No Data About the Future: The limits of pattern matching
How AI and humans both fail when facing the genuinely unprecedented
1In 2009, a graduate student named Stephane Ross at Carnegie Mellon University was trying to solve a problem that seemed straightforward: teach a computer to play SuperTuxKart, an open-source racing game similar to Mario Kart. His approach was elegantly simple. Ross would play the game himself while his software captured screenshots and recorded which buttons he pressed. Then he'd train a neural network to predict which buttons Ross would push in any given situation.
If the network could learn to mimic Ross's gameplay perfectly, it should be able to play the game itself—just push the same buttons Ross would have pushed in the same situations.
The initial results looked promising. The AI-controlled car would zip around the track, taking turns at reasonable speeds, staying roughly in the center of the road. For a few seconds, it looked like Ross had cracked the code.
Then something would go wrong. The car would drift slightly to the left. Or take a turn a fraction too wide. Or hit a small bump that pushed it off its ideal line. These weren't catastrophic errors—just tiny deviations from the perfect gameplay Ross had demonstrated.
But tiny deviations became bigger ones. The car would drift further off course, then make increasingly erratic corrections. Within moments, what had started as a small mistake would cascade into complete chaos. The animated vehicle would careen off the track entirely, tumbling into the virtual abyss while Ross watched in frustration.
The problem wasn't that the neural network was stupid. It had learned Ross's driving patterns remarkably well. The issue was more subtle and more fundamental: Ross was a pretty good SuperTuxKart player, which meant his car spent most of its time near the center of the road, driving at reasonable speeds, taking smooth turns.
The neural network had learned to drive perfectly—as long as everything went perfectly. But it had almost no training data showing what to do when things went wrong. When the car drifted off course, it found itself in situations that weren't well represented in its training data. And in those unfamiliar situations, it made increasingly poor decisions.
Ross had discovered something that would become crucial to understanding artificial intelligence: systems trained on past experience work beautifully until they encounter situations that weren't in their training data. Then they don't just fail—they fail catastrophically, with small errors compounding into complete breakdowns.
Ross, who now works on self-driving cars at Waymo (formerly Google's self-driving car project), had identified a fundamental challenge that would prove relevant far beyond video games.
This wasn't just a quirk of Ross's particular approach. It's how most AI systems learn—by studying massive amounts of data and finding patterns. Language models learn by processing billions of text examples. Image recognition systems train on millions of labeled photos. The fundamental approach is the same: show the system enough examples, and it will learn to recognize and respond to similar situations.
But Ross's SuperTuxKart car revealed the Achilles' heel of this approach: what happens when you encounter something that wasn't in the training data?
The Data Collection Solution
Ross's problem wasn't unique to video games. The same challenge was playing out in the real world, where the stakes were considerably higher than a virtual car tumbling into a digital abyss.
Google had started its self-driving car project in 2009, the same year Ross was wrestling with his racing game. In typical Google fashion they took a Google approach to the training data problem: they decided to collect massive amounts of it. For nearly a decade, from 2009 to 2018, Waymo vehicles drove millions of miles with human safety drivers behind the wheel, meticulously recording every scenario, every edge case, every moment when human judgment was required.
The strategy worked. By the time Waymo launched its first commercial service in December 2018, their cars had logged over 10 million miles on public roads. The vehicles that had once struggled with basic scenarios—like Ross's SuperTuxKart model drifting off course—could now navigate complex intersections, construction zones, and emergency vehicles with increasing confidence.
More data had largely solved the SuperTuxKart problem. Where Ross's model failed after encountering a single unfamiliar situation, Waymo's cars could handle thousands of edge cases they'd seen during their extensive training period. The compounding errors that plagued early AI systems became increasingly rare as the training datasets grew more comprehensive.
This approach represents how we've tackled most AI challenges: identify the gaps in training data, then systematically fill them. Can't recognize cats in photos? Train on more cat images. Struggling with medical diagnoses? Feed the system more patient records. Having trouble with language translation? Process more bilingual text pairs.
Today, we've taken this data collection strategy even further. AI systems can now manufacture their own training data through synthetic data generation. Need more examples of rare medical conditions? AI can generate realistic patient scenarios. Want to train a model on edge cases that rarely occur in real driving? AI can create thousands of synthetic scenarios with unusual weather conditions, unexpected obstacles, or complex traffic patterns. This ability to generate synthetic data means we can fill gaps in training datasets more systematically than ever before.
It's a powerful strategy, and it works remarkably well for a specific type of problem. Statisticians call this epistemic uncertainty—uncertainty that exists because we don't know something, but we could learn it if we gathered more information. Ross's SuperTuxKart model suffered from epistemic uncertainty. It didn't know how to recover from mistakes because it had never seen examples of recovery. Waymo's early struggles were also largely epistemic—their cars couldn't handle construction zones because they hadn't seen enough construction zones.
But epistemic uncertainty is solvable. You can collect more data, train on more examples, and gradually fill in the gaps. And here's where AI systems will eventually surpass human intelligence: they can process vastly more data than any human could ever experience. While a human driver might encounter a few dozen construction zones in their lifetime, an AI system can learn from millions of construction zone scenarios captured by thousands of vehicles.
The epistemic uncertainty gap between humans and AI is temporary. Humans currently have an advantage because we've been "training" on visual and spatial data since birth, accumulating decades of experience navigating the world. But AI systems are rapidly catching up, and they'll eventually have access to far more comprehensive datasets than any individual human could process.
For epistemic uncertainty, the solution is clear: more data, better algorithms, and time.
But There's No Data About the Future
Here's where the story takes a turn.
Our predictions of the future are reflections from our rear-view mirror. Even with over 20 million miles of real-world driving data, Waymo's systems still encounter situations that weren't in their training—not because of gaps in the dataset, but because the future draws from a vast theoretical space of unrealized possibilities.
Consider the Austin microburst from a few weeks ago. A sudden downdraft brought down trees and scattered debris across roads in patterns no training dataset had ever captured. Or the early days of the COVID-19 pandemic, when governments implemented lockdown policies based on historical precedent from vastly different eras.
In this theoretical space, tree limbs can land on roads, locusts can swarm highways, microbursts can scatter debris in patterns no algorithm has ever seen. These aren't just rare events—they're genuinely unprecedented combinations that emerge from the complex interactions of countless variables.
These events represent aleatory uncertainty—from the Latin word for dice—the inherent randomness and unpredictability in the world itself. Even our synthetic data generation can only create variations of what we can imagine from past experience. We extrapolate from history, but the future contains genuinely novel combinations that exist in theoretical possibility but have never manifested in any dataset.
You could collect data for eons, train on every scenario from the past, and still encounter situations that have never existed before. This is because there is no data about the future—not because we haven't collected enough data about the past, but because the future contains genuinely novel combinations of factors that have never existed before.
When Both Intelligences Fail
Here's what makes this distinction important: both human intelligence and artificial intelligence are sophisticated pattern-matching systems. We excel at recognizing situations we've encountered before and applying lessons from past experience. But when faced with genuinely unprecedented events, both types of intelligence hit the same fundamental wall.
Consider how humans handled the COVID-19 pandemic. We had data from previous pandemics—the 1918 flu outbreak, various smaller disease outbreaks, historical accounts of quarantine measures. But the world of 2020 was fundamentally different from 1918. Our interconnected global economy, modern healthcare systems, digital communication networks, and social structures created a context that had never existed before.
The result? Human decision-makers implemented policies that were essentially educated guesses. Lockdown strategies varied wildly between countries and regions. Some measures worked, others didn't, and many had unintended consequences that no amount of historical analysis could have predicted. We were pattern-matching from incomplete and often irrelevant historical data, just like an AI system trying to navigate an unprecedented scenario.
The same dynamic plays out in financial markets during genuine crises, in political systems facing novel challenges, and in any domain where the future presents combinations of factors that have never occurred before. Humans don't have some magical ability to handle aleatory uncertainty better than AI systems—we're just as dependent on pattern-matching from past experience.
This distinction matters because it challenges a common assumption about human versus artificial intelligence. We often think of humans as fundamentally superior to AI systems, possessing some ineffable quality that machines can never replicate. But when we separate epistemic uncertainty from aleatory uncertainty, a different picture emerges.
For epistemic uncertainty—situations where more data and better pattern recognition can solve the problem—AI systems will likely surpass human performance. They can process more information, learn from more examples, and update their models more systematically than biological intelligence.
For aleatory uncertainty—genuinely unprecedented events that emerge from novel combinations of factors—both human and artificial intelligence face the same fundamental limitation. Neither can predict what has never happened before. Neither can prepare for combinations of factors that have never existed.
But humans have developed something valuable over millennia of facing the unknown: mental models and frameworks for navigating uncertainty itself. We've learned to reason by analogy, to break down complex problems into manageable parts, to consider multiple scenarios simultaneously. These aren't perfect solutions, but they're evolved responses to the fundamental challenge of operating without complete information.
AI systems will likely learn these same mental models from the vast repository of human experience they train on. They'll absorb our approaches to uncertainty, our problem-solving frameworks, our ways of thinking through unprecedented situations. But they'll also inherit our biases, our blind spots, and our systematic errors. After all, they're learning from our exhaust—both the wisdom and the mistakes embedded in human decision-making.
The future will always contain surprises that no amount of historical data can anticipate. But this raises a deeper question about the nature of intelligence itself.
Is human decision making reducible to a set of mental models and rules? Or is there more to gut feeling than the total sum of carbon-based neurons buzzing in our heads?
This opening section draws from Timothy B. Lee's excellent explanation of reinforcement learning at Understanding AI.