On May 6, 2010, the Dow Jones dropped nearly 1,000 points in minutes. A trillion dollars vanished. And then the market bounced back. The whole thing took thirty minutes.
Today, we call it a flash crash. Flash crashes happen when unusual trades trigger chain reactions across interconnected algorithms. Each algorithmic trader behaves as trained. But the collective behavior spirals into something catastrophic. The market is a collective intelligence that breaks down in the face of unusual situations.
I spent 13 years in capital markets, supporting traders and deploying models in volatile FX environments. I was there in 2008 when credit markets froze, and in the 2010s when high-frequency algorithmic trading took over capital markets, leading to many small and large flash crashes.
Today, I work as a technology consultant supporting companies in their AI initiatives. A question I hear constantly: “What types of problems can AI actually solve?”
The honest answer: it depends. And also, how ready are you to face flash crashes?
Nassim Nicholas Taleb has a useful frame for thinking about this – Mediocristan and Extremistan.
Mediocristan and Extremistan
Taleb uses these terms to describe two different environments.
Mediocristan is where patterns are stable, and the past is a reliable guide to the future. Think of measuring human height. You’ll get a bell curve. No single person will be tall enough to affect the average. Outliers exist, but they don’t dominate.
Extremistan is where rare events have an outsized impact. Think of book sales or wealth distribution. A single outlier (Harry Potter, Jeff Bezos) can dwarf the sum of everyone else. The past doesn’t prepare you for what’s coming because what’s coming might be unprecedented.
AI systems are trained in Mediocristan. They learn statistical regularities from historical data. They work beautifully when deployed in Mediocristan. They break when deployed in Extremistan.
Another metaphor can also help us think about *what* AI does: the epicycle.
The Epicycles Problem
Rohit Krishnan’s recent essay “Epicycles All The Way Down” provides a helpful framework to think about how LLMs work.
Before Newton, astronomers predicted planetary motion by adding “epicycles” (circles on circles) to Ptolemy’s model. It worked for prediction. But it was the wrong underlying model. When Newton discovered the inverse-square law, the epicycles became unnecessary.
Krishnan says that AI is brilliant at learning epicycles. It struggles to discover gravity.
The technical version: given any dataset, many different underlying rules could have produced it. AI learns a rule that fits the data. Not necessarily the rule that actually generates reality.
When conditions shift outside the training data, the model doesn’t know it’s in trouble. It just keeps extrapolating. That’s a flash crash waiting to happen. Krishnan says that, just as capital markets are forms of collective intelligence, modern AI systems are as well. Their behavior stems from generating hypotheses that fit observed behavior without understanding cause and effect.
Taleb describes the risks of predicting the future based on the past. Krishnan warns that AI can make accurately-seeming predictions without a model of how the world works.
Both these ideas help us think about where and how to deploy AI capabilities.
A Practical Heuristic
Here’s how I think about evaluating AI investments: Is your project solving a problem that looks like a game of roulette, a hand of poker, or making decisions in the fog of war?
Roulette problems have fixed rules and known odds.
The wheel doesn’t change. The probabilities don’t shift based on what happened on the last spin.
Code generation for standard CRUD applications fits here. Authentication flows, database queries, REST endpoints. These are known domains with established patterns. Same with SQL generation from natural language and document summarization.
The problem space in a roulette-like problem is bounded, the inputs are structured, and the AI has seen thousands of examples that look almost exactly like what you need.
Provided you can create a robust testing strategy, AI is a good fit for this class of problems.
Poker problems have fixed rules but hidden information.
Other players adapt to your moves. The patterns shift because the environment responds to your actions.
Chatbots are the classic example of a poker-like problem. You’re building a conversational interface over a non-structured domain. You cannot predict everything a user might ask. But you can constrain the space through guardrails, deflection, and human-in-the-loop escalation.
Scheduling and pricing optimization fit here, too.
Poker-like problems have genuine uncertainty, but it’s uncertainty within a known structure. You can model the constraints and hedge the risks. The combination of tool use and human oversight keeps the system from wandering too far off course.
Use AI as an enabler or an accelerator to tackle poker-like problems. Keep humans in the loop for consequential decisions.
War problems are uncertain and chaotic
The rules themselves change. The terrain shifts while you’re navigating it. Adversaries rewrite the game while you’re playing. Your Roman legions may end up facing War Elephants crossing the Alps.
Making decisions about resource allocation in the face of competition, automating or using code-generation capabilities in a new domain, or making investment decisions in times of economic or political upheaval are war-like problems.
Here, second-order effects dominate, and relevant patterns don’t exist in the training set.
AI can help with subsets of these problems. It can generate scenarios, synthesize information, and explore options within known constraints.
But the human has to drive. The pattern-breaking is the problem, and no amount of training data prepares a model for out-of-band situations.
The Question
You don’t trust a market to self-regulate during a crisis. You build circuit breakers. The same logic applies to AI deployments.
The 2010 flash crash was a failure of collective intelligence in the face of unexpected inputs. Your AI deployments face the same risk.
Before your next AI investment, ask: Is this roulette, poker, or war?