Escaping Local Maximums: AI, Startups, and Personal Growth with Judah Taub

Speaker A: Foreign. Welcome to the Unrivited podcast, where we talk about artificial intelligence, digital transformation, and people. And as usual, this podcast is brought to you in part by our book, AI in a Weekend. John, do us a favor and introduce our guest.

Speaker B: Right. Thank you, Martin. All right, so, uh, as is tradition these days, we are joined by, uh, an extra special guest, Judah Taub, who is the managing partner at Hetz Ventures, as well as the author of how to Move up when the Only Way Is Down. Uh, so, Judah, thank you for joining us today. Typically, we try to keep things pretty conversational here on Unrivited, but, uh, sometimes we do have a few, uh, prepared questions for you. Let's, uh, jump in to the conversation. Or maybe I should turn it over to you for a little bit, if you want to just tell us a little bit about yourself before we get into the, uh, conversation.

Speaker C: So my claim to fame is that I've known Martin for a while, and I'm a big fan and I'm really happy to be here.

Speaker A: Wow, wow, wow.

Speaker C: I'm humbled that one out.

Speaker A: You know what? I'm going to quote that up on, uh, LinkedIn for you. Instead of as seen on YouTube, it will be as seen on LinkedIn. All right, Judah, thank you for, um, forwarding a copy of your book. Uh, I know I got it on Saturday and I scanned it, and it was fabulous. I'm going to have to actually read the whole thing cover to cover, but I did rip open to a few sections and found some fascinating places to start. Let me hit you right away with, um, some understanding concepts of local maxim maximum. Sorry, um, can you explain the concept of local maximum and how it applies to personal and professional growth?

Speaker C: Amazing. And so you both know this. But for those folks who don't know what a local maximum is, it comes from machine learning. It comes from mathematics. Uh, most data scientists will be aware of the concept. They typically call it local minima, but it's the same thing. The easiest way to sort of imagine it is imagine that you're dropped off in the middle of a desert, and your goal is to reach the highest mountain. That's your goal. You're in the middle of a desert, you don't know which mountain is the highest. So you're looking around, you're walking, you see sort of one mountain, another mountain. Suddenly you see a very big mountain, and you're really happy. You're like, this is the mountain that is going to take me. The highest, highest, in this case, could be the most money, the most happiness. The most, whatever we're trying to achieve. But you're climbing up the mountain. And when you reach the top of the mountain, you suddenly realize that behind the mountain you climbed or somewhere in the distance, there is a taller, a, uh, higher mountain than the one you're on. And the trap you've fallen into or the position you're now in is that you're in a local maximum. You had a point, which is a maximum. Why? Because if you take one step back, forwards, left or right, it will definitely be down, meaning you're at the top of a peak. You can only go down from here. But it's a local rather than a global maximum because it's not the tallest. It's not the best outcome you could have hoped for. You can see there's a better one. And people run into local maximums in a whole variety of ways. Whether it's in your profession where you suddenly say, one second I'm at a job, I realize I'm approaching my local maximum at this job, I can't get much better, or it's not that much better than where it's going to get to from here. But actually I could see in the distance a job that would make me much better, happier, or give more to society or whatever it is the playing field that I'm optimizing for. I see a higher mountain. Um, it could be at a startup, where your startup is sort of slowly plateauing or reaching a point where you can see that you're climbing a mountain. You don't have to be at the very top, but you can see you're approaching a top or even just on a mountain, and there's a better one elsewhere. And we can go into examples, whether it's in history or today, where people suddenly realize they're on the wrong mountain. And so the idea for me, somebody who spends a lot of time with entrepreneurs, uh, especially like in the evenings when they're winding down and they're asking, like, the bigger questions, are we heading in the right direction? Over time, I felt I was having a very similar conversation with founders who come from very different backgrounds and operating in different environments. And that is, how do they, they're the ones calling the shots in their own business, navigate their business towards, ideally, a, a global maximum, or at least away from very local maximum traps. And the whole idea behind the book was, are there techniques that we can learn from, specifically machines and AI for how it tries to avoid local maximum? So that's sort of 60 seconds in a nutshell.

Speaker A: Speak of nut Jobs myself. I mean, I think I hit my local maximum behind me in, in reference to the image of, of my four kids. Uh, I don't think I want to go any higher than that. If you don't mind segueing, ah, a little bit beyond, um, my personal local maxim. Lessons from artificial intelligence. Maybe, um, you could talk about some techniques used in AI to overcome the local maxima as they apply to human decisions. Give an example, maybe.

Speaker C: Yeah. Amazing. So, like, taking a step back, the passion for me for this specific book was that, uh, we humans talk a lot about decision making and we realize that decision making is really important and can make a big difference on outcomes of whatever we're making those decisions in. And at the same time, specifically AI or in general computers have gotten to a point where that in many ways it makes better decisions than we are. I'm not talking about better calculations, because calculations already, I don't know. A calculator could calculate more and more accurately than a human, but we're talking about actual decisions that computers are making. More often than not, computers are learning to make better decisions. And so let's just take an example. Um, think about chess players. So a grandmaster training at chess used to train against another grandmaster or read books on best moves, etc. Today, more and more grandmasters are, uh, using software to tell them, at this point, the probability of you winning with this move versus this move increases. You should be looking at these three moves that you weren't looking at before. Does it mean you have to listen to the software? No, not always. But they're learning to have like an AI assistant who's telling them you should be aware of decisions, not calculations, decisions you should be thinking about making. And so my idea was, is there some type of wisdom that computers have learned? Because once upon a time we were teaching computers logic, if not while, etcetera, are, uh, there certain wisdoms in that computers have learned that we can now implement? And I like most things, it's always best to sort of narrow things down. So I entirely focused the book on the problem of avoiding local maximums. And I took roughly 10 techniques that data scientists slash computers use. And I asked, could we humans replicate these techniques in our own lives? So that's the broad thinking. I'm happy to give an example just to sort of bring this home, if that makes sense.

Speaker A: Yeah, absolutely.

Speaker C: Like the first technique that I talk about, and I think it's literally like each chapter is a technique, um, is what I would call ABX testing versus a B testing. So one of the things that everybody's probably familiar with is a B testing. You're not sure about which path to take. You do AB testing, and then you say, this one scores better. Let's go back to that mountain climber. Um, if I was the mountain climber and I was like, I'm not sure if I should take one step back or one step forwards, a B testing would be one step forward versus one step back. Which one is higher? Take the step in that direction. And you can imagine the computer program that literally says, compare one step forwards, one step back. Whichever way it's higher, take that step. There probably isn't a better computer algorithm for finding yourself trapped in a local maximum than that one. What will naturally happen, you can imagine sort of the computer bot climbing the first mountain it finds. It will naturally just climb and get stuck at the top. And so the story I tell at the beginning of the chapter there is one of an American hero, in my opinion, that most people actually haven't heard of. And that is Dick Fosbury. Um, he was the high jumper, wasn't that good, won the gold Olympic medal. And the reason I like him even more is he was an engineer. And he talks about how in college he realized that he was, like, good at high jumping, but he wasn't even the best in his college. It's like that's how good he was at high jumping. He wasn't even the best in his college. And how did he end up, only a few years later, winning the gold medal is. Everybody was AB testing. So everybody was going home and saying, how can I gain an incremental half a centimeter? How can I gain an incremental sort of a quarter of a centimeter. And he was going home and saying, I'm never going to win if I just do incremental gains. And so he was drawing out, and he's got these, like, tiny pictures that he was drawing out sort of university of completely new ways he could potentially jump over the bar. And some of them were terrible. And he describes them, et cetera. And. And he was basically going down the mountain. He was going down, looking across the landscape, not climbing, just ab testing the hell out of the mountain he happens to be on. And he was searching for a new mountain while he was ab testing that mountain. And so if we go back to computers a second, and, yeah, like, literally a few years later, he wins just the golden medal. And just to finish the story, because it's quite cute, the next Olympics, not only does he not win again, I don't think he even makes the team. Because once everybody realizes there's a better mountain and this is how you should be climbing, suddenly you got to remember he's actually not such a great jumper anymore. So everybody completely destroys his new world record. And sort of, you could see how they break that record again. So going back to computers, what computers are learning to do more and more is not AB testing, but ABX testing. So when you're doing a Google search, when you're querying, I don't know, Walmart's website, and they're trying to work out what they should be putting on, um, your screen on the recommendation engine, they know they can't test every option. So they start with a good option and they a B test it. But along the way, literally every so often, M, they'll put in an X something completely different. Amazon's routing, Amazon's prime routing. It's the, it's the salesperson problem. They can't calculate every single route. So we'll do ab, AB testing. But every so often they'll try a completely different route. And if they see this new route scores fairly high, what does it tell the algorithm? We may not be on the right mountain. Let's explore more of the other. So routes that look much more like that one. And if it scores very poorly compared to my A B testing, I'll increase the A B testing that I'm doing right now. So it's quite a long example I've given here, but hopefully gives some idea of how computers are now implementing things that not only Dick Fosbury should be using to win Olympic medals, but hopefully some of us should be doing when we're not just AB testing, but adding X's into our decision making.

Speaker A: Excellent, Excellent. I'm going to put this in perspective and not flop in my career if I can help it.

Speaker B: I don't think you have anything to worry about. I think you've done, done a great job. So, uh, we. Everyone's entitled to a few flops here and there, right? So, um, we won't bring those ones to attention if they've occurred. So, uh, you know, they have. Okay. I don't know what they are. So I, you know, the mystery lies, uh, with you. But, um, Judah, some of the things that you mentioned there, you know, there's a lot to unpack, obviously. I think, you know, a lot of what you mentioned in terms of like ABX testing, uh, is, is analogous to like reinforcement learning like we see nowadays. Especially like reinforcement learning, uh, with human feedback, uh, that we use to train a lot of large language models, uh, today or even, um, kind of add that extra performance boost, um, for, for some of these, um, models if they're applied to, you know, particular applications. So you definitely see that, uh, even when, you know, I think of the analogy of, of, you know, being at a local maximum and then, um, you know, having to basically traverse down the hill, uh, in order to climb the next, you know, you know, achieve the global maximum or maybe even a higher local maximum. So, you know, thinking about it from, I guess the perspective of machine learning or AI, but, um, you know, going, I guess, continuing down that vein and I'm glad you gave us one of, you know, the example of, uh, the fosberry flop. So do you think that, you know, we're seeing any diminishing returns, um, in AI development today? Um, and if so, you know, what do you kind of believe is maybe the path forward?

Speaker C: Okay, so, I mean, it's a good question. Uh, I'll answer it sort of with, with something else, which I think is important. So one of the things that pisses me off. Let's start with that way. One of the things that gets me a bit annoyed is, is when companies refer to their tech teams as R and D. We have 150 people in R&D, we have 5,000 people in R and D. We have whatever. Um, and the reason that's so annoying to me is because the R and the D could not be more different. Okay? So the researchers are the ones who may fail. And actually if they aren't failing often, they're probably not doing their work correctly. These are, uh, individuals who are looking for higher mountains. They are trying to get their company, or maybe humanity, depends, whatever they're looking into to shift from the playing field we think we're in to a higher one to a better one to one with more opportunities. And so in a good research environment, you're seeing people floating around writing on the walls. That's what you're paying them to do. And every so often somebody will come up with something which will be like a thousand X better than everything else, like at least 10X but potentially a thousand. The developers could not be a, uh, more different creature. Okay? The developers work as, you know, in sprints. Like it's completely different. Three weeks, three weeks, three weeks. Every day, sort of line to go churn out. Now those guys are the ones who are the climbers. These are the engineers who are putting things into practice and are moving up, uh, the mountains. We're on and actually turning products out, making sure they don't fail. And I'm not here to say which one's more important because. Because you need both. Every company and every individual in your own life should ask yourself, have I got the right balance that I want between my researcher self sort of wandering off looking for potential new mountains, doing wacko ideas. And if you haven't got these big failures, uh, which Martin alluded to, then you probably haven't done enough sort of X's to reach amazing successes as well. And then every individual should not just be floating around. And so same thing with a company. So to your point, where I think we are in AI is because things have happened so quickly and things are literally like every six months things are moving ahead so much that a model or an idea from six months ago is like 10 years old in the previous sort of tech world. Uh, so I think there's a lot to be done both in the researcher world, which is looking for higher mountains, and the developer world. In the developer world, I would say most of it is things like you said, like fine tuning models or just making this actually provide value to an S&P 500 company. There's so much out there which is amazing, but like actual product generating real tangible value. I mean you talk to practitioners and they're like, yeah, yeah, I know that this version gazillion and one, I'm just trying to make something work and get it through security. Clients like that is all engineering and we have a ton of that because we've gotten to a enormous mountain so quickly that actually the engineers have a lot to do here. And picking what they want to do first and how quickly they're able to climb is a huge task that can literally take a number of years and each step will be significantly sort of up. So I don't think we're at a plateauing place yet on that front. On the researcher side, I think there is, I mean if it's not something that could fail miserably, it is probably not a huge sort of step up function as well. So like I would say the stuff that could really take us uh, from here to an absolutely new place is not going to be a few more sort of GPUs that you're running something on or eating the Internet twice over rather than once. It's going to be something of a completely different nature. Feeding it data to do with like physical objects, letting it wander around with a robot to collect data by, it's going to have to be something that is Most likely going to fail, but if it works, brings us to a, uh, new playing field. And the majority of the people are not doing that right now. Uh, they're talking about it, but the majority are actually just trying to get something working in production. Um, so to your answer. Um, I have this pet peeve between research and development. I think most of the benefit we'll see soon is on the development side. And I'm always curious to hear people doing sort of actual things on the research side that I think will shift. Most of them I wouldn't call step up functions because most of them haven't got a high enough chance of failure, in which my argument is that it's not a big enough step up.

Speaker A: That's a great, um, great way to look at this. And, and by the way, I share the same, um, belief on R and D M, because I've been on the R side and I've been on the D side. And you really do need to fail and you have to fail fast to know to change directions on the R side and then overlay my D side, which is my favorite part of my life. Um, there's, there's a lot going on in unpack. Um, um, just a quick, easy, um, thought here. I'm going to share or ask specifically about startups because I mentor at startups and one of the fun things to get into is how can startups identify, or do they really need to identify their local maximum to achieve sustainable growth? I mean, that's an interesting way to look at it just in a general sense.

Speaker C: So I don't think you ever know if it's a local maximum or global one. I think you always sort of, because the playing field is potentially bigger than whenever you describe it. So it's impossible to say that. So like in a geeky, technical way. But I think the key question is, are you climbing a mountain that you are satisfied with, or are you concerned that there's a bigger mountain that will suddenly dwarf yours? And, um, examples of that would be like Blockbuster versus Netflix. So like, Blockbuster were, uh, at the top or very near the top of the video rental mountain. And there's a beautiful quote from the CEO of Blockbuster. Um, and this was literally like a couple of months or maybe a year before Blockbuster went bust. And he said, I am fascinated by this ongoing sort of whatever it is with this company Netflix, because there literally isn't a single thing they have that we don't already do. And the truth is, I think he is 100% right. Everything Netflix had So did Blockbuster. Uh, they had streaming. They had everything. The big difference is, if I have to describe it, Blockbuster were at the top of this mountain, which was limited and capped, and Netflix were at the bottom, but climbing a completely different mountain. And, um, Blockbuster was so stuck up there with Capex, with stores with late payments and everything that for them to get onto the Netflix mountain was going to be so painful. Coming down off where they were and climbing this new mountain. And Netflix are this young, agile sort of company which doesn't have all the baggage from that mountain. So they can climb what turns out to be a mountain which is 100 times bigger and much more relevant. So, uh, to your point, Martin, I think you never know. I think as a founder especially, your job is to spend a decent portion of your time with your eyes. Call it binoculars, night vision, whatever we want it, scouting around and m. Making sure that you have a clear idea of what's happening around you. And it's, and it's one of the most important jobs of the founders to keep your head above the water and make sure you know where you're going.

Speaker A: Excellent. Excellent. And, uh, speak of keeping your head above the water and using your binoculars correctly, I can appreciate that. And for those listening to this that don't have the recall, uh, or knowledge, Netflix started with postal mailing those DVDs to your home. I know that's so foreign, uh, to most people today, but that's how they started. And I was an original Netflix customer before streaming. I know I'm that old.

Speaker B: Hey, me too. I, I remember getting the, uh, the DVDs as well. So, um, Yeah, I, that was, uh, I mean, that was the sole distribution for quite a few years when they started out, wasn't it? I mean, streaming wasn't really.

Speaker A: Not yet.

Speaker B: Yeah, thought yet. Okay.

Speaker A: Yeah.

Speaker B: So, you know, you're. I, I'm right there. I'm right there with you, Martin. But, you know, most people might not even know what a DVD player is. It's like trying to describe a laser disc player, uh, and notice someone back in the, the early 90s maybe, or maybe it was even earlier than that. So.

Speaker A: Well, for the, for the youngins listening, it's the, it's the place on an old computer that you push the button and the cup holder comes out to, to put your cup.

Speaker B: I bet they, most of them don't even have those on there anymore. So, uh, that's a stretch. So, going back into, uh, thinking about, um, specifically, you know, I want to talk about a, uh, issue that came up recently everyone's been talking about Deep Seek as kind of this big, you know, blockbuster, I guess, pun intended, uh, of AI models that came out and you know, the whole industry's rattled by it. And I think they said like the US stock market or tech stocks based on AI models lost like a trillion dollars in value or something because Deep Seat came out with comparable performance, uh, but obviously costing a lot, a lot less. Um, and then through some maybe cursory level investigation, I'm not sure where it stands right now, basically the accusation of Deep seq, uh, taking, uh, some of the training data, uh, or distilling down, uh, the OpenAI 01 or I think maybe it was the O1 model, but basically taking AI, uh, OpenAI's model and repurposing it under their own banner, uh, so kind of avoiding the entire training, uh, element to the capacity that you would having to train something like that from scratch. Anyway, a couple of questions here. Assuming, you know, that that happened and that OpenAI is complaining of some other company using their training data, you know, is there some level of irony in there that OpenAI also, you know, basically didn't ask anyone to use the data from the Internet that they used to train their models? Um, but I think the second question, uh, which is more important for me, and I think this goes back to the R and D discussion is, you know, should we really find that these types of events are, treat them as negative events, uh, when, if we think about the progress of science, science is based off of, you know, conversations and critique, um, and open, you know, dialogue where, you know, sure, someone might be the first one to rush and try to print a paper out and get their name on it. But ultimately scientific progress, and in this case AI progress, you know, could be hindered by companies like OpenAI saying, you know, don't take our model and, and don't use our, our, um, you know, our, our intelligence or our information, uh, to, you know, try to build something better. So, uh, you know, kind of, what do you think about the, uh, the elements of progress in AI and should we really be trying to hinder organizations from, you know, building off of other organizations? Or maybe. It's probably a very poorly worded question. This is one of my very long winded questions,

Speaker C: so I'll make two comments. The first and easy one is it's extremely, extremely difficult, borderline impossible to stop technology. I remember when I was head of data at a fairly large hedge fund and um, there was a debate and we brought in some lawyers to discuss Whether it is or isn't allowed to scrape websites. And, and like you look at what happens today, regardless of whether companies are admitting it or not, I think there's a function built into Excel that allows you to automatically scrape websites like it. It's just like you, you can, you can legislate a lot of this, but it's, it's really difficult. So, especially in the software world. So I don't know how this will play out and if it will between deep sick and OpenAI, swing them, etc. But that leads me to the second point, and that is, I think there's a very, very deep level sort of takeaway, and that is one that I would again put as a key technique for trying to overcome local maximums, and that is to prioritize agility over muscle. Okay? And we're seeing this again and again and again in technology and more often than not in business as well. If I go back to the example of Blockbuster versus Netflix, blockbuster are the muscle and Netflix with the, uh, agile. If you go to a gym and you imagine the guy in the gym who goes seven days a week and all the guy does is bench press and with like maximum amount of sort of weights on it and just does like three bench presses, goes home next day three, versus the guy who's going around the gym and doing a little bit of this, a little bit of that, a little bit, and sort of trying and training his bottle. Now, if the competition between the two individuals is bench pressing, it is very obvious who's going to win. Okay, it's the first guy who is just doing that. But if it's a completely different competition, and I'm stressing the fact, completely different because we're in a world where it is extremely difficult to predict what will be happening in one year's time, never mind five years time. So the competition between the two now is not in the gym. It's actually hurdle jumping. It's clear none of them prepared for hurdle jumping, I'll give you that, but we all know who's going to win at the hurdle jumping. Okay, it's the second one. It's going to be the agile one. We saw this with our startups in Covid. Nobody predicted Covid. Obviously some of them were naturally better placed because they were in sort of industries that Covid sort of boded well for them, and some of them were in industries that Covid was really difficult to adjust to. But if you ask me for the number one trait that was a predictor of which startups were going to thrive through Covid and which ones were going to find it difficult. It was the agile ones who were open to constantly adapting and to bring this to deep seek, et cetera. We're seeing this with algorithms and AI as well. If you look at GPT 1, 2, 3, 4, you will slowly, I mean, everything is increasing. The number of neurons, the number of layers, the number of everything but the number of neurons, meaning the size, the agility ness of the algorithm versus the number of layers that they're sort of crunching, which is so called the muscle. The agility is increasing dramatically faster than the muscle and along comes deep sequence is actually we can get to as good of a score and they're getting rid of muscle again. So if you want to train an algorithm or if you want to train a human, or if you want to train a business to do really, really, really well at one specific thing, that's fine. But if you want to try and make it work over a long period of time, and especially in an unpredictable world that is changing so dramatically from one day to the next, prioritizing agility, call it the opex versus the capex, call it the sort of the flexibleness versus the muscle, you can give it whatever title you want over time typically proves itself. So the takeaway for me, for our companies, or even for myself, is to try and make sure that we have a more agile approach to hopping between sort of imagine the mountains and suddenly there's an earthquake. So one mountain goes down a lot, one mountain goes up a lot. That's literally what happens when technology shifts the landscape. You want to be that human who may not be as good at climbing the biggest boulder, but is able to at any point in time, sort of go down and bounce back up sort of fairly easily.

Speaker A: Excellent. Um, well, Judah, I want to thank you for being on our unrivited, uh, conversation today. And I believe we would have much to talk about for another, another time, maybe potentially later this year or into next year. It'd be fun to pick up on, uh, the next level of topic. But, uh, from on behalf of our listeners who were able to watch or just listen to us, I think we have a big thank you. I want to say thank you for being on our revited podcast.

Speaker C: It's great being here. Martin, thank you very much.

Speaker B: Thanks, Julia.

Speaker A: Thank you, Judah.