The Future of AI Is Amazing

okay so I'm going to be uh covering this AI stuff everybody's talking about so I'm just going to spend about 10 minutes to talk about why we're so excited about it and then um Senator Todd young is going to be up here and we're going to have a discussion so I probably took my first AI course in the late '90s the stuff has actually been around with us for a very long time uh and during that you know 70 years um It's You Know by every metric has been a huge success it's been up and to the right we've solved a number of problems we didn't think computers were good at solving so for example expert systems in the 50s and 60s we used for medical diagnosis we got very good at beating Russians at chess in the 80s and 90s we're good at image detection we're good at robotics like we've solved a lot of problems that originally we thought you know computers were just like large calculators on top of just solving these problems for decades a lot of the solutions are actually better than than than humans like we're better than humans at handwriting detection we're better than humans at at uh identifying objects and images and and with all of this magic we've actually been able to add a lot of value to large companies right every time you go to Google and you get a search this is using AI anytime you get some personalization this is AI right so this stuff is like magic right it's been around for a long time it's solved all these problems so there's been this huge conundrum in the investment community and the conundrum is the following if this stuff is so magic and it solves all of these problems why haven't we seen a platform shift in the same way we saw a shift with mobile or with the internet like why hasn't this happened and we've actually done a lot of research with this as a firm and the answer is is that even though the capabilities have been fantastic like I talked about the economics just haven't been there in the same way there's a number of reasons for this I won't be exhaustive but I'll just cover a few of them so one of them a lot of the solutions just tend to apply to Niche markets there's not a lot of broad Market appeal the second one is probably the most important in Nuance which is a lot of the use cases that we apply it to correctness is really important like robotics but getting something absolutely correct is very very hard and requires a tremendous amount of investment so a number of the solutions require hardware and finally you know the the competition for AI it's not it's not another computer it's actually a human brain and you know maybe it'll be better maybe it's not as good the human Rin is incredibly efficient and it's incredibly cheap and one of the best examples of this is is autonomous vehicles or Robo taxi so when I joined Stanford to do my PhD in 2003 Sebastian thrud had just won the DARPA Grand Challenge right so he had driven a van autonomously across the desert and won this and we were like great news exciting like autonomous vehicle is a solved problem back in 2003 now if we go 20 years later we've invested 75 billion dollars as an industry and while we do have autonomous vehicles on the road and they're great and they're solving real problems the unit economics are still worse than say Uber and lift because they're competing against the human brain so while this is very important technology to date it's really remained in the realm of large companies right that can sync these types of Investments so the AI learnings of the last couple of decades is not that technology can't be built or even that we can't monetize it we're actually good at all of that is that this is very hard for startups to build businesses around so the reason that we're so excited and the industry is changing so quickly is this wave is very very different on exactly this issue economics so when I talk about kind of this wave I'm talking about the emergence of what we call large models or Foundation models or state-of-the-art models these are pieces of software that you put in text or you put in an image and outc comes something out can come an image or text or a conversation right there're these kind of like very very smart pieces of software that you can ask questions and they provide answers too and they've already entered a number of problem domains that we just haven't cracked in computers and certainly in AI right the AI has not been able to do this before so for example creativity you know these models are better than humans at creating images or creating music or creating um you know voice imitations it actually turns out they're great at natural language reasoning as well right they're great conversationalists they're great friends they're great romantic Partners they're great therapists um and they also have now serving this a thing which we call co-pilot which is this catchall phrase that they're pretty good at like mean online tasks and by mean I mean average right so if it's something that you do a lot of it can kind of get the hang of it and do it as well now remember when I said like traditional AI the economics didn't work and there was a set of reasons so those reasons just don't apply to this set of tasks right like these markets are enormous like whatever video games and movies alone are like a half trillion dollar market in many of these use cases correctness just isn't an issue right I mean what like there's no formal notion of correctness of creating like a fantasy image or like creating like a sonnet or something like that the use cases are primarily software and and the last point is the one that like I couldn't have predicted and it's the most surprising that it turns out that for these tasks the one that we think of as very human like you know communication and social interaction and creativity the computers are far cheaper and far better than than than humans are I want to give you a very specific example it may be silly but it actually generalizes so let's just say that I Martine wanted to create a a picture of me as a a a Pixar character so if I had one of these AI models do it the actual inference cost the cost of doing that is about one/ 100th of a penny and it takes about a second and it and I mean this is what we did here and this is the quality that you get if you were to compare that to hiring a graphic artist a graphic artist is what let's say 100 bucks in an hour it's it actually is much more I've gone down this road before so the AI you know it's just not a little bit better it's not like 20% better it's for orders of magnitude cheaper and faster this isn't limited to images this is also the true for like any sort of language understanding so like take a complex legal document I can take a complex legal document I can feed it into an llm and I can ask questions if you compare the analog would be me like whatever like working with my my lawyer so like the lawyer would have to read it would have to understand it you know I don't know how much you know you know the average lawyer cost is but like it's you know 500 tends to be pretty standard so again to use an llm is four to five orders of actitude cheaper and faster and it's exactly because of this that we as Venture investors and we on the private Market side are so exciting because we're seeing the fastest growing companies we've seen in the history of the internet including by the way the internet itself and this is by measured for revenue or the number of users Etc right it's exactly economic dislocations that create new startups not just new technology so if you take a step back historically when marginal costs have dropped this much this is what creates platform shifts and has changed the industry entirely it's happened twice you know that pretty concretely I want to walk through both of those so the first one was compute so in the creation of the microchip it brought the marginal cost of compute to zero like so before you had the microchip calculations were done by hand right so it was like people doing logarithm tables you know in in large rooms and then ANC was introduced which was forwarders of magnitude faster and then you had the computer Revolution and here came you know IBM and HP and everything else the internet brought down the marginal cost of distribution to zero right so before you like whatever You' send you know a box or you'd send a letter and then the price per bit dropped and you could send it over the Internet by the way OS forward is a magnitude Improvement and this ushered in kind of the internet Revolution right this is kind of Amazon and Google and Salesforce so it's pretty clear if you just take the fundamental economic analysis that these large models bring the marginal cost of creation to zero like creating that image and language understanding like reasoning over those documents and these are very very broad areas that they can be applied to now whenever we talk about economics we always kind of talk about Job dislocation as well it's very very important especially with an economic dislocation of this size we can learn from the last two Epoch both the microchip and the internet in that if demand is elastic so for example the demand for compute seems kind of unlimited and the demand for distribution seems kind of un limited that even though the costs drop the total throughput the total use increases by a lot because it becomes more accessible and so rather than removing drops or removing value these tends to EXP expand growth like the internet almost certainly expanded growth uh in the United States and so we think the same thing is going to happen here as well so get ready for a new wave of iconic companies it's almost certainly going to happen it's not just the technology which is solving problems that have never been solved before but the economic case is absolutely there um you know when this happened with the internet like we didn't really know what was going to be on the other side of it like we couldn't have predicted Google and we couldn't have predicted Yahoo but we knew it was going to be something and it's kind of one of those moments but we have some glimpses right we know like the social order is changing we know like this is a very real use case that's monetizing today and people are using we absolutely think that the creativity itself is going to change you know productivity these kind of mean workers like this is happening as well and if you really want some prognostication for this is all going I mean for the first time I say there is actually real line of sight to embodied AGI and by embodied AGI means something that's economically viable so you don't have a bunch of robots that can't work because they're too EXP expensive like actually like solving problems that we to have solved um and so like with that listen there's a lot to do this is going to be a major value driver I think it requires a ton of partnership from the VC Community from The Tech Community and certainly from DC so we appreciate all of you being here today and and your patience with my talk

Channel: a16z

Views: 10,463

Keywords:

Id: gKvVcEAua7s

Length: 10min 21sec (621 seconds)

Published: Tue Mar 05 2024