WEBVTT 00:00:00.000 --> 00:00:11.520 This is Essays on AI. I'm Tabrez Syed. Actually, no. I'm an AI voice reading his essays. But 00:00:11.520 --> 00:00:16.059 you probably guessed. This one is called, When the Map Runs Out. 00:00:16.059 --> 00:00:19.879 The opening section draws from Timothy B. Lee's excellent explanation of reinforcement 00:00:19.879 --> 00:00:25.959 learning at his newsletter, Understanding AI. In 2009, a graduate student named Stefan 00:00:25.959 --> 00:00:30.719 Ross at Carnegie Mellon University was trying to solve a problem that seemed straightforward 00:00:30.719 --> 00:00:36.400 – teach a computer to play Super Tux Kart, an open-source racing game similar to Mario 00:00:36.400 --> 00:00:42.340 Kart. His approach was elegantly simple. Ross would play the game himself, while his software 00:00:42.340 --> 00:00:47.279 captured screenshots and recorded which buttons he pressed. Then he'd train a neural network 00:00:47.279 --> 00:00:52.279 to predict which buttons Ross would push in any given situation. If the network could 00:00:52.279 --> 00:00:57.599 learn to mimic Ross' gameplay perfectly, it should be able to play the game itself. 00:00:57.599 --> 00:01:01.880 Just push the same buttons Ross would have pushed in the same situations. 00:01:01.880 --> 00:01:06.760 The initial results looked promising. The AI-controlled car would zip around the track, 00:01:06.760 --> 00:01:11.400 taking turns at reasonable speeds, staying roughly in the center of the road. For a few 00:01:11.400 --> 00:01:16.279 seconds, it looked like Ross had cracked the code. Then something would go wrong. The car 00:01:16.279 --> 00:01:21.000 would drift slightly to the left, or take a turn a fraction too wide, or hit a small 00:01:21.040 --> 00:01:26.639 bump that pushed it off its ideal line. These weren't catastrophic errors, just tiny deviations 00:01:26.639 --> 00:01:32.879 from the perfect gameplay Ross had demonstrated. But tiny deviations became bigger ones. The 00:01:32.879 --> 00:01:38.680 car would drift further off course, then make increasingly erratic corrections. Within moments, 00:01:38.680 --> 00:01:43.400 what had started as a small mistake would cascade into complete chaos. The animated 00:01:43.400 --> 00:01:48.120 vehicle would careen off the track entirely, tumbling into the virtual abyss while Ross 00:01:48.360 --> 00:01:53.239 watched in frustration. The problem wasn't that the neural network was stupid. It had learned 00:01:53.239 --> 00:01:59.320 Ross' driving patterns remarkably well. The issue was more subtle and more fundamental. Ross was a 00:01:59.320 --> 00:02:04.040 pretty good SuperTuxCart player, which meant his car spent most of its time near the center of the 00:02:04.040 --> 00:02:10.520 road, driving at reasonable speeds, taking smooth turns. The neural network had learned to drive 00:02:10.520 --> 00:02:15.720 perfectly, as long as everything went perfectly. But it had almost no training data showing what 00:02:15.720 --> 00:02:21.240 to do when things went wrong. When the car drifted off course, it found itself in situations 00:02:21.240 --> 00:02:26.199 that weren't well represented in its training data. And in those unfamiliar situations, 00:02:26.199 --> 00:02:31.160 it made increasingly poor decisions. Ross had discovered something that would become crucial 00:02:31.160 --> 00:02:36.759 to understanding artificial intelligence. Systems trained on past experience work beautifully 00:02:36.759 --> 00:02:41.559 until they encounter situations that weren't in their training data. Then they don't just fail, 00:02:41.559 --> 00:02:45.960 they fail catastrophically, with small errors compounding into complete breakdowns. 00:02:46.759 --> 00:02:52.360 Ross, who now works on self-driving cars at Waymo, formerly Google's self-driving car project, 00:02:52.360 --> 00:02:57.160 had identified a fundamental challenge that would prove relevant far beyond video games. 00:02:57.160 --> 00:03:02.199 This wasn't just a quirk of Ross' particular approach. It's how most AI systems learn, 00:03:02.199 --> 00:03:07.240 by studying massive amounts of data and finding patterns. Language models learn by processing 00:03:07.240 --> 00:03:11.960 billions of text examples. Image recognition systems train on millions of labeled photos. 00:03:12.520 --> 00:03:17.000 The fundamental approach is the same. Show the system enough examples, and it will learn to 00:03:17.000 --> 00:03:22.600 recognize and respond to similar situations. But Ross' SuperTuxKart car revealed the Achilles' 00:03:22.600 --> 00:03:26.520 heel of this approach. What happens when you encounter something that wasn't in the training 00:03:26.520 --> 00:03:31.880 data? Ross' problem wasn't unique to video games. The same challenge was playing out in the real 00:03:31.880 --> 00:03:36.919 world, where the stakes were considerably higher than a virtual car tumbling into a digital abyss. 00:03:37.240 --> 00:03:42.679 Google had started its self-driving car project in 2009, the same year Ross was wrestling with 00:03:42.679 --> 00:03:47.880 his racing game. In typical Google fashion, they took a Google approach to the training 00:03:47.880 --> 00:03:53.960 data problem. They decided to collect massive amounts of it. For nearly a decade, from 2009 00:03:53.960 --> 00:03:59.880 to 2018, Waymo vehicles drove millions of miles with human safety drivers behind the wheel, 00:04:00.520 --> 00:04:05.720 meticulously recording every scenario, every edge case, every moment when human judgment was 00:04:05.720 --> 00:04:11.880 required. The strategy worked. By the time Waymo launched its first commercial service in December 00:04:11.880 --> 00:04:18.200 2018, their cars had logged over 10 million miles on public roads. The vehicles that had once 00:04:18.200 --> 00:04:24.440 struggled with basic scenarios, like Ross' SuperTuxKart model drifting off course, more data had largely 00:04:24.440 --> 00:04:30.119 solved the SuperTuxKart problem. Where Ross' model failed after encountering a single unfamiliar 00:04:30.119 --> 00:04:35.160 situation, Waymo's cars could handle thousands of edge cases they'd seen during their extensive 00:04:35.160 --> 00:04:41.399 training period. The compounding errors that plagued early AI systems became increasingly rare 00:04:41.399 --> 00:04:47.079 as the training datasets grew more comprehensive. This approach represents how we've tackled most AI 00:04:47.079 --> 00:04:53.720 challenges. Identify the gaps in training data, then systematically fill them. Can't recognize cats 00:04:53.720 --> 00:05:00.679 in photos? Train on more cat images. Struggling with medical diagnoses? Feed the system more patient 00:05:00.679 --> 00:05:06.679 records. Having trouble with language translation? Process more bilingual text pairs. Today, we've 00:05:06.679 --> 00:05:12.200 taken this data collection strategy even further. AI systems can now manufacture their own training 00:05:12.200 --> 00:05:18.200 data through synthetic data generation. Need more examples of rare medical conditions? AI can 00:05:18.200 --> 00:05:23.640 generate realistic patient scenarios. Want to train a model on edge cases that rarely occur in real 00:05:23.640 --> 00:05:29.239 driving? AI can create thousands of synthetic scenarios with unusual weather conditions, 00:05:29.239 --> 00:05:35.399 unexpected obstacles, or complex traffic patterns. This ability to generate synthetic data means we 00:05:35.399 --> 00:05:40.040 can fill gaps in training datasets more systematically than ever before. It's a 00:05:40.040 --> 00:05:45.399 powerful strategy, and it works remarkably well for a specific type of problem. Statisticians 00:05:45.399 --> 00:05:50.600 call this epistemic uncertainty. Uncertainty that exists because we don't know something, but we 00:05:50.600 --> 00:05:56.359 could learn it if we gathered more information. Ross' SuperTuxKart model suffered from epistemic 00:05:56.359 --> 00:06:01.160 uncertainty. It didn't know how to recover from mistakes because it had never seen examples of 00:06:01.160 --> 00:06:07.320 recovery. Waymo's early struggles were also largely epistemic. Their cars couldn't handle construction 00:06:07.320 --> 00:06:14.119 zones because they hadn't seen enough construction zones. But epistemic uncertainty is solvable. You 00:06:14.119 --> 00:06:19.799 can collect more data, train on more examples, and gradually fill in the gaps. And here's where AI 00:06:19.799 --> 00:06:25.320 systems will eventually surpass human intelligence. They can process vastly more data than any human 00:06:25.320 --> 00:06:30.519 could ever experience. While a human driver might encounter a few dozen construction zones in their 00:06:30.519 --> 00:06:35.880 lifetime, an AI system can learn from millions of construction zone scenarios captured by thousands 00:06:35.880 --> 00:06:42.440 of vehicles. The epistemic uncertainty gap between humans and AI is temporary. Humans currently have 00:06:42.440 --> 00:06:47.399 an advantage because we've been training on visual and spatial data since birth, accumulating decades 00:06:47.399 --> 00:06:53.000 of experience navigating the world. But AI systems are rapidly catching up, and they'll eventually 00:06:53.000 --> 00:06:58.040 have access to far more comprehensive datasets than any individual human could process. For 00:06:58.040 --> 00:07:04.440 epistemic uncertainty, the solution is clear. More data, better algorithms, and time. Here's where the 00:07:04.440 --> 00:07:10.440 story takes a turn. Our predictions of the future are reflections from our rearview mirror. Even with 00:07:10.440 --> 00:07:15.640 over 20 million miles of real-world driving data, Waymo's systems still encounter situations that 00:07:15.640 --> 00:07:20.519 weren't in their training. Not because of gaps in the dataset, but because the future draws from a 00:07:20.519 --> 00:07:26.839 vast theoretical space of unrealized possibilities. Consider the microbursts that hit Austin in 2025. 00:07:27.559 --> 00:07:32.040 A sudden downdraft brought down trees and scattered debris across roads in patterns no 00:07:32.040 --> 00:07:37.959 training dataset had ever captured. Or the early days of the COVID-19 pandemic, when governments 00:07:37.959 --> 00:07:43.799 implemented lockdown policies based on historical precedent from vastly different eras. In this 00:07:43.799 --> 00:07:49.880 theoretical space, tree limbs can land on roads, locusts can swarm highways, microbursts can 00:07:49.880 --> 00:07:55.160 scatter debris in patterns no algorithm has ever seen. These aren't just rare events, they're 00:07:55.160 --> 00:08:00.839 genuinely unprecedented combinations that emerge from the complex interactions of countless variables. 00:08:00.839 --> 00:08:07.000 These events represent aleatory uncertainty, from the Latin word for dice, the inherent randomness 00:08:07.000 --> 00:08:11.959 and unpredictability in the world itself. Even our synthetic data generation can only create 00:08:11.959 --> 00:08:17.079 variations of what we can imagine from past experience. We extrapolate from history, but the 00:08:17.079 --> 00:08:22.519 future contains genuinely novel combinations that exist in theoretical possibility but have never 00:08:22.519 --> 00:08:28.760 manifested in any dataset. You could collect data for eons, train on every scenario from the past, 00:08:28.760 --> 00:08:34.280 and still encounter situations that have never existed before. This is because there is no data 00:08:34.280 --> 00:08:39.479 about the future, not because we haven't collected enough data about the past, but because the future 00:08:39.479 --> 00:08:45.080 contains genuinely novel combinations of factors that have never existed before. Here's what makes 00:08:45.080 --> 00:08:50.359 this distinction important. Both human intelligence and artificial intelligence are sophisticated 00:08:50.359 --> 00:08:55.719 pattern matching systems. We excel at recognizing situations we've encountered before and applying 00:08:55.719 --> 00:09:01.239 lessons from past experience. But when faced with genuinely unprecedented events, both types of 00:09:01.239 --> 00:09:06.840 intelligence hit the same fundamental wall. Consider how humans handled the COVID-19 pandemic. 00:09:06.840 --> 00:09:12.599 We had data from previous pandemics, the 1918 flu outbreak, various smaller disease outbreaks, 00:09:12.599 --> 00:09:17.880 historical accounts of quarantine measures. But the world of 2020 was fundamentally different 00:09:17.880 --> 00:09:23.960 from 1918. Our interconnected global economy, modern healthcare systems, digital communication 00:09:23.960 --> 00:09:30.440 networks, and social structures created a context that had never existed before. The result? Human 00:09:30.440 --> 00:09:35.799 decision makers implemented policies that were essentially educated guesses. Lockdown strategies 00:09:35.799 --> 00:09:41.239 varied wildly between countries and regions. Some measures worked, others didn't, and many had 00:09:41.239 --> 00:09:45.080 unintended consequences that no amount of historical analysis could have predicted. 00:09:45.719 --> 00:09:51.080 We were pattern matching from incomplete and often irrelevant historical data, just like an AI system 00:09:51.080 --> 00:09:56.200 trying to navigate an unprecedented scenario. The same dynamic plays out in financial markets 00:09:56.200 --> 00:10:01.559 during genuine crises, in political systems facing novel challenges, and in any domain where the 00:10:01.559 --> 00:10:06.919 future presents combinations of factors that have never occurred before. Humans don't have some 00:10:06.919 --> 00:10:12.679 magical ability to handle aleatory uncertainty better than AI systems. We're just as dependent 00:10:12.679 --> 00:10:17.719 on pattern matching from past experience. This distinction matters because it challenges a 00:10:17.719 --> 00:10:23.080 common assumption about human versus artificial intelligence. We often think of humans as 00:10:23.080 --> 00:10:28.359 fundamentally superior to AI systems, possessing some ineffable quality that machines can never 00:10:28.359 --> 00:10:33.719 replicate. But when we separate epistemic uncertainty from aleatory uncertainty, a 00:10:33.719 --> 00:10:38.679 different picture emerges. For epistemic uncertainty, situations where more data and 00:10:38.679 --> 00:10:43.559 better pattern recognition can solve the problem, AI systems will likely surpass human performance. 00:10:44.200 --> 00:10:48.599 They can process more information, learn from more examples, and update their models more 00:10:48.599 --> 00:10:54.599 systematically than biological intelligence. For aleatory uncertainty, genuinely unprecedented 00:10:54.599 --> 00:10:59.320 events that emerge from novel combinations of factors, both human and artificial intelligence 00:10:59.320 --> 00:11:03.719 face the same fundamental limitation. Neither can predict what has never happened before. 00:11:04.520 --> 00:11:07.479 Neither can prepare for combinations of factors that have never existed. 00:11:08.280 --> 00:11:12.840 But humans have developed something valuable over millennia of facing the unknown—mental 00:11:12.840 --> 00:11:17.799 models and frameworks for navigating uncertainty itself. We've learned to reason by analogy, 00:11:17.799 --> 00:11:22.840 to break down complex problems into manageable parts, to consider multiple scenarios simultaneously. 00:11:23.479 --> 00:11:27.640 These aren't perfect solutions, but they're evolved responses to the fundamental challenge 00:11:27.640 --> 00:11:32.520 of operating without complete information. AI systems will likely learn these same mental 00:11:32.520 --> 00:11:36.760 models from the vast repository of human experience they train on. They'll absorb our 00:11:36.760 --> 00:11:40.919 approaches to uncertainty, our problem-solving frameworks, our ways of thinking through 00:11:40.919 --> 00:11:46.599 unprecedented situations. But they'll also inherit our biases, our blind spots, and our systematic 00:11:46.599 --> 00:11:51.719 errors. After all, they're learning from our exhaust, both the wisdom and the mistakes embedded 00:11:51.719 --> 00:11:57.239 in human decision-making. The future will always contain surprises that no amount of historical 00:11:57.239 --> 00:12:02.520 data can anticipate. But this raises a deeper question about the nature of intelligence itself. 00:12:03.320 --> 00:12:06.919 Is human decision-making reducible to a set of mental models and rules? 00:12:07.559 --> 00:12:12.520 Or is there more to gut feeling than the total sum of carbon-based neurons buzzing in our heads? 00:12:13.320 --> 00:12:17.799 Thanks for listening. The full essay, with links to Timothy B. Lee's explainer 00:12:17.799 --> 00:12:21.719 and Stéphane Ross's paper, is on the episode page at mandolivia.com. 00:12:22.440 --> 00:12:25.799 If you'd like more essays and audio, subscribe wherever you're listening. 00:12:25.799 --> 00:12:30.520 Essays on AI is a Mandalivia production.