Zack Kass | Keynote Speaker | SpeakInc

The first thing I will say is I think we are on the verge of the most profound industrial revolution in human history. And I don't think it's close. I think that the fire and the wheel and the aqueduct and the Internet, these are all seminal moments in the human experience that moved us forward in step, functional ways. AI will present an even bigger opportunity probably than all of them combined. And the thought experiment I will leave you with. Well, open with. And leave you with is this. If you were asked to explain the Internet to someone in 1900, you would probably have an equivalent challenge to explaining the Internet to a caveman, which is to say our comprehension of the world is probably greater than someone in 1900 is to understanding a caveman's world. Which is to say we have probably been on an exponential journey for roughly the last 150 years at, as a human species that happens in sort of gradual and then suddenly very, very step, functional, incremental ways. And we're on the verge of a really big one. I mean really big. And it's going to happen in pretty incredible moments that will ultimately probably feel like, to our future selves, a somewhat slow journey. And the ask that I have of you is, please ask yourself tonight, spend serious time considering what someone in 50 years will struggle to explain to you today or 25, because that will help you understand how far we may have yet to go. Now, on that note, we are much further than you realize. So this is to say that the technology that comes out on a commercial basis, ChatGPT last week, can now speak to you and listen to you and provide images, and it has an encyclopedic understanding of the visual world. And that's amazing. Papers proved that that was possible about 18 months ago. So the science is moving much faster than the commercial applications. And this is a good thing, by the way. This is a good thing. And we can talk more about why this is a good thing. But from a scientific standpoint, we are much further than any of you realize, unless you're reading the papers. And my guess is most of you aren't because you're not scientists. And that's fine. But this is to say that we are making advancements in olfactory, in robotics, around touch, in video and 3D that you will start to see play out over the next two years in really profound commercial ways. We're also much earlier than you realize. I made the thought experiment earlier and I'll prove to you scientifically in a little bit that we are still beginning a pretty remarkable super linear Curve. We are still very early on the J, at least from a scientific standpoint. There are definitely some limitations, and we'll talk about those, but we are much, much earlier than anyone realizes on this AI journey. Okay? There are a bunch of major products in the market. I show this slide simply because if you have not used these. Good. Take photos. If you have not used these. No, seriously, please use these. Please, please, please. I am still troubled by a very simple fact. Two things. One, lots of people in this world who are aware of this technology don't use it. I don't understand why, and I really, really worry about this. And I'll tell you why. We accept the fact that our parents cannot use technology. It's just we accept this, and we accept it because the technological divide, the digital divide, didn't create an incredibly remarkable dichotomy between haves and have nots. Most people who don't have access to a phone or can't use it, we think of as sort of, you know, how nice of them to enjoy a pastoral life with nice bread in the morning and coffee in the afternoon. And that's fine. But the digital divide is going to cause a whole lot more inadequacy and disparity in the world if we're not careful. And this will happen across socioeconomic standards. So my plea to people is, please use the technology, if for no other reason, so that you're familiar with it, but also because it's really important that we bring everyone along with this. Okay, how did we get here? So we got here for four reasons, and these reasons are actually related, but mostly coincidental. And the first is the invention of the modern transformer. And I'll talk about the modern transformer in a second. But the next time you're at a dinner party and someone says, gosh, how did this happen? You can stand up and you can say the modern transformer, and I'll explain to you what to say next in a little bit. The second thing that happened was an explosion of computer. The zero interest rate environment was largely squandered, in my opinion, by most governments. We didn't build the infrastructure we needed, et cetera. However, we produced an enormous amount of compute. And that compute became effectively free against large loans, which created the way for companies like OpenAI and Meta and others to build massive, massive machines. And these machines also required an enormous amount of compute to run. And thanks to the explosion of compute at the end of the decade, we now have a lot of inference capacity. Without this, we probably don't arrive at this moment, without the zero interest rate environment, without the explosion available compute, we probably don't get here. The third is data. The Internet grew multiple orders of magnitude between 2010 and 2020. IP lawyers in the room sort of know this. You're like, you know, you're still trying to wrap your head around it. Now the most important thing isn't actually that the data grew tremendously. It's that it became much cleaner. And it became much cleaner because you can sell better ads against cleaner data. So the incentives are perfectly align. That being said, the explosion of clean available data to these researchers was really important. The fourth, make no mistake, AI is scientifically, is a scientific pursuit. It is driven by researchers who want to share their findings with others. And this is so, so important because without a collaborative environment, we probably don't arrive here. We almost certainly don't arrive here. And the paper that was written in 2017 by 12 Google researchers about the modern transformer is sort of exhibit A, okay, on the transformer. And this is super important. We went from building systems that thought in straight lines to building systems that thought in parallel. And the easiest way to conceptualize it is this. For those of you in the room who have read Proust and there are a bunch of attorneys, so I'm sure at least a few of you have, you don't have to self identify. We know you're the overachievers. When you finished a Proust novel, and I have never read one, I would never even think to. Somehow at the end of a Proust novel, you're able to build relationships between all of the characters and the moments in a million words. And you can actually sort of retell yourself these incredible, really descriptive stories that requires parallelized thinking, parallelized contexting systems used to require that you give a system a sentence and then another sentence and then another, and it could think forwards and backwards, but it couldn't build a relationship to something far back here. Now we're able to build relationships in parallel across dimensions. This is so critical, and this more than almost anything else, is what helped us arrive in 2017 to this incredible inflection in the quality of these machines, this ability to think in parallel. Another thing that's happening right now is really important and has been happening for the last three years basically quietly to most people. And that is that the quality of the model continues to improve and the cost continues to plummet much faster than Moore's Law. This is defying Moore's law. So in 2020, GPT3 came out. OpenAI launched what is now sort of a famous research product, GPT3, but of course made very little commercial headlines for reasons I couldn't understand. But I better understand today. GPT3 today is 1/100th of the cost that it was when it came out, and it's twice as good. GPT4 debuted six months ago. It's already half the cost than when it launched. And this isn't just gaming the pricing scheme. This is actual inference. Cost is coming down. These models are getting more efficient, they're getting smaller, they're getting better. This is great news for the democratization of AI. Critical because it means at some limit. Now, of course this asymptotes, but at some limit, the best models should approach very low cost. Another thing that's happening is really great. The hardware cost is declining. So while the models are getting more efficient in serving, we're getting much better at running them on the machines that we require. Now this is really critical because these machines are approaching a finite amount. Okay, this is going to drive the analysts in the room crazy, but if there's a designer in the room, you'll understand. I don't label the Y axis because it doesn't look good and it's impossible to read. But on the left is precision. So it's measuring the quality of the model on a scientific term called precision. It's just meant to show that from 2015 to 2021 the qualities of the models got so much better. And on the right is the size of the model, measured by billions of parameters, billions of tokens. So from 2015 to 2021, both of these saw super linear progress. And then from 2021 to 2023 we saw it again. Now there is no reason to believe that these curves will slow down. There is some consideration right now of actually the size of these models is going to start to fall. And I'll spend a little bit of time talking about that later. But there is reason to believe that we may actually be nearing the end of the large language model era, that the quality of the model will continue to improve as the size decreases, which would in theory also improve the efficiency. But this is critical and we now believe effectively that the transformer can get us to AGI. This was not well, this was not well understood about a year ago, maybe two years ago. It is well understood today that the transformer has material chance to get us to AGI. And the co founder and CEO of DeepMind actually talked about that this week. Okay, this is a philosophical journey. You should not necessarily agree with any of this, but please, please suspend disbelief for a moment and go with me on this journey, because I think it will help you ground your ideas first. The future is going to unfold so much faster than the past. And I say this as sort of a warning and also a note with a note of excitement, which is, if you were surprised by ChatGPT, just hold on. Because there is going to be a wave of scientific breakthroughs that lead to far more measurable and sort of meaningful outcomes in a commercial sense. And the reason that it won't happen, we'll talk about there's a policy risk, there's a compute risk, and there's an energy risk, but we will see a whole lot more progress made in a much shorter time. The second is, I think AI is going to become a utility, much like the Internet or the energy grid. I think we're going, we should expect to see this start to power critical infrastructure and ultimately become utilitized in the way that we treat most of our key infrastructure. Here's how I think this is going to happen. I think this is gonna happen in four phases. And the first phase is the phase that we are in today, which is enhanced applications. And the fun thing about enhanced applications is it comports with our model of the world. So today, if you wanna use AI, my mom, you know, my mom calls, she still calls, how do I use AI? And I say, okay, open up ChatGPT and do some things and then close the app. Now you've used AI, Congratulations, you're part of the, you know, the modern world. And, but that's, but that's how AI. That's, you know, it's funny, but that's actually how AI works. It's totally inefficient. But that's how we use AI today. And if you want to, you know, for those of you in the room who want to make sense of a long brief, the irony of that statement isn't lost on me. If those of you in the room who want to make sense of a big document and turn it into an abstract, you could put it in ChatGPT and then close the app, and that's how you use it. And we will be in that world for a little bit, probably 24 months. And I call this the enhanced application world. It comports to your mental model of how you use technology. You open up your phone, you open up your desktop, you open up an app, you use that technology and then you close the app and it works just fine. Now, I think this is terribly Inefficient based on where we could be. But again, we don't have to change any of our infrastructure and we don't really have to change the way we work. It just serves us, you know, it makes the things we do more efficient. And you're gonna see this embedded in all the technology that you use as attorneys, from Microsoft Word to Clio to the attorney specific stuff. The second phase is also comports to our world and I think is a lot of fun because it's much more efficient. And this is the agent powered phase. So agent powered applications is this idea that LLMs can actually drive our machines for us. And this is really exciting for a lot of reasons, namely that the world becomes a whole lot more natural language. So for example, let's say I'm having a great time. I get invited to a party tonight or tomorrow, and I don't want to fly home until, you know, I don't know what's today, Monday, I don't know, I've lost track of time. Let's say I don't want to fly home until Friday. So I say to my phone, okay, I'm going to fly home on Friday. Now, a bunch of things can happen automatically. A flight can be changed, my dog Archer can be rebooked, a hotel can be rebooked, and that can all happen automatically. This technology exists. It exists. I can show you the papers that prove it. I can show you the demos that prove it. The scary thing here is going to be passing the societal threshold. So we've met the technical threshold, we know that exists. When do we pass the societal threshold? When will you be comfortable letting a machine book your hotel, rebook your plane, rebook your dog walker, et cetera? When we arrive there, that societal threshold, we will have arrived at the agent powered phase. And that's gonna be really fun because again, we don't need to change our infrastructure. We will simply just automate most of the things that we do on the desktop on the iOS. I look forward to this. The third phase is the one that's really hard for people to grasp, in part because it's hard to make sense of. But it is a new operating system. And the reason this is hard to grasp is it probably requires a new hardware. It probably means that we get rid of the machines that we have been accustomed to using for 30 years, which, by the way, is an incredibly long time. If you think about it, the phone has largely remained unchanged. Of course, the mobile smartphone is just simply an iOS or simply a desktop in your Pocket and of course the computer is largely unchanged for 30 years. I think we are due for a massive update. This infrastructure is probably going to resemble like nothing that you have seen before. And you could imagine a wearable or an embeddable that exists in a natural language way because you probably won't need to do a bunch of manual typing. I say this, this is where I start to lose some people. Well, give it a second because I'm going to lose even more. But this is almost inevitable. This is almost inevitable. The current operating systems do not scale with the technology and we will have to update. I would bet that no one in this room uses an iPhone 21. I think we probably have five years left of the hardware as we know it, maybe six. And then phase four. Phase four is AGI. This, I have to just be honest, is an indefinite optimism phase. I can't tell you exactly what it will look like. How we interact with the super intelligence Ray Kurzweil talks about as being one of the most complicated parts of this. John Maynard Keynes in 1930 wrote a book called, or wrote a paper rather, called the Opportunities of Our Grandchildren and basically predicted that they would have access to technology that no one could comprehend and that their work week would be 30 hours and very prescient. Pretty remarkable. He and Ray Kurzweil somehow arrived at the same year. I'll talk about that in a second. It is very hard for you to comprehend because it's very hard to comprehend something that we truly have not seen before. But this is phase four. Okay, here are the risks. And by the way, I say all this with an incredible degree of humility, which is to say that all of these events have material potential of outcomes. And I don't want to talk flippantly about any of these things, but I think it's really important that we ground them in very specific terms. The first and I think greatest threat that AI presents us all is idiocracy. And if you want to know what that looks like, you don't have to look much further than TikTok. You don't have to look much further than what happens when someone stares at their phone for four hours mindlessly. We are starting to see the effects of surrendering our brains to technology, especially among younger generations. And I think it's pretty terrifying, to be honest. That being said, the greatest threat I think AI poses to us is that it creates a world so splendid that we stop being challenged and we simply surrender our minds and the problem solving skills and the entrepreneurial nature of the human condition to this technology. Pixar depicts this pretty well. In Wall E. There's a great Luke Wilson film called Idiocracy, where I borrowed this from. It's a satirical dystopian film. I recommend watching it. The idea that a very dumb man in the present wakes up brilliant in the future is sort of scary. The second is job displacement. And I want to be really clear here. I do not think that this is an actual risk. I do not lose sleep over job displacement. Every Industrial revolution sees a 50% net growth in jobs on the other side of it. That does not mean that living through that Industrial revolution is not hard at all. And it's really easy for us to say that the Industrial Revolution and the second Industrial Revolution and the Agrarian revolution were great for the human condition. And they were. But it also doesn't mean that if you were a miller, that it was somehow easy. That if you owned a company that produced small spools, that it was easy. That if you repaired sewing machines, that it was easy. No, it was incredibly hard. And there's a reason it was hard. It challenged your identity. We're not facing a job displacement crisis. We're facing an identity displacement crisis. We are facing a reality where people, a lot of people in this room will wake up day after day realizing that their job is changing radically. And that doesn't have any sign of slowing. And it's not that you won't be able to put food on the table. I don't predict that at all. There will be plenty of work. It's that you won't actually be able to reconcile why you went to school for eight years, all these skills that you acquired that you know, that are sort of innate to you, the things that define you. And I, when I talk to young people, I'll talk about this in a second. Basically warn about this and only this, which is that the human condition, the human experience, is going to change so radically that we need to ground ourselves in what it actually means to be human, because AI is going to start to displace a lot of the other things. So we'll talk more about this in a second. The third is widespread misinformation. And I'm here to just say it's gonna happen and it's gonna be bad. It's just gonna be really weird. And for probably a while, we're gonna deal with a lot of bad information. And the only good news right now is that we've never been more cynical about the things we read online. We're just really well prepared for this moment because we don't really believe the things we read on the Internet anymore. And this is a real shame. And when I talk to the press, my cry to the press is to recapture the trust of the average consumer, to stop polarizing the average consumer. Because if we can regain trust in our networks, in our news media, then we have a chance to actually have confidence in the outlets that are distributing us our news. Because if we don't trust the news media, we certainly won't trust Twitter and Reddit and our. And our sort of alternative sources, because that's just going to be flooded with fake content. And I worry about that quite a bit. The good news is that we've always, we are good at measures and countermeasures. Everyone in this room has been emailed by a Nigerian prince at some point. Everyone in this room has been emailed by their boss requesting them to wire them money. Right? We figure out how to, you know, how to catch fishers. We figure out how to stop scams, measures, produce, countermeasures. And by the way, at the limit, the good actors always win. This is the really cool thing about being human. The good actors always win at the limit, at least historically speaking.