What does it mean for investors when humans keep outsmarting a powerful AI program?
by Stephen Leeb, PHD and Donna Leeb
Optimists see the glass as half full. Pessimists see it as half empty. That cliche is a wonderful metaphor in many areas of life. But sometimes it fails to capture the full extent of the possible alternating realities.
Today, when it comes to the stock market, the optimists have faith that rising stocks are the gift that will keep on giving. The pessimists worry that the rise is threatened by any number of risks. We fall squarely into the pessimists’ group, as you know if you’ve been reading us. But we think things have gone well beyond the half full/half empty dichotomy.
Rather, we see the glass itself—i.e., the market—as teetering at the table’s edge. It may teeter a while longer, but inevitably it will topple and shatter, with the pieces cascading in all directions and cutting everyone who hasn’t fled.
What do we see that the optimists don’t? One reason for our pessimism is knowing that what largely motivates the market optimists is their faith in the unlimited prospects of AI. They are betting that AI will deliver in a big way, that it will justify in spades the hype and the huge amount of money companies are spending on it. AI, clearly, is the big story in today’s markets. The big tech companies that are making huge investments in AI account for a hefty portion of the market’s gains. They also account for a significant slice of GDP growth. They are basically the pillars on which the market stands. But if those pillars crumble, the market will have nothing holding it up.
We’ve written a lot already about why we think those expectations for AI are wildly misguided. But it’s instructive—if frustrating—to make the effort to understand why many investors continue to hold onto these expectations. We think a lot is explained by the enormous amount of propaganda that is generated daily, propaganda that borders on fantasy but that reassures these investors that they will be on the right side of history, with big gains as their reward. When sources that seem authoritative tell you that all is well, it’s all too easy to believe them. Unfortunately, everywhere you look today, you see examples of misinformation that is bolstering unrealistic expectations.
A misleading claim
A quintessential example was an August article in The Wall Street Journal that was a paean to the exciting and supposedly limitless powers of AI. Anyone who read it, and who took it at face value, might understandably want to dive in and invest hand over fist—without realizing that the article’s reporting concealed a huge gap. In fact, once you look more closely, the article demonstrated the exact opposite of what it supposedly was showing.
The headline was: “Move 37 Is the Moment AI Changes Everything. It’s Suddenly Happening Everywhere.” It was followed by the assertion: “A decade ago, a computer did something no human would have done. It was considered a breakthrough for AI. Now the world is full of them.”
It was a dramatic headline that caught our attention. Move 37 refers to a move that a computer—AlphaGo—made 10 years ago while playing the game of Go against the top human player at the time, Lee Sedol. Go is a complex two-person game that has often been compared to chess, but the number of possible games that can be created in Go far exceeds the number for chess. As described in the WSJ article, Move 37 was a deeply unconventional move that no human player would have made, and at first it was assumed to be a massive blunder. But it turned out to be a masterstroke that left Sedol dumbfounded and shaken
And this, according to the article, is comparable to what AI now is doing everywhere—coming up with original, surprising, brilliant ideas in areas that humans have studied for centuries and making one breakthrough after another.
So, what’s the problem? It’s that the article failed to mention that a few years after AlphaGo’s Move 37 triumph, a later, more advanced Go-playing computer, called KataGo, was utterly stymied by a human amateur Go player who employed an ingenious stratagem to outsmart the computer.
The human player was Kellin Pelrine. In 2016, he was an avid amateur Go player—not as good, then or now, as the very best humans at the game, but still very strong. Pelrine, a student at the time, knew he wanted to pursue academic research but hadn’t yet settled on a particular area. Watching AlphaGo beat Lee Sedol in the 2016 match inspired him to focus on data science and AI. By 2023, a team that he led, working with computers, had developed a stratagem—known as the “cyclic” stratagem—they believed would be able to beat the best AI Go system. It was a stratagem that any relatively strong human player could easily comprehend and apply, and Pelrine used it to win 14 out of 15 games against KataGo.
Blind spots vs. creativity
Since Pelrine’s victory three years ago, researchers have trained and retrained KataGo to focus on the cyclic groups at the heart of the stratagem used by Pelrine. They’ve had no luck, as strong human players armed with knowledge of the stratagem have continued to outplay the machine. We think the only plausible explanation for this continued failure by the machine is that the strategy, though easily grasped by a human, is combinatorially too complex for the machine to process. The ability of any computer—whether it hosts a large language model (LLM) like Claude or Gemini or is fully dedicated to a particular game—is only as good as the number of combinations it can recognize or create. Humans, by contrast, can find creative workarounds that are not bounded by the number of possible combinations in a problem or task.
And in case you’re thinking that as AI continues to develop, it will become able to power past these limitations, that’s extremely unlikely unless humans discover new laws of physics or computation. Keep in mind that the number of potential chess games is, give or take, equal to the number of atoms in the universe. The possible number of Go games is roughly equivalent to the number of atoms in three universes, each the size of ours.
AI itself confirms this fundamental weakness in AI. We posed the following question to the latest version of Gemini, a leading LLM: Why can’t KataGo and other Go AIs be trained to win against a human using a cyclic strategy? In its answer, it referred to “fundamental flaws in how modern AI learns” and pointed to “blind spots”—noting that AI is great at looking at individual local structures but unable to step back and to encompass a massive global structure.
Rather than say that Move 37 was the moment when AI changed everything, we think it’s more accurate to say that Move 37 was a moment in which humans realized they’d have to work hard to beat AI—which they did. And while we’ve focused here on Go, you can extend the same thinking to the LLMs that are at the heart of all the hopes that AI really will to be able to change everything.
It relates to a distinction we’ve made in previous articles between syntax and semantics. Syntax refers to the grammatical side of language—grammar, parts of speech, sequence of words. These for the most part are ruled by basic principles that, while they may be extensive, are fixed and relatively easy to master.
But semantics—the nuances and complexity of meaning—is an entirely different story. That’s not because new words are always being created but because the number of sentences that can be created in English, or in any other language, is infinite. Add to this the fact that the meaning of any given word is always changing and that the same word can mean different things to different people. A statement as simple as “I love music” can have as many meanings as there are people. And for any one person, the meaning of the same words and sentences can change over time.
One implication is that those who see AI as being able to substitute for humans are in effect arguing that all humans are basically the same. They are eliminating individual differences, which is tantamount to crushing individual human creativity, the driving force in human progress.
Language is immeasurably more complex than even enormously complex games like chess and Go. With chess and Go, you still can put a number on how many different combinations of moves are possible. But the consensus among language scholars is that the number of possible sentences can’t be counted. It’s one reason that hallucinations and other flaws in LLMs aren’t one-and-done weaknesses that can be fixed in future versions. They are foundational to AI’s entire structure.
There’s a lot more to say on these topics, but for now we’ll settle for saying that the giddy expectations for AI as seen in the market’s AI fever aren’t based on a realistic understanding of AI’s true capabilities. AI can be an extraordinarily helpful tool in enhancing human creativity, as illustrated by its role in helping Pelrine and his team devise the winning strategy for Go. But it can’t substitute for human creativity, and attempting to use it to sideline humans by putting AI in charge of human activities and endeavors will prove to be a self-defeating pursuit. And the risk is that it could short-circuit future progress if it ends up reducing the opportunities for human creativity—the indispensable force that has driven humanity’s progress for as long as we have walked the earth.
