We Humans Are Very Good at Inventing Games That We Suck At

https://www.youtube.com/watch?v=ykfQD1_WPBQ

The conversation between Yann LeCun and Adam Brown highlights a fascinating paradox: as AI systems continue to outperform humans in chess, law, medicine, coding, trading, drug discovery, protein folding, translation, and scientific analysis, we often interpret this as evidence of approaching artificial general intelligence. But there’s a more nuanced explanation: humans are exceptional architects of complex rule-based systems—and terrible at optimizing them once they scale beyond our computational capacity.

Consider what chess, law, and protein folding have in common. They’re all games in the broadest sense: systems with defined rules, objectives, and constraints. Humans invented chess, legal frameworks, medical diagnostics, and even the scientific method for analyzing proteins. Yet once we formalized these domains into structured problems, we discovered something humbling—computers, with their superior computational throughput and pattern-matching capabilities, could play these games better than their inventors.

This isn’t artificial general intelligence. This is domain-specific optimization at scale. When we created chess, we didn’t anticipate that brute-force evaluation of millions of positions could defeat grandmasters. When we formalized law into precedents and statutes, we didn’t realize that transformer models could identify patterns across thousands of legal documents faster than any human could read them. When we broke down protein folding into a three-dimensional optimization puzzle, we didn’t foresee that AI could model molecular interactions with superhuman accuracy.

The pattern is clear: wherever humans have reduced a complex phenomenon into a formalizable system with measurable inputs and outputs, AI excels. This is powerful—transformative, even—but it’s not general intelligence. It’s specialization at scale.

Examples of “Games” AI Dominates

Coding and Software Development

Writing code is another formalized game with clear rules: syntax, logic, and measurable success (does the program run? Do the tests pass?). GitHub Copilot exemplifies how AI excels here. Research shows that code written with Copilot passes 53% more unit tests and developers using AI assistance code 55% faster. The code is rated as more functional, readable, maintainable, and receives higher approval rates from human reviewers. Like chess, coding follows a set of rules and patterns—and once we formalized those rules into programming languages, machines could optimize them faster than humans.

What’s telling: developers aren’t being replaced, but assisted. The “game” of translating human intent into working code has proven amenable to optimization once it’s been abstracted into a language the machine understands.

Algorithmic Trading and Finance

The financial markets are perhaps the most explicit “game” humans created—a system of rules, incentives, and measurable outcomes (profit/loss). AI excels precisely because finance is already heavily quantified and rule-based. Modern AI trading systems analyze real-time data, detect sentiment across financial news in seconds, and execute trades at microsecond speeds with minimal human emotion interfering. Traditional human traders simply cannot match the computational throughput required to identify and exploit patterns across multiple asset classes simultaneously. AI reduces emotional decision-making and human bias, enabling precision execution at scale—trading across numerous markets and asset classes concurrently without degradation in accuracy.

Here again, AI doesn’t “understand” markets the way a seasoned trader might claim to—it optimizes within the formal constraints of the trading game faster than any human ever could.

Drug Discovery and Pharmaceutical Development

Drug discovery is a massive optimization problem dressed as science: identify disease targets (proteins), design molecules that interact with those targets, predict safety profiles, and run clinical trials. Once formalized into data—molecular structures, protein interactions, historical trial outcomes—it becomes a game AI can play better than humans.

DeepMind’s AlphaFold solved protein structure prediction, a problem that had stumped researchers for 50 years. More broadly, AI-discovered drugs in Phase 1 clinical trials show success rates of 80-90%, compared to 40-65% for traditionally discovered drugs. AI accelerates target identification by analyzing vast biological datasets in hours rather than months, predicts which drug candidates will be safe, and optimizes clinical trial protocols. The entire process—which once took 10-15 years and billions of dollars—is being compressed.

This is perhaps the clearest example: humans invented the rules of molecular biology and pharmacology. We formalized them. And now that they’re formalized, machines can optimize them at a scale we never could, bringing life-saving treatments to market faster.

Image Recognition and Classification

Whether identifying tumors in medical imaging, classifying satellite imagery for agriculture or urban planning, or moderating content at scale, AI has surpassed human performance. These tasks are “games” with clear objectives: is this image containing X or Y? Once the problem is reduced to pixel data and category labels, neural networks optimize the classification task faster and more consistently than human experts.

Protein Folding (Beyond AlphaFold)

Your article already mentions protein folding, but it deserves emphasis as the ur-example. For decades, scientists manually analyzed protein structures through X-ray crystallography—painstaking work. The problem was formalized into: given an amino acid sequence, predict the 3D structure. Within a few years of serious AI attention, the problem was essentially solved. What took human scientists years to determine, AI systems now predict in seconds.

Language Translation and Natural Language Processing

Modern translation systems handle thousands of language pairs—a feat no human polyglot could match. The “game” here is: convert text from Language A to Language B while preserving meaning. Once that problem was formalized with enough training data, transformers solved it at superhuman scale. The conversation you referenced notes that some models now translate “a thousand languages to another thousand languages in any direction.”

Pattern Recognition in Scientific Data

From genome analysis to particle physics, AI excels at finding patterns in massive datasets. These are all “games”—problems humans formalized and then underestimated the complexity of once we had enough data to play them at scale. Whether it’s identifying genetic markers for disease, discovering new materials with desired properties, or sifting through astronomical data, AI’s throughput advantage is decisive.

The Unifying Pattern

What connects all these domains—chess, law, medicine, coding, trading, drug discovery, protein folding, translation, and scientific analysis? They’re all formalized games:

Rule-based: The domain has clear rules and constraints

Optimizable: There’s a measurable objective (winning, accuracy, speed, efficiency)

Data-rich: Enough historical examples exist to train on

Computationally tractable: The problem can be reduced to mathematics

The moment humans formalized a complex domain into such a system, we inadvertently created a game that machines could eventually play better. Not because machines are more intelligent or “understand” the domain, but because they can execute billions of evaluations per second and detect patterns at scales no human brain could match.

What AI still struggles with: Problems that aren’t clearly formalized, domains that require genuine novelty and creativity, situations where the rules keep changing, and contexts demanding abstract reasoning about human values, ethics, or meaning. These remain stubbornly human domains—for now.

The real question isn’t “Will AI become superintelligent?” but rather “What domains have we formalized well enough for AI to exploit, and which remain irreducibly human?”