Full Transcript

Full Transcript

Transcript

Nathan Labenz: (0:00) Hello and welcome to the Cognitive Revolution, where we interview visionary researchers, entrepreneurs, and builders working on the frontier of artificial intelligence. Each week, we'll explore their revolutionary ideas and together we'll build a picture of how AI technology will transform work, life, and society in the coming years. I'm Nathan Labenz, joined by my cohost, Erik Torenberg. Hello, and welcome back to the Cognitive Revolution. In today's episode, I'm excited to share an AI scouting report that I recently presented to a group of retail investors. This fast paced, high level overview is obviously quite different from our usual deep dive format, but I definitely think it's worthwhile for all of us to occasionally take a step back to review the fundamentals and survey the current AI landscape. I also hope that this can be a compelling introduction for intellectually curious people who haven't been tracking the field and who need a fast on ramp to understand where we are with AI today. Because I was told that the group was made up primarily of doctors and lawyers, I opened this presentation with a walkthrough of the current state of the art in medicine. After that, I gave a brief overview of what I consider to be the most important concepts in AI, including what goes into these systems, how they learn, and what we can say about what they really understand, their strengths and weaknesses relative to humans on a practical level, how one might think about investing in this space, though, of course, this is not investment advice, and most importantly, how these developments might affect society at large. Before we dive in, I wanted to mention a few other things as well. First, I am available for more speaking engagements along these lines. So if you have a sophisticated audience that could use some help catching up on AI in general or on a certain focus area in particular, drop me a line and I'll see if I can help. With 3 little boys at home, I will be far more able to accept remote invitations. But for the right audiences, I am open to occasional travel as well. Second, in addition to my ongoing work at Waymark and as an AI adviser to Athena, I recently got involved with a new AI advisory and implementation firm called Stellus AI. This company aims to serve small and medium sized businesses, you can think 50 to 1,000 employees, with a mix of guidance and solutions that drive AI adoption, efficiency and productivity gains, and ultimately business growth. I love that we're really getting into the weeds with this work. One of our very first projects is for a plumbing parts supply company that has decades worth of old part catalogs currently sitting on a giant bookshelf and is really struggling to help people find what they need. We're working on an AI solution to make this information far more accessible, and that could create major new opportunities for them as a business. If you would like to explore being an early customer of Stellus AI, please be in touch. Finally, I'm considering starting a very niche Cognitive Revolution job board where past guests and other companies I believe in can post their open positions. Right now, for example, Elicit is looking for seasoned software engineers to join their team and help scale their systems. Athena will soon be ramping up a hiring process for a number of AI and other technology roles. And Stellus will be hiring both AI engineers and AI advisers whose responsibility it will be to translate AI technology into practical results for businesses. If you're interested in any of those roles, please contact me and I'll help you get connected. I am super curious to see how many audience members are interested in such opportunities, and obviously, your response will help me determine whether or not this is a worthwhile project. If you have any roles that you'd like to advertise, you can ping me about that as well. As always, we appreciate it when folks share the show. This episode is more accessible than most and could be a great starting point for any friends of yours who've been sleeping on AI. So if you know someone who fits that description, please consider sharing this episode with them. And, of course, we invite your feedback. You can contact us anytime at our website, cognitiverevolution.ai, and I do read all my DMs on all of the social networks. With that, I hope this AI scouting report provides a solid foundation for understanding the exciting and at least somewhat terrifying AI developments that are unfolding around us every single day. I'm excited and look forward to sharing what I hope will be an interesting and super fast paced journey through the modern AI landscape. I call this the AI scouting report, the not investment advice edition. I'm going to really not pull any punches in terms of being very in the weeds about some of the important details of what is happening in AI now. We're gonna start off with just a quick rundown. Obviously, can't cover anywhere close to everything, but a couple of key points specifically from the field of medicine to describe the current state of the art. When you see SOTA in AI, that refers to state of the art. We'll talk a little bit about how that's already impacting the economy, what the limitations are of the current systems, then I'll get a little bit theoretical for a bit and talk about the inputs to how AIs are made and some of the science that actually goes into making them work as well as they do. And then toward the end, we'll get into the true not investment advice section where I try to and my crystal ball gets very foggy after just a pretty short time into the future, but where I at least have a (5:02) little bit of a sense (5:03) for how I'm thinking about investing and what companies are in the best position. I can close it with a few key questions that I think will really be super important to watch as we try to get a sense for where this whole AI thing is headed. I'm not gonna use analogies. I often call my content an analogy free zone because I think that the AI phenomenon is something that is really worth the time and effort to understand on its own terms. I think it is as important of a phenomenon and as fundamental to understanding the world as, like, understanding DNA and natural selection. To walk around the world without those concepts just leaves so much unexplained and puts you in a position of confusion. I think there are some core concepts in AI that are really becoming similarly important. And so I don't use analogies to describe them. I'm gonna try to describe them in very literal terms. Simple as possible, but very literal without being wrong is what I try to do. And then I won't do any, like, big predictions, although I do have few thoughts toward the end on at least how I'm thinking about allocating capital. So I guess I'll break my no predictions rule just ever so slightly. Okay. Here is a little tap dance through the current state of the art, specifically with a focus medicine. So where are we in AI today? I think it's really important just to try to keep tabs on that. It's a big part of what I do is try to have an up to date sense of what our current AI system's capable of. So it's really remarkable, but it is undeniably true at this point that since the 2022, the best AI systems have been better than the average human at the average task. That is to say, if you go find some random task that might be typical of white collar computer work and you take a typical intern or young employee and ask the AI to do it and ask the employee to do it, you're gonna get better results from the AI. That's been true for almost 2 years now. In addition to that, the best AI systems are really now closing in on expert performance for routine tasks. And the word routine there is really important. Routine means it's obviously common and that we know what good looks like. In medicine, you might think of something like the standard of care, something that is well established. So that's a really remarkable phenomenon and it's happening super quick. GPT-two was released in 2019. I read about that the hospital when I was there with my wife as she was about to deliver our son, and he's just turned 5. 5 years later, we've gone through a series of different generations of the models to now the current one, GPT-four, is closing in on human expert performance on really demanding tests of cognitive ability. This test, the MMLU test, is comprised of undergrad and graduate school exams across a super wide range of topics. And the average GPT-four score at this 86 compares to the human expert in their domain. So it's coming close to matching what human experts can do in domain, but it's beating the humans. If you were to have one human try to answer all the questions across all the domains, then the GPT-four result would be notably stronger. So that is pretty crazy, and it's happened in a pretty short period of time. Now to get more specific in terms of how this is playing out in medicine, here is a similar curve focused on a US medical licensing exam. To pass is somewhere in the range of 60, low sixties. So the first system that pretty clearly passed the medical licensing exam was MedPalm. That is one from Google. Not too long after that, really just a few months, look at this scale. December 2022 and March 2023, they came out with MedPalm 2 and it hit 86 percent. And again, that is closing in on expert performance on the medical licensing exam. It goes further in medical question answering. Here are 9 different dimensions on which they evaluated. This is, again, out of Google. They evaluated an AI and a human doctor head to head, and they asked human doctors to evaluate both the AI and the human doctor participants. The top section here are the good things, the things that you want to see in your answers to your medical questions, things like reflecting the consensus of the medical establishment, reading comprehension, knowledge recall, better reasoning. The orange is how often the AI was judged to have done a better job, and the blue is how often the human doctor was judged by other doctors to have done a better job. The gray was in the in between is when it was considered to be a tie. So you can see that the AIs are dominating on the good things, and then here are the bad things. There's this is the one of 9 categories on which the AI was judged to be worse. It did more of bringing up inaccurate or irrelevant information. Not a ton more than the human, but a bit more. The humans did all of the other bad things more than the AI did. So we're, again, getting into this strange zone where state of the art AIs are closing in on the performance of experts even in cognitively demanding areas like medicine. Here's another study. Again, is medicine. This is differential diagnosis. On the left is the AI. On the right is the human. There are 4 bars here that refer to whether they had the exact match on the proper diagnosis or some increasingly permissive definition of accuracy. And then the top k is how many guesses they got before they got to the right one. So, again, you see the AI is at a significantly higher accuracy percentage across the board as compared to the human doctor. These are primary care physicians specifically in this study. So I tell people in all seriousness these days that if I had a medical condition that genuinely worried me, I would not be without my AI doctor. I would also, of course, want to have a human doctor. I wouldn't put all my trust into the AI, but I would absolutely be cross referencing their answers and checking them against one another to try to get the best of both worlds. But I would not be without my AI doctor with a serious medical condition in today's world. A lot of other interesting applications just in medicine as well. This is a project on virtual tissue staining. I don't know a lot about the traditional way that this is done, but biopsy and slicing tissue and staining it and looking at it under a microscope, I'm told that this can take hours and that it's too slow to be done while the person while the patient is in surgery. Now there's a new technique that includes this in situ imaging, and then they add this virtual stain to it. And now they can get these sorts of images out of patients in just seconds while they're still in the Operating Room. So if you're doing some sort of procedure to try to remove a tumor, you can get a lot of these images in real time as you're going. And that, of course, can lead to better patient outcomes. So the creativity that people are bringing to all these different applications of AI is super, super interesting. Here's one that flies under the radar and always blows my mind. This is literal mind reading. I have a full episode, actually 2, on this project. But what they've done here is taken fMRI scan data and taught an AI to reconstruct the image that the patient was looking at when the scan was captured. So you're in an fMRI, you're shown these images, they capture the data, and they record what you were looking at the time. And then you do a bunch of that. And then after that, the AI can take the brain data, the fMRI data, and reconstruct the image that you see here on the right where the original one was the one on the left. So you can literally decode brain activity and reconstruct images with this level of fidelity. And I often think, man, I feel like when I was a kid, this would have been major news. And somehow today, we live in a world where it just crosses over and gets a few retweets or whatever, but most people don't even realize that this kind of thing is out there. So literal mind reading is happening. And with this original one, it took 30 to 40 hours of scan data to train the AI to code one person's brain. With the latest update, they're now down to 1 hour of scan data required. So it is actually starting to get feasible where you could put somebody into the thing for an hour and get this type of process up and running without it being a huge undertaking. So, obviously, we do a lot of things. I think one thing that's important to keep in mind, and I'll get into a little bit more detail in just a second comparing what the AIs are best at and what the humans are best at, but they can't do everything, of course. So this is a study of what they can do and how that breaks down by different kinds of jobs. Interestingly, and this is definitely a surprise, and I think this has become common knowledge at this point, the AIs that we're getting are not really the AIs that we expected. We expected that they would automate relatively low wage and low status labor first, but on the contrary, they seem to be impacting high wage and high status labor first. Your proverbial plumber or custodian of a building is, like, not feeling much impact yet from the current wave of AI, but your doctors and lawyers are definitely seeing significant impact. So this study broke jobs down into 5 different categories. The green here is the lowest, let's say, status wages and lowest impact. As you go out, it's the fourth quintile that's the highest one out. So the most impacted are the fourth quintile, and then the fifth quintile comes in a little bit again. And to read this, you can basically think of all the area under the curve is amount of work that AIs seem to be poised to be able to do. Everything outside of the curve, they would not be able to do. So if you're, again, a plumber or a custodian, there's a relatively small amount of your overall work that is likely to be impacted by the current wave of AIs, whereas if you're a doctor or lawyer, you're maybe out here somewhere. Like, this is to say about half of people have about half of their tasks such that AI could impact them, plausibly do a significant part of them. I think that's really something that we are going to be adjusting to for quite some time. And again, we're just getting started in this AI era. Here's one thing that was really just announced and is coming to market right now. This is AI nursing. This company Hippocratic AI has a pretty interesting go to market proposition where they are charging by the hour for their AI system. So you're starting to see AIs that are actually competing in essentially a labor market, not by API calls, not by number of tokens generated, but literally by the hour that they're working for you and doing stuff. Of course, they're, like, scalable and cloneable in the way that software systems are, but they're actually starting to charge an hourly rate. They charge $9 for their AI nurses. And this company specifically is, like, trying to be very cautious, very safety focused. Even though you saw the result above on the relative strength of AI systems on diagnosis, this company is taking a strong we're not trying to even get into diagnosis, but we do think that we can get AI to do a very reliable job on a lot of these, like, follow-up tasks that either nurses do or that, frankly, today nobody does because there's just no time in the day to do them. So they have all sorts of things where they'll call you and work through your pre op thing or follow-up on your discharge instructions or make sure you're taking your medicine or checking on how you're feeling, all that kind of stuff. And $9 an hour is the going rate for an AI nurse in today's world.

Hey. We'll continue our interview in a moment after a word from our sponsors.

(19:07) Okay. So where does it stop? This is where we get to limitations. Where it stops is things that are not routine. So can AI do new science? In general, the answer today is no. There are a few very engineered and purpose built systems like your AlphaFold that I'm sure people have heard of that are specifically designed and highly tailored to do a narrow task in science, and those are high impact systems as well. But the general purpose systems are, like, chat GPT like things that you can talk to and just have do whatever you want. Those things today cannot do new science, and that basically extends to general, novel, unfamiliar, somebody really has to figure something out for the first time situations across the board. They are not good at those. I often I used to say, actually, no eureka moments. That was one of my summary bottom lines for the state of AI. Now I have to say precious few eureka moments because this is actually one instance where GPT-four was doing a better job on some pretty novel stuff compared to even human experts. I'll just try to play this video. And the scenario here is they're trying to train a robot hand to do various tasks. The task in this case is twirling a pencil, and they use a technique called reinforcement learning. So this gets a little meta because it's AI training other AIs. But to do these sorts of reinforcement learning programs, you have to have what's called a reward function. When the AI is just beginning, just starting to learn what's what, it doesn't do very well. And so it needs some score to learn, like, when am I doing better? When am I doing worse? I can do more of the stuff that's getting me the high score and gradually find my way to figuring out how to do the task. So the reward function, could imagine sitting down and trying to come up with, okay, if I'm looking at this, how would I score it so that I could give the AI a signal from which to learn how to get better at this task? And might you think, okay. Maybe I could measure the angular momentum of the pencil and turn that into a score somehow, something along those lines. There are people that do this. Right? This is an expert task to write reward functions for robot reinforcement learning. Turns out that GPT-four is significantly better at writing these reward functions for robot reinforcement learning even than the people who did it. And this is not like you're a person on the street sort of task. Right? Often just, again, step back and think, go talk to somebody on the street and say, hey. Could you write me a reward function for this robotics reinforcement learning project that I have? You will not they will not even parse that sentence, let alone give you a serviceable reward function. So you're, by definition, in expert territory here, and yet GPT-four still able to significantly outperform the human reward function writers. That doesn't happen very often. This is the exception, not the rule, but I have had to update my summary position from no Eureka moments to precious few Eureka moments because of a number of little projects like this. Here's one other one that kind of I'm still wrapping my head around. This is a example of essentially a ChatGPT like model trained on just tons of DNA sequences, 300,000,000,000 base pairs worth of DNA sequences all drawn from bacteria and phage DNA data. And what they find is that now they can use this thing to generate new stuff. So they've found that they are able to generate a variety of CRISPR like complexes. These are all complexes generated by the DNA generating AI. And the other thing that they can do that's really interesting is they can start to perturb the AI and see how confused it gets and use that to determine how essential different genes are to the survival of the cell. They call this in silico gene essentiality experiments. And apparently, this is, again, quite a cumbersome process to do if you're doing it in an actual traditional lab context. But now they can run it in just, like, a couple seconds for because they can, you know, obviously do it in the computer. This is just from the last few weeks. So this is really some of the latest and greatest, and I think we're, again, just beginning to see the impact of all these advances across a wide range of fields. So, hopefully, that has you motivated to understand a little bit better how it all works under the hood. The core inputs to AI systems are data, compute, and algorithms. That's like your trinity, your 3 legged stool. Without any of these 3, it's not gonna work. So you need data from which to learn. You need compute to crunch the numbers, and you need an algorithm that actually translates that data and that number crunching into something that actually works. So I'll take you through each of the 3. For starters, all these systems are really just information processing systems. You might call them circuits. They are inspired by human neuroanatomy, but they're definitely not much like human neuroanatomy. One thing that they all have in common is that they are all unidirectional. They are all differentiable, which means there's no, like, cyclicality. In the human brain, you have different parts regulating one another and a lot of crosstalk. In the AI systems that we have today, they proceed very linearly from one layer to the next. So this is just an example of a really simple thing where you might have a number and your goal is to classify this image as what number is this image. This one is obviously a 7. You can break this up into pixels. You can feed raw pixel data in here, and it progresses through the layers. And then your goal is to get the 7 to be the one that lights up. And if that happens and it's working effectively, then you have an AI now that can identify which number it is. So it's a very simple system. I strongly recommend if you wanna go see great visualizations of this, the YouTube channel 3 blue 1 brown does some of the best visualizations, not just on this, but on all sorts of different mathematical things. But this short clip is drawn from a longer series that is really good for building intuition if you wanna go check something like that out. Okay. So there are these information circuits, but let's get a little deeper now on how do you actually get it to learn. A super important concept that's sitting at the heart of this is the loss function. The loss function is the way that you score the output from the AI. So let's imagine you are at the beginning of a training process. And by the way, they usually just randomly assign the numbers at the beginning. There's not like there are some tricks, but traditionally, the numbers in all these different positions are just randomly assigned. So what happens at first is basically garbage in, garbage out. You have your kind of matrices that are crunching all these numbers. It starts in a very random position. Let's say you go ahead and put a 2 in. You crunch all the way through the thing, and you get this kind of result out. And it gave a high rating to a 3 and also a high rating to a 6, very strangely, and a high rating to a 1, and the actual 2 was a very low rating. Okay. That's a bad prediction from the ads. It's not able to do the task at this point, but you can score it. So you can say, hey. Your score on 2 should have been higher, and everything else should have been lower. And now you have a score. You can put that into a single number, and you can say, this is your score. And then you can do what is called back propagation, which is saying, if I go back through all the different parts of this system, all the different numbers in the matrix, and I look at each one and I ask, if I tweak this one, should I tweak it up or should I tweak it down to make the overall score a bit better or a bit worse? This is ultimately chain rule from calculus. For each number in the thing and if you're talking about a GPT-three scale system with famously a 175,000,000,000 parameters, then you're doing this literally a 175,000,000,000 times. In these small systems, it can happen very quickly, but it does take a lot of number crunching as the systems get big. But for literally every single parameter in the system, you're just saying, okay. I can either move this up a little bit or down a little bit, and I wanna just do that in such a way that makes the final score a bit better. It's a really simple concept, actually. You have a loss function, which gives you a score. You use back propagation to work your way using the chain rule of calculus from the end all the way back to the beginning of the network, tweaking every single little parameter along the way, and then you do that in a loop. And that loop is called gradient descent. So I'm sure you've heard some of these terms flying around. Gradient descent is the process of just through all these little incremental changes, gradually finding your way to something that works. So this is a visualization of what's called a loss landscape, which is like at any given point in the vast space of all the possibilities for the numbers in this information processing network, you're gonna have some loss value. And as you work your way down to the lowest possible loss, that is the process of gradient descent. And this gets complicated. You've got all sorts of different strategies for how to find your way down the hill, and you can imagine people have been working on this for a long time, so there's lots of different little complications to it. But this is basically the core thing. You have to have something that is differentiable so that you can do this backpropagation. You have to have a score so that you have the loss function. And if you have those and you run this in a loop, you are doing gradient descent, and you are on your way to training an AI. What you'll typically see are these loss curves where as you go step by step, the loss gets better and better. You're finding your way to the best possible solution. Traditionally, this was done with curated datasets. So this is like a classic dataset, a dataset of just nothing but small images of hand drawn numbers. I think it's called MNIST, but that's all it is. Right? Small images of hand drawn numbers, and there are some that you're meant to train on, and then there are some that are meant to test how well you did from the learning process. And typically, what people would notice is that the more you would train, your training loss would go down. You would continue to see better and better scores. But then when you actually tried it on the test set or also sometimes that's known as the validation set, you would see that at some point, the validation results would get worse. And this was typically considered to be the point at which the model is overfitting. That's another term that you'll potentially hear, if you spend more time in this area. When it's overfitting, it is learning now idiosyncrasies of the dataset that don't actually mean anything for the general case. So, essentially, you can imagine memorizing a bunch of strange weird cases and getting, you know, really good on the test, but this is teaching to the test. Now you, you know, go and show other handwritten numbers. It's actually getting worse because it's learning the wrong stuff. So past this line where the validation gets worse, that is known as overfitting. Okay? Here's where the paradigm has changed. Right? That was the traditional thing. You'd have a dataset. The dataset was finite. You train on a certain amount, and then you test on a different amount. And the game was like, who (28:19) can come up with (28:20) the best algorithms, the best architectures to get the highest validation score? That was the pinnacle of the field. Now things have changed very significantly, and the big unlock has been figuring out clever loss functions that allow you to tap into huge amounts of unstructured data. There's only so much labeled data in the world. And the key thing about these datasets is that, again, you know what the number is. Right? There's a the dataset includes the image, but it also has the label. So you can check your answers against the actual answer key. But what they've done now with these, what's called unsupervised learning, is they've come up with clever formulations to allow the world's web scale data to serve as training data. And there's 2 big things that have been really popular recently. One is called next word or next token prediction, and the other is image denoising. So in next word prediction, and this is how we're now getting to ChatGPT and that sort of model, the whole body of all human text becomes available because your job is now just predict the next word. So you take all of the all possible text. You cast that into a vocabulary. There are 50,000, sometimes a 100,000 possible outputs in these text generation systems. And you do something very similar to what we showed here, except instead of 10 numbers to choose from, there's now 50,000 possible NEXT tokens that you could output. And you get a very similar score and you can do backpropagation. But the key is that the text itself provides the answer. So now you can just crunch all available text, for every single bit of text, you have the opportunity to score yourself with a loss function and to do that back propagation. So this is how the datasets have become huge. A similar thing has happened with images where somebody came up with a very clever idea of what if we first added a bunch of noise to existing images? What if we degraded the images purposefully with noise and then we trained an AI to denoise the image? That would allow us to use all of the images that are out there for training data. So this has been I can't overstate the importance of this paradigm shift from small curated labeled datasets to web scale unlabeled but self documenting datasets that is known as unsupervised learning. And this is how all the main scaled systems in today's world are trained.

Hey. We'll continue our interview in a moment after a word from our sponsors.

(31:02) Okay. Then you also need so that was all data. Right? Where is the data coming from? What is the nature of the data? How are using the data? You also need compute. Of course, there's been massive amounts of investment in computing infrastructure. NVIDIA's stock price has gone up tremendously because they are the leading company that is providing the GPUs that are doing all the compute. The key thing about GPUs as opposed to CPUs that you need to know is just that they allow for massive parallelization of the calculation. You can imagine that if you're gonna ask a question about a 175,000,000,000 numbers, do I tweak it up or down? You don't wanna have to do that sequentially because it'll take a really long time even if you have a really fast computer. So you need to come up with some way to ask that in parallel for as many of those different numbers as you can and then batch the results together and then make the updates. So that's part of where the algorithm, progress comes in, and that takes you to the transformer. Transformer is the number 1 pretty much dominant approach to AI today in the world. I'm trying to not use too many advanced vocabulary terms because I know that people get lost in the jargon. So here's a quick glossary of terms. The transformer is just a type of information processing circuit. The tokens are is another word for the words, or it could be like a little part of an image. When you embed something, that is the process of converting that text or image input to numbers. The neurons are the nodes in the network. They're generally organized into layers. I said earlier, everything has to be sequential. You can't have, like, loops in there. Sometimes these are also called activations or the values at the particular neuron places are called activations. There are 3 different parts of the transformer that are important. The attention mechanism is the one that gets all the attention because it was one of the big breakthroughs. The MLP is the multilayer perceptron, and then you have these things called nonlinearities. Typically, each layer will have all 3 of these. Then you finally get to the parameters and weights. These are the numbers that are actually in the thing sitting there that were trained, that were learned, that are going to do the number crunching. You'll also hear words like logits. These are the final numbers that get output before they're finally converted back into actual text. The forward pass is the process of actually running the thing once, and the model is the thing that is actually run. So I'd just like to go through these real quick because there's so much jargon, and it's very easy to get lost in all the jargon. Okay. So I mentioned that the transition to unsupervised learning and the massive amount of data that comes available when you switch to that paradigm has been super important. So let's talk about scaling and then a couple of really interesting observations as people have started to scale up these systems. They're getting really big. There has been a whole study of what are known as scaling laws, which is to say how good are these things gonna get if we put a certain amount of resources into them? And what you'll notice across the board in these sorts of system designs is that there's always a Pareto curve. There's always some optimum point where you can make the model bigger in terms of parameters, but then it takes more compute to do each process. So you can't for a given compute budget, then you can't process as much text. If you make it smaller, then you can do more loops, but the smaller models don't tend to learn as fast. So there's always this sort of optimum point where you're choosing the right balance between how big you want the model to be and how much data you're gonna run through it in the training process for a given budget. So each of these different colors represents a given compute budget. The we're here at 6 to the 18 times 10 to the 18. That's big, but it's still 8 orders of magnitude below the White House threshold for reporting large training runs, which they set at 10 to the 26 flops. A flop is a floating point operation. It's literally just a small unit of compute. You're actually doing a thing with number. You're doing that 10 to the 26 times to trigger the White House reporting requirements. And here, they're going from, like, 8 to 5 orders of magnitude less than that. But what's key about this is these small studies have allowed them to then draw these lines, which allows them to project out into the future as they're really scaling these systems. How good do we think they're gonna get? We can predict the loss based on this scaling law for a given compute budget. And this is how people are starting to back into these things like, okay. We should have this many billion parameters, and we should have this many tokens if we wanna run through it. And we it's gonna cost us this much money. So how big is this getting today? The big systems, GPT-four was trained on 10,000,000,000,000. That's essentially 10,000,000,000,000 words, and it cost reportedly some tens of millions of dollars just to run the compute. That does not include, like, the salaries of the researchers or collecting all the data, but literally just to run the servers to do the work. GPT-five, allegedly in training or getting close to the end of training, is expected to be trained on 10 times more than that, a 100,000,000,000,000 tokens, and is expected to cost a few billion in compute. So that's pretty crazy. It is bringing also some crazy results. So one of the things that is most remarkable to me about these systems is that they seem to learn things that they were not explicitly told to learn. The first sign of emergence, this is a story from the OpenAI team from back in, so this is way smaller. They didn't have all the compute, all the resources they have now, but they trained a model just to predict the next character in Amazon reviews. And then as they were looking inside and looking at the activations, which parts which nodes in the architecture had high values and under what situations did those positions have high values, What they found is that there was one little neuron that was lighting up strong if it was a positive review and was, like, very strongly negative if it was a negative review. And people have been trying to do sentiment classification for a long time. This is, like, a pretty well established thing, and people had all sorts of techniques. What they were amazed to see is that this system, which they had only trained with a loss function of predicting the next token, internally had this neuron that if it was on its own, a sentiment classifier would have been the state of the art sentiment classifier at the time. So this is like a really profound observation because it's like, wait a second. We just trained this thing to predict the next token, and yet internally, without us ever giving it any signal or information about the sentiment, even the concept of sentiment, it knows nothing about that, or at least we didn't tell it to know anything about that. Somehow it has come to represent sentiment as a means to do what it is ultimately doing, which is predicting the next token. So somehow representing or understanding what the sentiment is helps you predict the next token, which that intuitively makes sense that it could help. But what's remarkable is that it is learning that. So as Greg Brockman, the president of OpenAI, put it, semantics emerged from a syntactic process. Even though it was only trained to predict the next token, it learned a higher order human recognizable concept, which is sentiment. That is really the sort of breakthrough that is at the heart of the advances now that we're seeing that make these systems so general and powerful. There's been a lot of other study. I'll go quickly over this because I don't have a ton of time for it, but similar phenomena have been observed when you train a model to play a game and you maybe just train it to predict the next move, and all it gets to see is a sequence of moves like this. Like f 4, this is Othello. F 4, f 3, d 2, f 5, that's all it ever sees. And yet they can look inside the model and begin to decode that it is actually representing a 2 dimensional board state and updating that board state with each move as pieces get captured and territory switches hands. So all it ever saw was this, but it learns to develop its own conception that looks more like this. This is also happening in image understanding of computer vision models where you have these different neurons that seem to respond to a particular concept. And then people figured out how to reverse engineer what would be an image that would maximize that, and what would that look like? And you'd see these sort of trippy things where it's, okay. That's a window, and that's a something, and that's a wheel, and that's what maximizes these particular neurons. And then as they feed forward to the next layer, these sort of window and car body and wheel detectors all send a strong signal to what appears to be a car detector. And this is the image that would maximize the response from the car detector neuron. Now there was no engineering to say this should be the CAR detector neuron. All of this is an emergent process that just results from this definition of a loss function, the process of gradient descent, and the massive scale at which these things are now operating. These models are learning these high order human recognizable semantic concepts from this purely syntactic process. That is something that I think way more people should understand. And if there's one takeaway, it would that I would want you to leave from today with, it would be that. Okay. So what this is a really another kind of profound example of it. Here's a task called modular addition. You may remember this from your MCAT studying days or whatever. The idea here is to take 2 numbers, add them together, then modulo divide by a third number and take the remainder. Right? So 5 plus 12, modulo 12, it's 12 divided by 12, remainder is 0. 17 plus 8 is 25 divided by 10 is 2, remainder is 5. Okay? So they train a model to do that. It learns to do the task. But right in here, what is the process where you would traditionally call overfitting because it, like, learned some, but then the validation performance falls off. The test performance continues to get better and eventually memorizes the whole test or the whole training set, but the validation is bad. So for the longest time, people have just said, this is useless. Right? You've just trained a model and memorized things, but it didn't actually learn the real task, and so they would have stopped long before you're notice this is a log scale. Right? So this happens at, like, hundreds of optimization steps. Finally, somebody said, what if we just run this a really long time? What if we give it, like, 10,000 optimization steps? What they find is they get out toward 1000000 optimization steps is that it actually generalizes and learns a real solution. They call this grokking, generalization beyond overfitting. So it's overfitting here. Traditionally, people would have stopped here because, okay, the validation loss never gets better once it starts to get worse, except that it actually does. It just takes a lot longer for it to finally learn to do this task. And now at the end of this, it can actually do modular addition even on examples that it hasn't seen. It learns the examples that it sees in the training early, but takes a long time for it to grok the general concept of modular addition and do it for unseen examples. But that's really profound. Right? That it's able to make this shift from memorizing the examples to learning an underlying algorithm. The algorithm that it learns is, like, quite weird. I'll show that in just a second. But this is a really key point that we don't know in the process of training these things when something will be grokked, when the model will go from just predicting whatever next token might seem plausible to actually figuring out the underlying concepts and being able to use them across the board. So for my money in the GPT-four technical report, which is the paper that they put out, this is the most important sentence. Certain capabilities remain hard to predict. We showed the scaling laws that show how they can predict what the loss is gonna be. That is the aggregate score. But what's really still a mystery is how does that aggregate score translate to particular behaviors, to particular abilities, to which kinds of problems it's gonna be able to solve and which kinds of problems it's not gonna be able to solve? And here you see one where as they got models bigger and bigger, this was from a paper called the inverse scaling. There's actually a contest to find, are there certain tasks where as things get bigger and bigger, the performance gets worse? And they found some of these tasks. One was called the hindsight neglect problem. I won't bore you with the details. But, basically, across several generations, the models got worse at this task with increasing scale. And then with the next generation, it seems to have grokked that concept, and now it's perfect from being bad to worse to good in a surprising way. So capabilities remain hard to predict. We know that we can scale things up, and we know that the loss will get predictably better, but we don't know for any given task when this process of generalization will kick in. It's a surprise for all of the different tasks that we might be interested in. So this is a really profound challenge in terms of understanding what these systems are doing and also, like, how to control them. Okay. So I said that they're human level. They're definitely not human like. Here's just a quick I won't even go into the details of this, but somebody went in and reverse engineered. How is it learning to do modular addition? And it turns out it's doing it in a very weird way. It's doing it's like learning sine and cosine functions, and it's essentially rotating around the circle. No human would really do it this way. It involves, like, Fourier transform math. So it's a very alien process, but it does add up to a reliable algorithm. And now that this has been reverse engineered, you can actually see how it is doing it. You don't have to like it. I think you should probably be a little bit scared of the fact that it's learning such alien solutions to these problems, but, nevertheless, like, this is what it does for this particular thing. So I really emphasize how while they are very human like in their abilities in some ways, they are also not at all human like in their underlying mechanism, and they and we should not think of them as human like where it really counts. Okay. So here's the tale of the tape. This is important. What are AIs good at? What are humans good at? AIs are much better at breadth. They've read all the information on the Internet. They can speak all the languages. Even human experts don't have anywhere the breadth of a modern AI, but the human experts do still have better depth, especially in their area of expertise. Breakthrough insight is humanity's biggest edge right now. As I said earlier, few eureka moments, precious few eureka moments. The AIs are a lot faster, and they are a lot cheaper. Typically, I would tell people that you can expect they'll be, like, at least 10 times faster and usually 90% cheaper if they can (45:36) do what you need them (45:37) to do. They're also, like, super parallelizable, super clonable. They're, you know, ready whenever you are. You can pick up where you left off 24/7, so they have, like, lots of availability advantages. Memory is something that I've really come to appreciate in myself. My memory is certainly not perfect, but it does give me a coherent sense of who I am, what I'm trying to do, and the AIs don't have that in the same way. They have a very limited and brittle memory as it stands today. They do have great technology diffusion speed. A paper that I think should give everyone a little pause is natural selection favors AIs over humans. The key thing there is that when an advance is made, it can often be ported to all the different architectures, all the different models very quickly, whereas, like, we struggle to learn new things, especially as we get a little older. Bedside Manor is another one too that AIs are actually beating humans on in evaluations, but they're really not good at adversarial robustness, meaning they're quite easy to trick. You can confuse them. You can trick them. This is sometimes known as jailbreaking. If you go on AI Twitter, you'll see lots of examples of people getting the AIs to make fools of themselves in ways that humans would never do. This is one of their main weaknesses is that they're quite brittle and not really robust to being tricked. Okay. I know I'm going super long. I'm skipping over best practices for business, and you might be able to check that out on my podcast feed if you're interested in that. But just to cash this out a little bit to some investment advice, or I should say not investment advice. The question that a friend of mine poses is, are we approaching a big tech singularity? The way that he defines that is a moment at which the big tech companies have such advantage because of their dominance in the core inputs to AI. Those are, again, data, compute, and algorithms. Are we reaching a point where they might be able to enter into all sorts of different industries and dominate those new industries because of their core strength in AI and in the inputs to AI? And I would say that's looking increasingly plausible, and certainly the stock prices of the big tech companies would seem to reflect that. This is a chart of how much compute different companies in the world have. This is Google, Microsoft, Amazon, Meta across the top. Apple comes in at number 5. Alibaba is the top entrant out of China. And then things get, like, relatively small. And this is Oracle down here, by the way, which is a multi $100,000,000,000 company. But these are the dominant players in compute. The big 4 in compute today, Google, Microsoft, Amazon, Meta. Nobody else is on their level. And there's really no path for any company that's not already on this list to get into the compute game at anything approaching similar scale to what the big guys already have. If you're prepared and able to spend billions to train a single AI model, it's gonna be very hard for anyone to enter into that market and rival you. So I don't pick stocks. I genuinely don't pick stocks, but I do have a little investment club with some friends of mine, And I've literally made one individual stock recommendation to the group all of history, and that was NVIDIA about 2 years ago because I could just see where this was going, where everybody's saying, man, look at these scaling laws. Right? To compete, we're gonna need to hit this compute budget at this given time. We're gonna need to collect this amount of data. Of course, we're gonna have to do a lot of research on the algorithms too, but you're not the algorithms right now are, like, incremental improvements. Nobody has recently come up with an algorithm that just blows everything out of the water. You gotta have all 3 elements. You gotta have the data. You gotta have the compute, and you gotta have the algorithms. And so everybody's buying compute from the same place. That's NVIDIA. I think all the chipmakers are probably going to be doing pretty well. People are interested NVIDIA is making insane margins, so people are very interested in, like, AMD and Intel as well. And, obviously, the government is getting involved and saying, hey. Maybe it's not such a good thing that this is all getting manufactured in Taiwan. Maybe we should bring some of it back to The United States. But I would say that probably the safest bet in the space is the fundamental compute providers, both at the layer of manufacturing the chips and then also at the big tech layer because they have all 3. They have the data, the compute, and the AI researchers that are constantly working on new algorithms. This is my list of, like, live players. That is to say, who are the companies that really have a chance at shaping the future? And these are pretty familiar names at this point. OpenAI is the leader. Google still has the deepest bench in terms of the most researchers, the broadest research agenda. Anthropic is a real dark horse. They're not very big, but they're partnered with both Amazon and Google. And there are a bunch of people that left OpenAI to found this new company. Meta is certainly a big one. They're putting out a big open source model you'll hear a lot about called Llama 3 right now and over the next few weeks. Microsoft is not quite as cutting edge in terms of research, but they've got it where it counts in terms of datasets or data centers, and they're partnered with OpenAI. Never count out Elon. I think Tesla and X will be interesting. A dark horse that is not a public company, you can't invest in it, but definitely one to watch is called Character AI. They basically make AI girlfriends, and people are spending a lot of time talking to their AI girlfriends. That's a whole other dimension of sort of societal impact. What's gonna happen to the birth rate is, I think, a very interesting long term question. Mistral is the European champion. They're based in France. I would say, you know, they probably wouldn't have a chance at really competing except for the fact that the European Union may say, you know what? Here's a few billion dollars because we wanna have one of these leaders on our soil. And the geopolitics of this are gonna be very interesting to watch. Obviously, what's going on in China is a big deal as well. They're not at the same level as The United States, but they're not too far behind. If you were to go to the big AI conference in New Orleans called NeurIPS a few months ago, I didn't personally attend, but I was told you heard a lot of Mandarin conversations on the floor. There are tons of talented Chinese researchers. A lot of them are here in The US, but also, obviously, a lot are in China. The chip ban, though, may be a big deal for them because they don't have their own domestic chip industry. And if they can't import the chips, they may have a hard time reaching those next levels of scale. Apple, hopefully, is gonna make Siri good at some point, but I would say they're currently behind. But they're you can't count them out too much because they are you look here on the list, they are the number 5 on compute and certainly they have more money than anybody could possibly know what to do with and that can pay a lot of researchers. So, alright, I'm gonna land the plane here in the next 2 minutes and then I'll be happy to take a few questions. So are we all going to die is obviously a question that people are asking. I think that really depends on what happens next in terms of how much things scale and whether the scaling laws hold. They're called scaling laws, but they're not fundamental laws of nature that we at least it doesn't we don't have that level of justification for them. They would be better maybe described as scaling trends. And so the trend could break. So are we gonna see something where the sort of capabilities level off at a human expert level, or do they just blow past human expert? I think that's probably the single biggest question, and it's hotly debated. There's not really a clear answer there. My guess is that we're gonna get something alien that'll be superhuman in some ways, but it'll still be weird and weak in other ways. And how that plays out is ultimately gonna be weird and very hard to predict. One thing that I do think is strange, though, is that we're racing into this without a real plan. Nobody really has a sense for how these systems are going to be controlled if they do become superhumanly powerful. Even the leading developers are specifically saying that it could lead to human extinction. Sam Altman, the CEO of OpenAI, has said the worst case scenario is lights out for all of us. So it's, like, bizarre that they're chasing this goal with seemingly as much speed as they can muster while they don't really have a plan to make sure that things don't go totally awry. The best I can say is at least they know that they're playing with fire. This is a survey of AI researchers that just goes to show how much radical uncertainty there is in the field. 48% of the respondents these are people that had published in top journals in AI in the last few years. 48% of them gave at least a 10% chance of human extinction from AI. So half gave a 10% chance. For my money, that is very much worth taking seriously. Tons of this is another thing, mitigating the risk of human extinction. This was signed by all of the leaders. Sam Altman, the CEO of OpenAI, the CEO of Anthropic. Bill Gates is on there, the chief scientist from OpenAI, the chief scientist from Google DeepMind, the CEO of Google DeepMind. These 2 are Turing Award winners, which is basically like the Nobel Prize in computer science. So it's really a who's who of people signing on to the idea that mitigating the risk of extinction from AI should be a global priority. There's just really radical uncertainty, and all bets are really off, I think, for how the future is gonna shape up. How fast is this gonna happen? It's not that long of a time. The average there's a couple different ways you can ask the question, but the average among forecasters is maybe in the next few years, maybe it's closer to 8 to 10 years. Who knows? But that's the range in which the field is expecting these breakthroughs. Of course, it may never happen, but the consensus view is, like, a few to a handful of years from now. And so what will we be watching to figure out where we're going from here? Will this raw scaling continue to work? Will the scaling laws hold all the way to, sorry, to superhuman systems? That is very hotly debated. Will a new architecture come along and be even better than the transformer? Recent results suggest that yes. Actually, there will even be better architectures than the one that has taken us this far. Will mechanistic interpretability catch up? That's like the reverse engineering and figuring out what's actually going on inside. That was one thing that the survey of the 2,000 AI researchers did ask, the consensus was no, that we're not close enough to being able to understand what's going on inside to be able to count on that to save us, even though great progress is being made. What new modalities will matter? You saw the DNA one earlier. I think there's going to be a lot of kind of immediately superhuman capabilities coming out of AI when it comes to things like modeling cells, breeding brains, predicting the weather. This is the kind of stuff that humans just can't do, but we're already starting to see that AI systems can do it. And as soon as they can do it, they're pretty much immediately superhuman at it. Will these things start to become agentic? That means will they be able to have their own goals and pursue their plans over medium time horizons? That seems likely, and a lot of people are working on it. That's like you might think, gee, that's crazy. We wouldn't want to turn these things into, like, independent agents. I could point you to a 100 startups that are doing exactly that for all kinds of different use cases. Are these frontier labs, like your OpenAI's, your Google's, Anthropic's, are they going to pause their research or stop scaling if things start to get scary? They say they are. They've signed that statement. They have other various policies that they've put in place, but will they follow through on that? Obviously, we'll see. When will the government get involved? It's already starting to. I would expect a lot more focus on AI among not just the US government, all sorts of governments around the world. And what's going go on in China is another big one because people are really concerned that we might end up in a sort of AI arms race vis a vis China. And if it were to go that way, that would definitely be really bad. So my final thought is I love AI. I feel like I sound a note of caution here, but I love building apps. What AI has done for Waymark has been transformative. Our product is literally 10 times better than it used to be. I use ChatGPT tons of times every single day, and I love having these AI servants at my disposal. But I would just suggest that while we should all be figuring out how to use the current systems in our lives and in our work, the hyperscaling, the scaling up of 10x and 100x past where we've gone already is something that I think society would be wise to slow down on. Let's figure out what we have. Let's figure out how best to use it before we try to make something that is a superhuman AI scientist because I don't think we have the strategies that we will ultimately need to keep that under control and to make sure it goes well for us. If you wanna find more from me elsewhere, all these different places are good places to find more. So thank you. It is both energizing and enlightening to hear why people listen and learn what they value about the show. So please don't hesitate to reach out via email at tcr@turpentine.co, or you can DM me on the social media platform of your choice.