Centaurs and the Future of AI
Livestream & Transcript
It's been a year of revolutionary advancements: computers can now write business letters (GPT-3), create photos and drawings (DALL-E), and hold a conversation (Chat-GPT). The underlying technology is being used all over in less obvious ways, to write software and recognise faces and even for psychotherapy. It may seem that machines are about to take over, that there's nothing left for humans to do.
But these innovations aren't sui generis—we've been here before, in many fields where computers today perform tasks that Turing and von Neumann couldn't have imagined: running factories, simulating fantasy worlds, planning driving routes. And in each case, the "human in the loop" has been a key element of the solution, not a useless appendage. Think of grocery-store self-scanning checkouts or the map app on your phone: we don't blindly trust the machine, we attach ourselves to it and guide and check its actions, forming what chessplayers call a "centaur", a human-computer hybrid. The skills and behaviours needed to birth, train, and manage these centaurs are different from what we needed to run teams of knowledge workers, which means you need to upgrade your hiring, operations, and strategy—but don't close down your call centre or development team just yet.
Join me on this free livestream and learn how these changes are going to affect your business this year:
Why the history of chess computers encapsulates the evolution of machine "intelligence" and how humans interact with it.
How to think about the apparent sentience of chatbots, which are actually "method actors".
What my best clients are doing to create flocks of centaurs that stay well ahead of the evolution of AI, in industries from construction to customer service.
Here’s the transcript:
Speaker (00:03):
It says we're live. Okay. I believe we're live. Laura will signal me if anything is not working as it should. But as far as I know, you're here with a Squirrel Squadron. Hello Squadron. We'll give everybody a moment or two to turn up. In the meantime, I will say hello to folks watching this on a recording. And this is always available on the Squirrel Squadron Forum, as well as on the platform you're watching on it now. So watching it on now. So if you need to go somewhere else or if something comes up, don't worry. The recording will always be here and you can talk to, talk to me there. Well, you can talk to me on the recording, but I won't answer. What's the good news is this is live and interactive. So I'm looking for questions and comments and thoughts from you guys about what is all this crazy artificial intelligence,GPT three stuff.
Speaker (00:48):
So we'll cover a lot of that we'll get into it for, for sure. If you're here live let me know what brought you here. Give me a quick chat message. Just let me understand what are your questions? What are the topics that may are making you interested? Are you using artificial intelligence? Are you interested? Do you know what a Centaur is? That's the crucial trivia question for you of the day. Do you know what the ancient Greeks called a Centaur? I suspect most of you do, right? So the place I asked people to comment beforehand was on the Squirrel Squadron Forum. We had a little poll going there, so I'll report on the results of that later. That's from the Squirrel Squadron community, which is over a thousand, over 1200 now people who are both technical and non-technical who are joined together to learn from each other.
Speaker (01:41):
So that's why I started the Squirrel Squadron. I don't know anybody else who, who gets that particular group of people together. We have the forum. We have these weekly live events. Next week I'll be live in London, so you can come along. All this stuff is free. I never, this is my way of giving back. So we'll have an event next week on how to Get Your Tech Team to Deliver every day So you can see results instantly. If you're tired of waiting if you're an frustrated engineer what techniques can you use? How can you deliver every single day using something called Elephant Carpatio? Then I'm going to be in Crackow just got this up on the website. So I will be there the 18th of May. Not sure what the topic will be yet, but it'll be really fascinating.
Speaker (02:24):
It's part of the ACE Conference. So more and more exciting things coming in the Squirrel Squadron. Lots of opportunities looking forward to, to hearing from you guys and seeing many of you at the live events. Okay, so I'm going to start with a, a little information not about the squadron, but about the, the poll I took. So I was interested in who what and why people were thinking about artificial intelligence. Now. It's all over the press, of course. So you're getting to hear about it pretty much every time you turn around. Thanks, Steven. Glad to see you. And somebody put you up on the screen, so that's great. But I'll answer your question. I'll deal with your comments in just a second. The poll I put up was all about how you are actually using artificial intelligence today.
Speaker (03:15):
And some of you may not be using it at all. That's great. We'll talk about how you might want to use it in the future, and what things to watch out for, how it's going to change things for you. So, the results were that most people seem to be using the artificial intelligence tools like chat,GPT for answering a question or composing something. They're writing some text. And of course, this is the great worry at the universities is that everyone will not only, they used to look it up on Wikipedia now, they'll just look it up in chat,GPT, and then they'll ask chatGPT to write a five paragraph essay on the topic and, and go lie on the beach. There's some argument that might actually be better than what students do today, but it sounds like most people are using these tools in that way.
Speaker (03:57):
Today, we have some examples where people are correcting some text or summarizing what they have to say, or, or they're using it for something else. They didn't tell me what it was. Nobody is using it for one of my favorite purposes, which is practicing conversations. So, I want to say more about that when I get onto specifics on, on ChatGPT. But more generally, it seems like the thing people are interested in is not quite the right thing. I think it's not actually playing to the strengths of the AI. And there are lots of ways to use artificial intelligence, use the new tools that we have these days in more clever ways. And that's what I want to talk about next. So, I'll come back to this survey, but I just wanted to start with the results since I thought people might be interested.
Speaker (04:39):
Okay. So let me tell you a story now about chess. So I used to play chess. I wouldn't call myself even a semi-pro, but I played it very seriously as a teenager. And I got to learn about how people developed chess computers. And the history of chess computers is really interesting. In the 1950s or sixties, it was astonishing that anybody could play chess, that the computer could actually make a move and it wouldn't be an illegal move In the seventies or eighties started to be able to beat some not very good humans but still the world champion would clean the floor with any chess computer. The chess computer really couldn't play against humans in any meaningful way. Then there was this huge leap forward. And the chess computers actually got to the point very, very quickly, almost overnight, like in months that they could beat even grand masters.
Speaker (05:31):
Even the world champion would get beat by a chess computer. And this is where it became super interesting and very strange. And where Centaur come into the picture, because somebody discovered the very interesting idea that if you took a chess computer and you sat it next to a chess Grand master, and you had them play together cooperatively playing on the same side of the board, you would get much better results. And that's what we call a Centaur in chess. A human who is sat at a computer, and the computer has a representation of the board in front of both of them. And the human makes the moves according to what the computer says, but also gives the computer input and says, Hey, computer, what if we tried this? Hey, computer, I know you'd like to go this way. What if we tried that? And the human plus the computer for quite a long time was better than the computer.
Speaker (06:17):
And of course, also humans. So you can think of this almost like when we invented cars and humans used to run pretty fast, and then humans were getting beat by cars. But actually a human plus a car could drive faster than if you just you know, sort of pressed down on the accelerator and left the car, right? The human can steer around potholes, the human can provide very useful input that can get you from point A to point B even faster than the person plus the car. And that's what happened with chess and, and really conclusively, it wasn't just, you know, the computer's just a little bit behind, or the human's just a little bit better. When you put a human plus a computer together, you get much, much better results than either one alone. Now, that was true until a few years ago.
Speaker (07:00):
And here's the interesting, and somewhat disquieting further part of the story, which I'm not going to focus on, but I did want to tell you. And that is, now, when you get a human plus a computer against a computer, it doesn't matter. The human is not adding anything anymore because chess computers have advanced so far that in fact, the computer is the superior being in the world of chess. Whether you have a human stuck on the back of it or not, you might as well just ride off as a horse, you don't need the horse plus the human to make a Centaur. By the way, for anybody who doesn't know, a Centaur Greek mythology is a half worse, half human, and that's where the name came from in chess. So, I'm telling you this story about chess because it tells us an awful lot about pretty much every other area of human endeavor, which is behind chess.
Speaker (07:47):
And the reason for that is very simple. Chess is a super controlled environment. When we start putting computers in airplanes and cars and your desktop and lamps and microphones and all kinds of other things around in our world the world is much more complicated, right? It's not 64 squares. And every piece has a location and a precise way of moving the situation when we have that we have in chess is so much simpler that the computer is ahead of the game. And so we can look to chess to see what's going to happen to the rest of us, what's going to happen as the rest of artificial intelligence catches up to the complexity of the world. So I think that's a very useful way of thinking about all the different examples that you see of artificial intelligence being used.
Speaker (08:37):
Look for the Centaur. Where's the human in the loop? And only give you just a couple of examples, and then I'm going to come to some people who've made some comments in the chat. The one example is a very simple one. If you go to any modern Western supermarket, there's one option, which is to go scan the groceries yourself and put it in a bag yourself, and pay and go home. You don't need a cashier, but you notice there's actually a human who walks around at those stations. There's fewer humans. You don't need as many as you do to check everybody out individually, but there's somebody who comes around and says, oh, yes, I have to approve that, and I have to change that, and know this machine is stuck and so on. And so, you need a human to keep the machines going.
Speaker (09:14):
That's a very simple example. A much more sophisticated example comes from one of my clients. And what they do is they help you when you are building a skyscraper. So this comes from the world of construction. And one of the things you want to do in your skyscrapers, make sure you do things consistently that for example, you make sure that you put up the ceiling before you put the fire alarms on it. If you put the fire alarm first, it's underneath the ceiling and it can't detect the fire. So that's, that's not so good. But the problem is, you're might be putting up thousands of apartment blocks, apartment rooms. And each of these in this high rise needs to have a smoke alarm. Each one needs to have a certain number of electrical plugs and so on.
Speaker (09:52):
There are certain rules. And when you see a lot of them, it's easy to forget some. If you're doing this over and over again, as you might be as an electrician, say, in constructing such a building, but they said, oh, this isn't a problem. Don't worry. What we're going to use do is we're going to have somebody just walk around with a video camera and film the entire building once a week, and then we'll have an artificial intelligence program that will go in and check all the different video and make sure that every single room has all of the correct elements. And then we kick it out to a human. Here's the Centaur, right? So there's a Centaur, a little bit of a Centaur. Somebody gathers the data by walking around with a camera. That's kind of dumb. There's no real additional brain power other than navigating the building.
Speaker (10:32):
But where you really add value is when somebody gets the anomalous photograph. This doesn't look quite right. I think this smoke alarm is not on the right part of the ceiling. The, the AI says that, and then a human does the final check and says, yep, actually we have a problem in the 38 B, get somebody over there and fix the smoke alarm. So that's a, a much more sophisticated example. We'll come across some others as we talk today. But what you want to look for, and the challenge I have for you is every time you're thinking about some application of ai, every time you see a startup that's trying it, you see somebody boasting about how great it is for their brother-in-law, ask yourself, where's the Centaur? And if the answer is as say, for Tesla, oh yes, we have full self-driving, everything's perfectly fine.
Speaker (11:16):
You don't need a human go ahead and go to sleep in the backseat. You should not believe them, and maybe you should sell that share short, right? That is not a company to invest in, because they believe that somehow they don't need the human in the loop. And we are not at that stage of evolution. We'll get there if we follow the path that we saw for chess computers, but we're, we're a long way from it. Yet it's certainly in places like architecture and driving and places where you have to deal with a very messy world, not the very simple world of chess. Okay? So that's my first topic for the day. Let me deal with some questions. By the way, please jump in with questions. It's my favorite thing. I wish I could do this interactively.
Speaker (11:54):
I'm sure we'll get that eventually. Maybe I'll go on Twitch. No, you can't even do it there. Can you, we need something so you guys can talk to me. But please, without being able to talk, stick questions in the chat, argue with me bring things up. Those of you who know me know that that's going to work a lot better. We'll have a better interaction. Steven says he's built his AI ml ecosystem, and now he's coming up with the use cases. Wait a minute, Steven, when we put it that way, it sounds like it might be backwards. I'm a little worried here. Maybe we should come up with the use cases before we come up with what we're going to do with it. People are not yet thinking about the art of the possible. Yeah, I'm not so sure.
Speaker (12:29):
I'm interested in possible, I'm interested in what people want. So Steven, we might have an interesting debate about that one. Roland good to see you as well, curious on the podcast about chat,GPT and using it to practice conversations. One of my favorite topics. I'm definitely going to talk about that as we go here, Roland, if for any reason I don't, you stop me and make sure I do I've not yet had a chance to try it, but it feels like it might be easier than finding a willing person. Well, I think that's fascinating, and there's an awful lot to say about that topic. Let me see if I can go there now. Yes, I would like to, so I can, I can shift things a little bit because Roland's brought up this very important point. For those of you who don't listen to my podcast, I do that with Jeffrey Frederick, my co-author. We talk about conversations and how they can improve your tech team. And we did a little experiment. We said, let's use this chatGPT thing, and let's see if we can't use it to learn more about conversations.
Speaker (13:20):
Because one of the main things I coach people on and Steven knows this for example is, is how to have better difficult conversations. You might want to talk with somebody about, for instance, using Steven's example, how maybe it would be a good idea to figure out what we're going to do with the technology before we build it. And maybe we might want to change that. That's a difficult conversation, especially for somebody who's raring ahead to, to go try something fancy and new. And when you have that difficult conversation, it's super helpful to practice it ahead of time and also to practice it after, like, it went badly. And you say, man, I wonder how I could have done that better. And practicing is good. Well, it turns out that these artificial intelligence chat tools and, and being as the most prominent recent one just before that we had chat, GPT, and there's more in the pipeline chat.
Speaker (14:08):
GPT just announced their API. So now you're going to be able to build your own little chat bot and pay open AI the company for having a chatbot powered by this on your website. So using any of these to practice a conversation turns out to work pretty well. And there's a particular reason for this as a way of thinking about it, which I find really fascinating, I think is just really helpful. And that is to think of whoever is talking to you whatever chat material it is I'm just going to say chat, GPT, because that's the most common one. When that entity is speaking to you think of it as if that entity was an actor. And the tricky part to think about is let's imagine that this was Tom Cruise and Tom Cruise is playing.
Speaker (14:56):
There's somebody maybe you guys will can remind me, there's a maverick. Is that the character in Top Gun? I'll pick that one because I think probably people know it. I haven't seen the movie, but Tom Cruise plays a person who's a pilot. Tom Cruise is not a pilot, as far as I know, doesn't fly fighter jets, but Tom Cruise plays one. And so if you were to ask Tom Cruise to be maverick for a little while, then you might have a conversation with Maverick, the character, but not with Tom Cruise himself. You might then say, well, it's not being maverick for a minute, Tom, can I talk to you about, about you Tom Cruise? And you'd have a different conversation. A good actor might also switch accents, might switch body posture, might move differently.
Speaker (15:34):
Certainly bring up different topics and would act in a way that matches the character and not the actor. Well, the interesting thing about these chat tools is there is no actor. There's nothing behind there. All the thing is doing is figuring out what the next word is that makes sense in according to the context it knows about. There, there's no conscious action so far as we can tell, but you can think of the entity on the other side of the screen as being somebody, somebody is sort of not there, sort of a ghost playing a character, and they can play the character really well. So I saw one just earlier this morning. Somebody was said, well, would you pretend to be Donald Trump, please? And the machine, of course produced word salad, right? Completely confusing, rambling sort of stuff.
Speaker (16:19):
You know, the kind of thing Donald Trump says all the time that that's kind of slightly unhinged and kind of wild. And, and it matched his characteristics pretty well. It wasn't perfect, but it was good enough that if, say, I was going to, if I was a world leader and I wanted to negotiate with Donald Trump, if Kevin fend, he gets back in you might want to practice with this, right? You might want to practice. If you were going to be his debate opponent, maybe Joe Biden should use this. But more seriously, if you have a difficult conversation coming up, I have a client I was just talking to earlier this afternoon, who's going to be going into a high pressure sales situation? Well, it sure be useful if he could say to chat, GPT Hey, would you pretend to be this tough skeptical buyer who, who's not sure what the connection is between what I'm going to do and, and the value he's going to get?
Speaker (17:06):
And what would it sound like? Well, the person might say things like, well, you seem overpriced for the market. I think I could get this more cheaply. Or I, I'm I'm not sure that you're the right person for the job. I think we ought to do this internally. Those are the kinds of objections that a salesperson might hear. And boy, it is helpful to have something, an entity on the other side of the screen responding as if it were the person you were going to be conversing with. So Roland, I can talk about this for hours. I may come back to it later, but it's one of my favorite new techniques. I haven't quite got my head round exactly how to have my my clients use it yet. And I'm interested to hear if anybody is using it this way.
Speaker (17:47):
I put it on the pole specifically because I was wondering if anybody had tried it. I've only tried it a as a toy, but I think it could be extremely powerful because the machine is such a good actor, there's nothing behind it, and it'll make up the details, right? And this is one of the things people object to. They say, oh, man, you know, I asked, GPT question, and it made up the answer, right? There's a famous one, Bing was trying to convince the person talking to it that the year was 2022. Of course, if anybody's noticed, it's 2023. But Bing was just making up that detail and, and defended it very strongly. More common kinds of things are, he'll ask it for some citations about cold fusion or, you know, some black holes or some topic that it doesn't really know very much about. And it'll make up a whole bunch of our academic papers and professors who've written things and so on. It'll make up an answer. But if you think about it, that's what Tom Cruise would do. If you said, Hey, Tom Cruise you know, what high school did Maverick go to? He'd, he'd make one up. He'd say, oh, yes. Well, if he's answering as Tom as Maverick, he would say, well, yes, I went to this high school, and you know, this is the cheerleader I dated in senior year, and this is the class I flunked and so on. It'd give you a lot of detail completely made up, because what he's trying to do is sound like Maverick. He's not trying to give you accurate information about Maverick who doesn't exist. Anyway. So that's how I encourage you to think about these chat tools, is they're, they're not sources of accurate information.
Speaker (19:10):
There might be a guide might give you interesting ideas, but they're actors, they're playing, they're almost like a method actor, right? Someone who just inhabits the character and performs as the character all the time. Sean Penn was famous on the scene of fast Times at Ridgemont High. He went around as the kind of surfer smoking pot dude the whole time. And then on the last day of filming, the last day of Philman, he went around and shook hands with everybody. He said, hi, I'm Sean Penn. Because he had been acting as the character all the way through. And so he imagining they'd never met him. Well, of course they had, they'd met his character. So this is one way of thinking about it that I find really helpful. And then of course, that helps you in giving a prompt to the to chat GPT I've got it up here, by the way.
Speaker (19:53):
We might try this if, if we want to, but you know, I was giving it a prompt. I was saying, be, be Alan Weiss, the the, the guy who gives advice about solo consulting. And I said, how would you respond to this? And, and how would you answer these questions? And it, it gave okay responses that it certainly was no replacement for the actual expert because it was making up its answers. But if I were going to practice something, if I wanted to understand how could I have this conversation better I think there's real potential there. I think there, there are really great things you can do with it. Steven says responding to my ribbing him for inventing inventing the solution before the pr understanding the problem in order to convince the business that there is value for our business.
Speaker (20:36):
We are developing some solutions that we can show results on to get the results. We built the ecosystem for it in a controlled, speculative way. I don't mind spikes and trying things out, but I always want to understand why I'm trying to, what I'm trying to accomplish first. So I may still give you a hard time for that, Steven, right? Let me give you another application of these tools. So this is not the chat version. This is the sort of text generation version that people were using back in the summer. Remember the summer, you know, long ago when the stone age of artificial intelligence, it seems like it's moving very quickly here. So I did due diligence technical due diligence for a client whose business is really fascinating.
Speaker (21:19):
And it was amazing that they, they're able to do this with such profitable results. That what they do is they take manuals for appliances. So the easiest thing to think of is like a refrigerator or an oven, but these also might be industrial tools. So things used in a factory. And of course, all of these come with these complicated technical manuals, and they have schematics and diagrams and differential diagnosis tools and other things. There's the things that are very useful, but to sort of the ordinary human, they're quite abstruse, they're hard to find. And if you've ever looked through one of these, trying to figure out what error seven means on your, on your microwave oven you'll have that idea what I'm talking about. This, this can be tough. And so these folks said, well, it'd be really great if the manufacturer had a chatbot on its home webpage that said, Hey, just ask me your question about your microwave oven or the industrial paint robot, or whatever it is that you, you want to know about, and I'll look up the answer for you.
Speaker (22:20):
And so what they do is they takeGPT, the original tool that does the, the text. They're not so worried about using it in a chatty kind of way but they want to get sort of through a tree of steps. So they use the machine to go through tons and tons and tons of manuals and come up with a series of questions which sort of form a chain. So you can go, well which brand of microwave oven do you have? Is it silver or is it black? It does the bottle number start with q is the error flashing or steady you know, a series of questions. And as you answer each one, it knows where to go to the next question.
Speaker (23:03):
It generates this tree of questions. And then of course, at the end it says well try turning it off and on again, or whatever the, the correct solution is. You know, don't change the, the sprinkled grommet or, or whatever you're supposed to do. So the fascinating thing is, and here's the Centaur, right? So practice, what are you supposed to look for when you hear about ai? I'm telling you about it. Where's the Centaur? The Centaur is humans in customer service who are busy answering these kinds of questions because if you don't have the chatbot, what do you do? Well, phone the support number or whatever, and you say, Hey, you know, I got error 17. The, the paint robot will only paint white. What do I do? All my food is burnt. How should I progress?
Speaker (23:43):
So the the tool produces this chain of reasoning, all these questions. And the human who's been answering lots of these and knows what the right answers are and, and what the question should be goes through and says, oh yeah, we forgot step seven, and we need to tune this one. We need to update that. So we're not trusting the machine to just have conversations with humans. It might make up the answers, right? Remember, it's a method actor, so we don't want to trust it to come up with things. So where it goes off piece and starts saying, well, I don't know. So I'm just going to ask, you know what kind of what are you wearing? What, what shirt color shirt are you wearing? Not a relevant question for diagnosing your oven, but it might ask and we can say, oh, yeah, we don't, we don't want that one.
Speaker (24:24):
Or, or if it gets confused between different models or in some way and makes another error, the human corrects it. But just like the architect example, it's much, much easier for the computer to go to go through the many, many, many examples. All the huge number of manuals. Just like it's easy for the, the computer to go through the video of all the different apartments and then to kick out the ones that look important. And for a human to go through and say, okay, we need to adjust this. This is the one we need to update. And the combination of using the computer for the volume and the human for the creativity is just like chess, right? So we're back to chess again, where chess has, you know, many, many paths you can follow and many, many different moves that you can try and so on.
Speaker (25:11):
The computer sifts through all of those and says, Hey, human, which one of these looks interesting? The human says, oh, applying some creativity. What could we do with that one? That Centaur approach, I think is the best one that I've seen so far. We're going to hit the stage where computers can do these things themselves. We're definitely not there yet. Right, so we've got Stephen coming back again. Great. Good dialogue. Steven. Consumer now prefers solutions that are powered by AI ML result, regardless of whether, whether the results are better than competitors. Buzzword bingo. However, it's reality in software development. I completely agree. And the unfortunate thing I'm seeing now is I've got folks who,it's certainly not fraudulent, but they're lying to themselves and they'll come to an investor who then comes to me to do technical due diligence and say Hey, I've got this wonderful new AI powered software.
Speaker (26:06):
It does these amazing things and it does do amazing things. It's just not got any AI in it. It's just a bunch of rules. And the computer follows the rules really well and produces a result. And that's fantastic. It's just not any of these generative machine learning high high volume data things. It just does a really good job. I remember one in the FinTech area in FinTech, they were somebody who was doing lots and lots of trading. They were giving tools to traders who, who were processing lots of data. And the main thing they did is they took the messages from the market, which were very messy and in lots of different formats, and very hard to understand. And the traders, of course, had developed many years of experience in understanding these messages. But they, they were limited and how fast they could react as they were humans.
Speaker (26:48):
And, and they just cleaned up all the data, then they put it all in a very standard format and then applied rules to it and so on. And they said well, it's artificial intelligence. And I said, great. Show me the machine learning. They said, oh yeah, we didn't do that. We're getting to that. We're not there yet. But what we did is we cleaned up the data that my advice was forget the machine learning. You don't need it. Keep cleaning up the data cuz the data is so messy that actually you pr providing tremendous DA value in simply cleaning it up and, and providing something valuable that way. So Steven, you're exactly right. There's lots and lots of fashions in software. You wouldn't think so, but just like people change what style shirt they wear people like to change and say, okay, it's AI.
Speaker (27:26):
You know, a couple of years ago it was crypto and everything was a crypto cryptocurrency startup or tool or idea or, or method. We were all going to use blockchain to keep track of what was in our sock drawer. So this sort of thing comes and goes, but there's real value here as there is in cryptocurrency. I go back to the livestream I did on that almost a year ago now there's real value in these things. It's just often not where people think it is and not where people cite it. So very good contribution there, Stephen. The capabilities of valuation multiplier, whether or not it actually adds value. Exactly.
Speaker (28:03):
Good business to get into if you don't mind a bit of a crash at some point when people discover it's not actually doing what you think it is. So let me comment on, on just a couple other elements. What do you do with this? So now we have the idea of Centaur. We know that if we're going to be doing artificial intelligence, we want humans in the loop. We know that the computer we should think of when it's chatting with us as a method actor. But what does all that mean? For instance, does this mean that Steven should go out and fire a bunch of programmers and start using GitHub co-pilot, which can help write computer programs? Maybe we should use that instead of some of our programmers.
Speaker (28:45):
Hint, no, if you've been listening, the point is you want more skilled humans to be the center, the human, half of the Centaur. So in fact what we're not going to have I believe is a wholesale replacement of humans by computers that may come in the future. But what we're going to have, and again, the analogy to cars is pretty close. You know, we didn't eliminate humans from the car equation. We just put humans in cars and they could move faster. The same is going to happen with things like writing business letters. So when you want to write a routine email, and Gmail, by the way, already does this, I think Outlook does too. It'll start to complete more of the words for you right now. It just says you know, if I say it was good to, it'll say, see you today.
Speaker (29:29):
And then if I say, see you yesterday, it'll, you know, finish the word yesterday for me. That's as far as it'll do it today. But soon you'll be able to sort of get a blanket thank you email. It'll know who the person is you're writing to. It'll have a little bit of the context. It'll get you started. And that'll save a lot of time for people who write a lot of thank you emails. So yeah, but they will always be correcting, right? They'll always be updating the information. There's always going to be that human in the loop. So we're going to have greater productivity in certain areas, especially routine text routine answering of questions about ovens, for example, taking my example from before. But where it's really going to make a bigger difference is in operational processes.
Speaker (30:12):
So you can imagine that the way that people are going to look at things like fraud detection or security analysis or things like that may change substantially. Again, because the computer's going to help you, it's going to do that first check. Again, like the apartment example. It's going to look through many, many pieces of data that you hand it in an undifferentiated form. And it's going to be able to say, well, here are the anomalies, here are the important areas here. Let me give you a summary of what I'm seeing here. I'm minded of I just thought of this example. I was a CTO at an e-commerce startup, and we had 4,000 new products every week. What we did was flash sales.
Speaker (30:56):
So we would get new stuff in, it would come into the warehouse sometimes we had no idea what it was, and then people would sell it, get it to us, and we'd get an invoice for it. And it wasn't what we ordered and we didn't know what it was. And we had an army of humans and, and there they were called the processors. Nobody knew what they did. But we finally went over and sat next to them for a day and watched it. What they were doing was taking the invoice and comparing it very, very closely against the actual item. And they'd say, ah, this is a pair of green shoes with buckles. And we ordered purple shoes with bows. So either we need to send it back or find a way to sell it. Often we would just sell it because we, we wanted to move the product.
Speaker (31:34):
But those invoices, of course, came in illegibly. They were faxed to us. They were printed sideways on weird pieces of paper. They were in foreign languages. It was a complete mess. And, and we did look at, in 2012, whether we could do any automatic processing of this. We couldn't find a way to do it. I tried to use a thing called Mechanical Turk where at that time you think Amazon's still running it, you, you can get an army of sort of anonymous humans to go to a task for you. I've tried that, didn't get great results. I suspect you'd get really good results. And I suspect people are using this these tools this way right now to streamline that kind of process and reduce a lot of the errors. You're not going to get rid of the people who are experts in understanding invoices and warehousing and what you ordered for of the shoes.
Speaker (32:18):
But man, it would certainly be easier if you got a summary that said, we received 50 invoices, 45 of them will look normal. Here are the five anomalous ones. And this one looks like you know, it's it's cost us twice as much as we as we expected, do that one first. That's going to make a huge difference to all of you who do operational workflow. And I'm going to talk you know, I'm always talking about the operations teams. They're one of my favorite friends of the technology team. The operational workflow is one of the places where engineers can make a huge, huge difference. And, and here's a tool that, that can really shift that. If you have questions about that, ask me, because I'd be happy to say more.
Speaker (32:58):
And the last place, and Steven alluded to this, I is in strategy, right? So how is this going to upend your marketplace and your, your competitors? As Steven says, maybe just trying to throw AI at everything and see what's happening. You can be smarter. And when you're thinking about the strategy, I do this, you know, strategy work with with companies all the time to try to help them figure out where the market's going, what their advantages are your advantage can be. For example, just as the examples I've been listing, where are your expert humans who can ride on the Centaur, who can be the human half and provide huge value on top of chatGPT G P T itself? Bing whatever tool you're going to use, stable diffusion. We haven't even talked about the ones that can draw for you.
Speaker (33:44):
And maybe if you have artists or if you have architects or people who are producing visual outputs, can you shorten their process? Can you give them better tools again so that they can filter through much more a greater quantity and with greater skill? How could you put those humans in the loop to get there? And we've had a perfect person appear in the chat. Who knows all about this, which is Eleanor. Hi, Eleanor, good to see you always have a person in the loop. Why would we ask Eleanor and I know this because Eleanor's in computer vision and he's doing really exciting things with ai. And I believe, Eleanor, tell me if I'm wrong, you don't have a human in the loop. So this is very interesting. I'm very glad you're here.
Speaker (34:23):
All the signs are that we are accelerating to get to a state of no humans in the loop and more and more scenarios. Eleanor, not sure if you were here at the beginning. If not, go back and, and listen, I did a little story from chess where you're absolutely right, we're moving toward that world. And you may be at the front of it. I want to say more about your world in just a second. But the kind of overall trend is yes, absolutely we're going toward no humans in the loop, but we're not that far yet. And for a long time in chess, the situation was that the human plus the computer could beat the, the human and the computer. And that's true in most areas. Now, the thing that Eleanor does, I'm sure he won't mind me mentioning it, is detect emotions.
Speaker (35:03):
So he can tell whether I'm really happy or very sad or if I'm confused or something else. He can see that on the screen, which is tremendously valuable in learning situations in interview situations, lots and lots of different situations. It's very helpful to be able to get a signal about whether the person at the other end of the screen is what they're feeling, what are their emotions? Really, really interesting and very valuable. I suspect elna that there are cases that in your world where you have automated responses, I know one of the areas you do is education, I think, so that you're you're trying to help somebody learn and you want to know whether they're feeling confused or happy or frustrated by the material or something else there, you can have a kind of instant no human in the loop reaction.
Speaker (35:51):
But I suspect there are lots of cases where you could get even more from, say, a customer service interaction or a classroom virtual classroom or, or any other circumstance where you have a group together and the people with the strongest emotions or the strongest reactions or the most interest or something like that. The computer can pick those out. And that seems like that would be really useful. Then the human can react, right? The teacher of the class can say, oh, the person in the back there looks really unhappy. Let me see if I can help them. And it might be hard to pick up for the human themselves, but with a computer assisting saying, oh, look back in the back of the room on the right I would think that would be, that would be better. Eleanor, I'm very interested in what you think.
Speaker (36:31):
I don't know if you're able to post back here, but if you can I'd be happy to hear what you think about whether, not just in your world, but in computer vision generally. Are we at the point where we can completely rely on the computer? Would, would we want to act, say, in a in a mental health facility on somebody's emotions based on what your computer vision can do? That would be great. I'd be really interested if you think it's that far ahead. My suspicion is that what we want is we want to flag up, Hey, nurse, this person over here looks very down in the dumps. We need to get some intervention for them to help them out where it's life and death, where it's got medical implications bigger than, say in a class where you can recover quickly. I suspect you would want a human in the loop, but maybe you see that differently.
Speaker (37:18):
Okay. I'm not sure Eleanor's responding or maybe he's busy frantically typing, or he's thrown his computer out the window in frustration. I'm not sure, but Eleanor's my old friend doing amazing stuff. So I'm going to bring us to a conclusion here. To summarize what we talked about was several different ideas how computers got used in chess to start with and, and how important it was that humans could add a lot to the computer and, and give you much better results. And then we talked about how computer sort of on the other side of the screen, on the other side of the chat conversation is like a method actor, somebody who performs a role and can be really great for helping you to practice with someone who's performing that role or to produce sample output or in some other way to pretend to be a person.
Speaker (38:04):
It's not a person and of course the data it has is not accurate, and that's the important thing to remember. It can fool you quite well, It can make you believe that it's really a human that you're really talking to, to someone. But that's just because it's a very good actor. Eleanor come back. He says, I'm all for human in the loop. I'm actually worried about more and more scenarios with no human in the loop. I agree with you, Eleanor. And classic example of course is Tesla cars driving themselves into police vehicles because they can't tell the difference between a police vehicle stopped on a motorway and a crossing or something else that you can drive through or can't detect roadworks or these other kinds of problems. So we've gone ahead too quickly and I'm always very nervous when I drive on American roads about whether the car coming toward me has actually got a human in it.
Speaker (38:51):
And that there isn't a problem that I think I, I was very conscious of until just a couple of years ago. So to finish the summary we have a number of examples from customer service, from construction, from e-commerce, all over the place where people are using computers and AI in this very sophisticated Centaur human plus computer way. And I think that's really the future. That's where I see the most value. When I do technical due diligence, that's what I'm looking for to advise my investor and the person who's commissioning the report from me on where the real value is. Because the, the computer by itself makes a lot of mistakes. We're not advanced enough yet to, to get it out of the loop. It still crashes the car, but man can add a huge amount when you have a large volume where you have something that can be repetitively managed and then you can kick out the exceptions to the humans.
Speaker (39:42):
So I hope that was a helpful view. I had fun thinking through these issues with you. Really appreciate the comments and the thoughts and the chat. I will mention again that I'm live in London next week, so less than a week away. I'll see many of you in person. So we're looking forward to that. We're going to be talking about helping your tech team to deliver every single day. How can you get a new result, new features, new value every single day? So you'll make the maximum profit from your tech team. And then I have many other events coming up. Just head to let me get that up here on the screen. Squirrel squadron.com is the place to go. We have events coming up talking about mob programming, about which is a way of getting all your engineers to write programs on the same computer at the same time.
Speaker (40:26):
One computer, many programmers. We're going to be talking about recruitment and how to improve your recruitment very, very quickly. How to use an internal recruiter, which is an idea that's getting more and more mainstream now. But we're going to have a, a world expert on this Stevie Buckley best recruiter of this kind in London that I know and lots of other good conversations coming up soon. So looking forward to seeing you at more events like this. And I'm very glad to see all of you here. This recording will be available both here and on the Squirrel Squadron Forum, where there'll be more discussion. One thing I want to do, by the way, just so you know, is I'll be sharing a very good article on this kind of method acting idea. So it's not my idea, but it's a very, very valuable one I've, I've been noodling on for ever since I read the article. So, head on over to the forum to read more on this interesting topic of computers as actors and computers as Centaur. Okay, wonderful to see you see many of you next week in London Live on Wednesday. Wednesday the eighth. It's not Thursday, it's Wednesday. Yeah. And if you haven't signed up for that, right there on the screen, squirrel squadron.com. Thanks so much. Have a wonderful day, everybody. Take care.