AI and Everything Else

[Music] so oh I see I'm a chairman I've been promoted possibly um Global vice president as well um so thank you for that introduction um and I'm obviously equally reflective on the loss of Charlie Munga who's always much clever and much pithier than me talking at events like this um but I was asked to come and explain uh Ai and everything else and um only to take about 20 minutes and not talk too fast um I can probably manage two of those although maybe not all three um but I thought it would be kind of useful to start with a quote from another um important influence in our lives Bill Gates who said this spring that in his life he had seen two technology that he thought was revolutionary the first was the graphical user interface um back at xero Park in the 1970s and the second was chat GPT and um you may have noticed that people have got quite excited about chat GPT these are headlines for the the main technology news site this is how excited people are about um Ai and how they excited they were about the iPhone um back um now almost 10 over 10 years ago um so this has become something that is the entire tech industry has become hugely excited about it's also now what pretty much every new startup is working on these are startups at y combinator the main incubator in Silicon Valley um today if you're not working on generative AI there's probably something wrong with you and people will will kind of go and and talk to somebody else um and even as the Venture industry in general is now in something of a crash as we come out of all of the excitement of the pandemic um all of the investment is now surging into into AI um and even outside of the tech industry there's now a huge amount of excitement um this is a survey from McKenzie um on how many Executives outside of the tech industry um have at least played with generative AI which mostly means chat GPT so again everybody's interested everybody's excited um a few years ago you may have heard big companies saying that they need needed to have a 5G strategy um now the answer is definitely you need an AI strategy um and open AI itself um despite the drama last week it seems to be doing pretty well 100 million people using chat GPT every week something over a billion dollars of Revenue which is kind of a record for something that didn't exist a year ago um and indeed of course we have a rush to regulate um as a global policy makers try and couch up with this um and of course try and get some headlines for themselves however we still have to kind of work out well what actually is this and what does it mean everybody in the tech industry has spent the last 12 months walking around so to speak holding on to the top of their head with both hands saying oh my God this is amazing long pause and then yes but what is it and so I think there's a sort of a base case view that says this is sort of like the iPhone or the Internet it's another of these platform shifts that roll through the tech industry every 10 or 15 years then you have the Bill Gates view that says no this is a rather more important change this is more like a once every 20 or 30 or 40e shift in what technology could be and then there's a minority of view that says no this will take us all the way to artificial general intelligence this will take us to how 9,000 and the Terminator this will take us something that will sort of fundamentally change the nature of the world and everyone is kind of trying to work out how we would fit this within those kinds of Frameworks so first I thought it would be useful just to sit and think about how we try and understand what generative AI might be and to do that I wanted to go back and talk about the last web of machine learning in 2013 um large langu um neural networks machine learning deep learning started working and you would have demos like this and I would show them to big companies and they would say Well done that's very clever why do we care you can recognize a dog or a chair and so it wasn't obvious that this was a scientific breakthrough that would have much broader consequences and we went through a period of kind of trying to work out what are the right levels of abstraction to understand machine learning so it's not just image recognition it's not just a whole other class of Unsolved problem like langu translation natural language processing speech recognition really the way to understand this was that this was patent recognition and then as a big company you had to work out well what is it that we could turn into patent recognition and there were obvious things like fraud detection or credit scoring but then there were also a lot of other things that didn't look like patent recognition that you understood you you slowly realized you could turn into pattern recognition and then you could automate them with machine learning um we now go through I think the same process with the second wave of machine learning with large language models again you get the amazing demo so I use mid journey to make a picture of cats this is very cool it's also not obvious why this is useful I can also ask chat gbt to write me a song about Tesla and SpaceX and again it's it works it's a song it Rhymes it scans it's not obvious what this is good for um you push a little bit further supposing you're going on holiday somewhere next week and you want to know what to do you could go to Google and open 30 tabs in Chrome or you could just go and ask Jack GPT what should I do that week which gets you this phrase generative search and so that starts getting a little bit more interesting and starting to challenge some kind of hundred billion dollar Revenue models um and then you can push it a little bit further and say can you maybe explain something what happens if you can explain a joke or explain a mass problem and so this starts to look like maybe this is doing some reasoning which seems like it might be more interesting again and are there people pap stop and think okay where exactly is this going to go how far will this take us is this patent creation or is this something else however you also have to be a little bit careful so imagine for some bizarre reason you actually wanted to know something about me and so you went to chat gbt and you asked it and it would give you this biography and it says that Benedict Evans is a hugely important and influential person which is obviously true um and that I went to Cambridge and I worked at a couple of different companies worked at Andre and hor this is all roughly right hit reload oops no I went to a different University different degree different jobs um different companies reload oh no back to Cambridge again went to work for no I didn't work for the guardian I think I applied for a job at Atlas Ventures but I certainly didn't get hired there and so what is it that this is doing well these all look like biographies of someone like me but they aren't biographies of me so what this is doing is it's not answering the question it's matching a pattern um this is a great example from earlier this year somebody asked Jack GPT what's the most cited economics paper of all time the paper that it describes here doesn't actually exist but it's what the paper would be likely to be so what kinds of papers by what kinds of authors about what kinds of subjects so it's matching a pattern I think a really good way to understand what's Happening Here is this is I asked mid journey to make a fantasy 1960s FR sports car this looks like a sports car it looks French it's got kind of a citon Vibe it also has two steering wheels and no door and that's not lying exactly it doesn't know what steering wheels are it knows that images like that have a shape like that in roughly that place this can be kind of deceptive um you've probably seen some news stories about people relying on these systems and getting them wrong this is a great survey that Dee did people who've used these things tend to think that they're accurate because they haven't really thought about what they're doing or tested them properly they're kind of misleading and so everyone in is kind of trying to work out how would you think about this the the most interesting framing I heard is that this tends to be bad at things that computers are good at and good at things that computers tend to be bad at so it's not predictable it doesn't do logic um it's not a database it doesn't do kind of reliable information retrieval on the other hand it's very good at a bunch of things that you could never use computers to do before um and so if we come back to that framing like what's the right level of abstraction to understand this you can start by saying maybe this is patent creation and that seems like a kind of a good technical description of what's happening and then you could push a little further and say well maybe this is doing synthesis or summary or search or ex explanation it's a way of retrieving information and describing it um and then you can push further again and say maybe what this is doing is doing some kind of reasoning or explanation or understanding maybe it's some kind of engine for reasoning um the framing that I always like talking about the last wave of machine learning was that it gives you infinite interns so you would like to listen to every coming into your call center and tell me if the customer sounds nervous or if the service agent is being rude and you don't need an expert to that for that you could probably get a 10-year-old to do that but you don't have enough interns um machine learning gives you infinite interns anything you could get an intern to do now you could automate and that seems like kind of a useful first wave of thinking about what you might be able to to do with generative AI um the how is it that you you think about things that you could automate things that match a pattern things where you can align with that error rate that question of probability and so the kind of the first two things where people are doing doing a lot is on the one side um using it as a coding assistant um do create effectively creating Auto suggest and generating code things where everything has to be exactly right but you can see the mistakes and on the other side things where there aren't really mistakes things where there's no such thing as a wrong answer just good or bad answers and so that gets you things like brainstorming creation lists marketing um marketing copy and things like that um and so that as a sort of a base case of thinking about what you can do with this fits quite nicely into the classic framing of Technology as moving forward in a series of s-curves a series of platform shifts and so in the last 50 or 60 years we had mainframes and PCs and the web and smartphones and each of these took 10 or 15 years um and then there was another platform shift after that and so the base case is to say that generative machine learning kind of fits into that framing it's the next platform shift and as you go through these platform shifts you kind of get engineering questions and you get value capture questions and you get kind of questions for outside of the tech industry is what will change once you've deployed this and those obviously kind of intersect and reinforce each other and so we have engineering questions right now um gold rush to get GPU capacity from Nvidia um to get compute compute capacity to run these models you also get a gold rush to try and create more and more models and to get them to catch up with the performance of chat GPT and there's a whole question about how many models there will be and who will have them and whether they'll be open or closed um but for the rest of this um the rest of this this looks like product and so you ask well what sort of products come out of this um and the classic framing with each of these platform shifts is that the incumbents always try and make the new thing a feature and so Google and Microsoft today are rapidly um implementing large language models as part of their Office Products as part of Microsoft Office as part of Google docs then you have startups that try and use it to un bundle that try and kind of pull apart some piece of the income and business and then every now then um people find a way to actually change the entire nature of the market so today every L every text box on the Internet is getting an llm this is LinkedIn because obviously LinkedIn doesn't have enough predictable generic content spread all over it so we need AI to make more of it um slightly more interesting Amazon has already started deploying image generators to help Merchants create product shots so you can customize your product shot create another product shot this is a perfect example of AI is giving you interns this is exactly what you would get the intern to do now you can get AI to do it um but I think it's more interesting to push a bit further and say well how is it that the tool actually changes what we do because with every new tool to begin with you make the tool fit the way you're already doing the job and then over time you change the job in order to fit the new tool this is the process we went through with PCS and with the web and with smartphones um with cloud with the last wave of machine learning it should be the process that we go through with this this is a still from a movie called the apartment from 1960 Jack Lemon in the middle in the second row is a clerk in an insurance company in Midtown Manhattan everybody you see here has a typewriter and a Rolodex and an electromechanical adding machine made by our company called friden basically everybody you see here is a cell in a spreadsheet and the whole building is a giant Excel file once a week someone of the top floor presses F9 and the whole building kind of recalculates from top to bottom and generates new insurance prices um 1965 they bought a Mainframe they threw away all of those adding machines they automated all those jobs but somehow they employed just as many people but they changed how the insurance business worked to begin with they use the Mainframe to automate what Jack Lemon was doing then over time you kind of change how you might run the business or what kind of things you might be able to do with that this is kind of the process that we're going through now but if I go back to talk to that quote from Bill Gates there's kind of another theory here which is to say no this isn't like the web or like smartphones or the PC this is more like a fundamental shift in what you can do with computers so what he's saying about the guoy here is that when you what the guey did was it meant that you changed who could use computers because you didn't have to memorize keyboard commands and type them in instead you could just click on what you wanted and so that was a fundamental change in who could use computers um but then somebody has to have made the thing you're clicking on someone has to have made that tool and so if you want a piece of software to do a thing you need people to click on it what um what chat GPT does is provide you all of those tools without you having to create them one by one so there is this possibility that this is a fundamental change in the levels of abstraction that with guies you have to with command line you have to learn the commands with the guey you see the options and click on them with generative AA maybe you can just tell the computer what you want at least that's what Bill Gates is sort of proposing um I'm not entirely sure this is true when I look at chat GPT today I'm not sure this is a product so much as a technology demo and in fact it kind of reminds me of excel you open Excel and you think what is it that I'm supposed to do and Microsoft gives you all of these suggestions but then all of those suggestions have been unbundled into companies and at a certain point you don't use Excel anymore you use d a dedicated piece of software in fact kind of the classic quote here I think is from Jim bdale that there are two ways you can make money you can bundle or you can unbundle what seems to be happening now is that open AI are bundling everything else into chat GPT and every other tech company is working out how to unbundle chat GPT now I'm just kind of wrap up in one or two minutes more before I stop talking um if we zoom out from Ai and think about everything else that's happening it seems to me there's a sort of a sequence in how technology hits the rest of the world the tech industry spends a lot of time talking about what will happen in 2025 or 2030 um nobody in Tech wants to talk about crypto or metaverse today but they're actually still kind of interesting important Technologies now everyone wants to talk about generative AI meanwhile all actual software companies are deploying ideas from 2010 SAS automation workflow collaboration machine learning and meanwhile most of the rest of the economy is being overturned by ideas from 2000 ideas like maybe people will buy things on the internet or watch video on the internet and so we have e-commerce continuing to accelerate upwards um we had this kind of great blip in the pandemic but is now almost exactly back on the trend line but back on a trend line at something over 20% um and I think we kind of forget how small the internet used to be and how much bigger it's become um when Apple launched the original Mac they only sold 200,000 units um when MOS kicked off the consumer internet there were maybe 50 to 60 million consumer PCS on the entire planet now something over 5 billion people have a smartphone and I think the best way to express what that means is this study from the US of online dating that 50% of new relationships in the USA now start online so technology has gone from being something that was always kind of exciting and interesting but not actually a basic part of everybody's life but just a kind of a basic part of how part of how we all live and how we all do business um so we have entirely new companies being created that just presume everybody is online she sold $30 billion of fast fashion last year overtaking Zara and H&M and that's just a company that presumes every teenage girl has a smartphone um Ika is now opening Flagship stores in the center of major cities in shifting away from their platform of being on the edge of town as they think about what it means that all of the customers are online and maybe don't want to go to the edge of a city um we now have um Insurgent Banks Neo Banks using smartphones as a route to Market to bypass Legacy Banks and using that to build a better user experience um this is showing many Taps it takes to open a bank account for Insurgent banks on the left as opposed to Legacy banks on the right um and as I look at those questions it seems to me that none of these questions are really software questions anymore what's going to happen to television I don't know ask a TV analyst what's going to happen to the car industry as the Chinese start exporting I don't know ask a car analyst these are all questions that come from the industries being changed by Tech these are no longer software questions and indeed I think cars are kind of a useful analogy for a lot of this stuff that you could say the first 50 years of the car industry was what's a car but the second 50 years was what happens when everybody has a car how does every other industry get changed that's kind of what's happening with software now the first 50 years was what's a computer the second 50 years is what happens now that everybody on Earth is online um with that I'll will say thank you