtering I mean, in the whole thing gets filtered in in some ways. I mean, take Google as an example, let's go back to search. They call it the dark web, but Google knows about It knows about all the websites that exist that it can Um, just websites that exist behind firewalls, enterprise doesn't know about those. It knows there's something there, but it can't get it. But it knows about everything out there. And then it chooses which subset it's going to go crawl and make a copy of. I know you were surprised to hear that it's a copy. Yeah. But it goes and gets that copy, but it's a subset. So that's filtering because it's choosing which ones to back. Um, it's choosing how often to go back and get a fresh So that's filtering. So that's on the index side. It's choosing what information to bring back. Um the answers you see, the type sports questions, they've which types of answers they're gonna give. So again, they're filtering which sports you can see. You might not get women's cricket, but you might get men's It's like, okay, well, there's something happening there. It's filtering. That's on the index side. LMs are different because it's got everything. Yeah, it's also look, it's they have to manage their cost So they're also gonna choose which subset of documents gonna use for their models. So you're gonna get a subset, still pretty big.

Speaker 2: 29:31

Yeah.

Derrick Connell: 29:32

Then on the question side, the filtering we did, and Hugh I worked on this together, let's say on image and Yeah, you have to choose as a question gets asked to you, you have to decide what is this question? What's it likely to be? And one of the first things you will decide is is this a question or not? And that's a really important decision.

Hannah Clayton-Langton: 29:50

As in legal versus illegal.

Derrick Connell: 29:52

Yes.

Hannah Clayton-Langton: 29:53

Okay.

Derrick Connell: 29:53

You know, you can imagine types of things that people might be asking that are I don't want to.

Hannah Clayton-Langton: 29:57

Exactly.

Derrick Connell: 29:57

But you choose, you have to choose whether or not the is something we should answer. Yeah. Um, and then there are things that are more sensitive. Like if someone types, you know, how would I commit That's a question that you have to treat very sensitively. Yes. And you're not going to tell them. Yeah. You're going to give them help.

Speaker 2: 30:14

Yeah.

Derrick Connell: 30:15

And so you're choosing to answer, but not in the way that asked the question. You're interpreting it. And so that's a good thing to do. So you're always doing filtering, classification, and those are all very sophisticated models that are using learning to figure out what is this question? Should I answer it? And then how should I answer it? Should I give the user an answer or give them advice? And LMs are the same.

Hannah Clayton-Langton: 30:37

And just to take us back to your Tay disaster and the learnings, was the pivot point for the industry on ethics and AI around or did it that conversation exists and this sort of played into what not to do?

Derrick Connell: 30:50

Well, that's a good question because two weeks later, one of the big insights for Satya, the CEO at the time, was okay, we made a mistake. We learned a lot of things. One of the things we learned was we don't have a set of We decided to call it ethics. And there was a whole bunch of senior people got together, legal people, PR people, marketing people, engineering Satya, and came up with Microsoft's first set of AI ethics, which two weeks after Tay, we announced. To the world.

Hannah Clayton-Langton: 31:19

Externally, okay.

Derrick Connell: 31:19

Yeah, we created them and announced them because it was a good opportunity where it was in the news, people were listening.

Hannah Clayton-Langton: 31:26

Well, and a good PR redemption arc, I suppose.

Derrick Connell: 31:28

Yeah, and also Satya is a great leader because he thought this is a good opportunity for us to actually do the right thing, like produce a set of ethics, some guidelines. They were adopted by the industry.

Hannah Clayton-Langton: 31:39

So do the industry you think didn't have much of a finger pulse on that stuff before?

Derrick Connell: 31:44

No, because at the time, how many real AI products were out there? I mean, you could say search was an AI product, but it using LLMs because they didn't exist at the time, but it did have the most sophisticated machine learning. Google and Microsoft had a lot of that. Amazon had some machine learning, but there wasn't a lot of like of actual AI products at the time. I mean, now we're thinking, of course, everything's an But even 10 years ago, there were only really there was one product that was AI, which was search. And then everything else was kind of a little bit AI, but not much.

Hannah Clayton-Langton: 32:17

And you're quite relaxed about the the mistake, as it were. Because of the learning opportunity and because it was well managed.

Derrick Connell: 32:24

There was there any point in that whole period where you were really like No, I I don't want to appear nonchalant about it, but with AI, you have to think about the potential damage.

Speaker 2: 32:35

Yeah.

Derrick Connell: 32:36

And you think about Tay, nobody was hurt. There was nobody who was physically hurt. It wasn't a product that somebody could get injured, like cars. There's a potential for someone to get hurt. Like if you're building a self-driving car, that's a real that somebody might get physically injured. Um now I know there will be people who use Tay who were and that's a that was a problem, but they weren't physically hurt. You know, company didn't lose any money, we had no revenue

Hannah Clayton-Langton: 33:02

And you shut it down quickly, which shut it down in 18 yeah.

Derrick Connell: 33:04

And then didn't bring it back again. So yeah, okay, a company like Microsoft, it's a PR problem. You know, but we're big boys, we can take the hit. So that's why I look back and go, yeah, it was a mistake, but it forced us to think about the ethics of AI, not hurting not damaging people, being mindful about the guidelines, governance. You know, even today I was advising a company and they about the person who's their responsible AI leader. And every company now has a either an AI ethics team or a AI officer, and that's good. Someone is thinking about what damage could we do with with AI and be mindful about it, put some guardrails in place, from the mistake we made, which is unexpected things will unless you plan for it. So mistakes will happen, not might happen. So be careful.

Hannah Clayton-Langton: 33:50

Be careful, but I guess it's good to go early with something and experimental because you can take a learning that can be a hundred times over to everything that goes next.

Derrick Connell: 34:00

Yeah. We were doing hundreds of experiments every month. Yeah. Trying new things because we were up against Google.

Hannah Clayton-Langton: 34:06

Yeah.

Derrick Connell: 34:06

We weren't expected to win. We didn't. But you know, there's no reason why we should have. They had a dominant position, so we we were forced to It was great. Yeah, it sounds like some. I'll tell you, you know, some people say, wouldn't you have rather have been at Google at the time, like working for And I know Hugh mentioned he had an offer from Google in 99 and I had one in 2001, but no, because I competed against for a long time. So this might feel like being competitive, which is which it probably is. Um, but we had to be innovative.

Hannah Clayton-Langton: 34:32

And scrappy almost. We had to be scrappy. Yeah, yeah.

Derrick Connell: 34:34

You know, we were one-third in terms of the number of We had a lot of people, you know, the rest of people at used to say we were too big, you know, four and a half people were like, How many people? Yeah, yeah. Well, Google were 15,000, so we were small, and they had money to spend on bigger indexes and et cetera. But we had to be scrappy, which meant we had to innovate. We had to find ways to have an index as good as Google's less money. Yeah. We had to find ways to be as good at better at ranking in some places, and you know, we were competing, we were innovating. And then we had to find ways to introduce new products. We were like, you know, we had to, which was, I think, more fun than being the guy on top watching us and kind of having to react to what we did. Yeah. Um, whereas they had their their business model and they steadily just going out there and and using it and making good money. But I don't know, were they innovating as much? I don't think so.

Hannah Clayton-Langton: 35:21

But well, innovation comes out of problem solving. And if your competitor's got 90% market share, you can't do You bet. I mean look at Gemini.

Derrick Connell: 35:27

But look at Gemini right now. Yeah. Why is why is ChatGPT or Claude the most used product by Why isn't Gemini? It's a great product. They have much bigger data set, they have lots of users. Why? Because they're protecting their search business. So I would not like to be on the Gemini team because you're out there thinking, oh, we could be number one.

Hannah Clayton-Langton: 35:45

Yeah, yeah. And any other before we leave this Tay topic, any other big either personally or that you took to your next projects.

Derrick Connell: 35:55

I mean, I mentioned them, but uh just to summarize it. I think there's the leadership. If you're leading an innovation team, your range needs to we could hit, but you also need to be aware of the fact it could be a huge failure. And if you restrict your innovation to safe, and then if you are leading that team, when you fail, I hope that I would the leader who would say, What did you learn? Yeah, as the first question.

Hannah Clayton-Langton: 36:18

Not who do I fire.

Derrick Connell: 36:19

Who do I fire? Screaming and roaring. It's like, what did you? I mean, I still to this day think that is one of the single best questions I was ever asked in a moment like that, yeah, when only a handful of people would ask that question. Great question. So leadership. Um, I'm gonna quote Mez, who was on two weeks ago. He said he's an AI optimist. I am too, because those experiments showed that there is a interaction model. Yeah, and how cool is that? Yeah. I mean, come on. I mean, we're if you'd asked somebody ten years ago, will ever not use Google search?

Hannah Clayton-Langton: 36:52

Not ten years ago.

Derrick Connell: 36:53

Five years ago. Two years ago. Yeah. You would say, of course not, because it's perfect. Yeah, like brushing my teeth. Why would I change? It's perfect for the job. And now the world has changed.

Speaker 2: 37:03

Yeah.

Derrick Connell: 37:04

There's something better, and that's great. I mean, come on. As we progress, this technology is getting better and And then the third learning is we do need to be Yeah. You know, for the people who are AI optimists like I am, you can be optimistic and innovate on the edges, but be Yeah. And think about what's the worst thing that could happen. And be that's the thing to be careful about, which is if you have the potential to hurt people, be very careful. Yeah. Is what you're doing going to you know cause a lot of to lose their jobs? Be careful, be thoughtful, be mindful.

Hannah Clayton-Langton: 37:35

And do you think that the big LLM players, the open AIs of the world, are leading or cognizant of some of those questions you talk about people's jobs? Um It's an interesting discussion.

Derrick Connell: 37:46

Yeah, no, of course. They're they're thinking about it. Um, but they're not the ones who are thinking about it the because they're more of a platform. It's all the people building, all these apps that are built now in enterprise. Like if you're a lawyer, you're probably thinking, ooh, is a company, a startup that's coming along that's gonna the legal business where people can get access to legal without needing to come to a human. And then do we need more lawyers? Do we need less lawyers? I mean, that those are questions that that startup will think about. But there's money to be made, there's a business opportunity I could solve a lot of problems, I could democratize access to those.

Hannah Clayton-Langton: 38:21

Well, I was gonna say actually, there's there's two sides because dem democratizing access to things that sit behind a service does have benefit. I think Mes talked about this a little bit in the episode week. So it's it's a difficult line to walk.

Derrick Connell: 38:33

Yeah. But the world changes. You know, we used to ride horses, yeah, and then we drove yeah. Now we're in driverless cars like Waymo. Yeah. I think he was using them exclusively as well. He's a bamo. A big Waymo fan, you know. Um and then what's next? Who knows? So things change and humans adapt to that change. And and and we just have to be mindful of, I mean, mostly it would be for the governments to think about for its what's going to change, what could happen for our for our people, so that we're part of this and that we're people for the future. You know, the worst thing you could do is stick your head in the sand and pretend it's that nothing's changing. Yeah. You know, for the next five years, apps are gonna come out. I mean, I built four apps yesterday.

Hannah Clayton-Langton: 39:14

Well, even even for you, that's pretty good going, isn't it?

Derrick Connell: 39:17

Yeah, but I uh when I talk about innovation, yeah, is it a or a flop? It's probably somewhere in the middle.

Hannah Clayton-Langton: 39:22

Yeah. Yeah, yeah.

Derrick Connell: 39:23

So I haven't done anything interesting yesterday, but that I could do four apps in parallel, I just sent them off and they got built by uh AI, and then I had them at the end of the day, it was great.

Hannah Clayton-Langton: 39:34

Yeah.

Derrick Connell: 39:34

So maybe one of them will be a hit.

Hannah Clayton-Langton: 39:36

And but just back to Tay, because I did do my research, as you say, this was a moment that the the world noticed.

Derrick Connell: 39:43

Yeah.

Hannah Clayton-Langton: 39:46

March, did it? There was uh a little bit of a hiccup a couple weeks later.

Derrick Connell: 39:49

It was um it was a dark moment and it did come back to life. It was like a zombie, you know, coming back from the depths of hell, climbing out in your dreams. It was. It was late at night. We were, you know, it was um we were down at San Francisco. In fact, it was the next day Sathy was going to announce, did announce our ethics and so Tay's happened.

Hannah Clayton-Langton: 40:11

Yep. We've we've learned. Yeah, we turned it off, we've learned, and actually, pretty those ethical guidelines were spun up by the business. That's great. But the day before it's being launched to the not just the to the world, it comes back from the dead.

Derrick Connell: 40:27

It woke up uh about 10 30 p.m. And then we were told about it an hour later. I got a call. I thought it was a joke.

Hannah Clayton-Langton: 40:36

Tay is back on the corporate sabotage.

Derrick Connell: 40:38

No, I I I at that moment at 11 30 p.m. I thought I don't want to swear, but I was like WTF. Yeah. It's like 11 30. I thought it was a joke. I honestly thought someone was pranking me. So I get online and lo and behold, there it is. There it is. And so it's like, what the heck? And so, you know, we gathered a few people that had a PR, couple of people got together and 'cause you you were in at least you were in the same place because you're actually Thankfully we're in the same place. It wouldn't have mattered where we were in the world, we would have been on the phone call together. But luckily we were in a got together sitting in a hotel room on my bed with a little speaker phone on the bed. We got in touch with the developer in China who uh lo and it decided to turn the server back on to diagnose the code, but not realize that when he turned it back on, it reconnected to Twitter.

Hannah Clayton-Langton: 41:24

And then someone noticed and the press and so you guys found via the press that it was back online.

Derrick Connell: 41:28

We someone tweeted it. Tay is back. Return of Tay. And we got on the phone, and you know, by the time we got in touch with this developer, and it's like, what the on over there? He's like, What are you talking about? It's it's back online.

Hannah Clayton-Langton: 41:41

And how do you get it offline?

Derrick Connell: 41:43

Well, I gave the instruction go to the wall, find the plug it out, turn that thing off. Yeah, it's done. It was on one server.

Hannah Clayton-Langton: 41:52

Okay.

Derrick Connell: 41:53

And it was done. Time of death, 1.38 a.m.

Hannah Clayton-Langton: 41:56

Oh my gosh. And why couldn't he unplug it at 1034 or whatever? Like take away.

Derrick Connell: 42:00

Well, he didn't know it was back online. He was just sitting at his machine looking at the looking the code. Oh my god. Hadn't realized that it connected back to Twitter.

Hannah Clayton-Langton: 42:07

That's a bad day for him.

Derrick Connell: 42:08

It it yeah. What I asked him was, what did you learn?

unknown: 42:11

Yeah.

Hannah Clayton-Langton: 42:13

Okay. Derek, you have a book coming out this week or next week at time of recording. Uh can you tell us a little bit more about that?

Derrick Connell: 42:21

Uh yeah, thank you for asking. Um it's a different kind of book. So since I, you know, stopped um working full-time for I've started pursuing a whole bunch of things that I didn't have time to do when I was working, because you know what like. You're working 80 hours a week, seven days a week, you know, you're just working, which is fun. But the book is called 21 Summers. It's co-authored with Thomas Warner, who's a guy I met. He's a New York Film Academy professor of photography and arts. And when we met, we realized, I mean, I think he was meeting me because you know, I went there to do a workshop. So I love fashion photography. Okay. Uh weirdly. And I went there for a workshop and was working with him and he heard there's this tech exec coming from Microsoft, to learn about you know, photography. He told me later that he was dreading it, but within an hour, I was telling, you know, he started asking me something about technology. I was like, oh yeah, you know, your most personal is with your search engine. And he was fascinating.

Hannah Clayton-Langton: 43:14

Well, you told me that earlier, and it's it's really rung with me. There is nothing more personal than my Google searches

Derrick Connell: 43:20

You should always be aware of the fact that the most relationship in your life is with your search engine.

Hannah Clayton-Langton: 43:25

Well, maybe it's now with my LLM of choice.

Derrick Connell: 43:28

Um Maybe. I don't think so yet. I don't think the user model I I have a feeling that that model, the conversational model, is more like a human Trevor Burrus, Jr. So you self-regulate. You self-regulate. You self-regulate because you're expressing yourself the way you would to another person. I mean it's not anthropomorphic, but it feels a bit So I still think the search engine is the place you can The cold ten blue links with the sorry, we digress. Oh no, that's okay. But the uh then we realized that we had a lot in though he's a professor of fine arts and photography. And I'm a tech guy. And then we started talking about pursuing dreams and I wanted to do and the things he wanted to do. So we decided to work on this book about exactly that. It's called 21 Summers because you know when you're a kid, you're thinking about the summers that you the big things you do during the summer, you know, go traveling or take up a new skill or whatever it might be. But it's like three months of just go do something, with vigor and energy and passion. And when you think about our life, it's fixed. We're all exiting the same way. Yeah. And you know, in reality, we've probably got 21 of those left. So every time you miss an opportunity to pursue one of those 21, it just gets ticked off, you know, 19, 18, 17.

Hannah Clayton-Langton: 44:40

I spent the last summer of my life in Cincinnati, Ohio. So you've given me some some cause for thought. That was one of my 21.

Derrick Connell: 44:46

That was one of your 21s. Was it the one you would have chosen?

Hannah Clayton-Langton: 44:48

No, it was interesting. It was interesting. Yeah, yeah, yeah. An adventure.

Derrick Connell: 44:51

But I I realized that I had been self-regulating. I mean, when I got into fashion photography, I was telling nobody else was stopping me.

Speaker 2: 44:58

Yeah.

Derrick Connell: 44:59

Nobody. Of course. I could do it. Yeah. But I was telling myself all these reasons why I couldn't, I'm a tech guy. No one expects me to be a fashion photographer. And so we went down and we talked to about 17 different around the world, different kinds of people, EMTs, a lady in Africa who runs an NGO, uh Buddhist uh makeup artist from Japan. And the thing that came out was that most people, there's stopping them from pursuing their dreams apart from spots we call them. The voices in your head giving you reasons not to do it. So we just explored that idea in the book, look at the blind spots, and then give some examples of people who have overcome those blind spots in their life, some techniques for them. So that the goal is to get people moving so that they can a fun life.

Hannah Clayton-Langton: 45:46

And you're going to be spending one of your summers, I on a book tour. So that's pretty cool.

Derrick Connell: 45:50

Yeah. Well, actually, my the book is a 21 summer for me. Like just publishing a book, being an author. I'll be an author next week. By the time this is published, I will be an author, wasn't before. So that's a 21 summer. But I am going to Turkey, going to Istanbul for two months on a fashion photography residency. Wow. That's a month of 21. Yeah. And the book, that this idea of just go for it. Get out of your own way. There's lots of other reasons that are you can't do Yeah. But don't limit yourself.

Hannah Clayton-Langton: 46:18

Yeah. And um, can the listeners find your photography anywhere?

Derrick Connell: 46:22

Oh, yeah, on my Instagram. On your Instagram. My Instagram is dedicated to my photography.

Hannah Clayton-Langton: 46:27

And what's the tag?

Derrick Connell: 46:28

It's at Derek underscore Connell.

Hannah Clayton-Langton: 46:30

Okay, perfect. Very simple. Okay, great. a wrap, guys. Thank you so much for listening. As always, you can find us on socials, including LinkedIn, X, TikTok, or on YouTube. And of course, you can go to techoverflowpodcast.com. See you next week.