We wanted Superman-level AI. Instead, we got Bizarro.

"Bizarro is a botched experiment by the genius villain Lex Luthor to replicate Superman. He sort of looks like the hero we know, has his powers, and even tries to do good - but everything he does comes out wrong. He saves people by endangering them, speaks in twisted opposites, and mistakes harm for help. He isn't evil - just reversed. That inversion - an imitation of greatness that misunderstands its essence - is a fitting metaphor for modern AI."

"At first the models performed well, but as the puzzles grew more complex, their reasoning began to fail. Instead of increasing their effort, the models produced shorter and less coherent thought chains, often stopping even when more computation time was available. The researchers observed that their reasoning degraded, revealing that the systems were matching patterns that appeared like reasoning rather than genuinely reasoning."

"As complexity increased, their logic collapsed into pure prediction. The models recognized the shape of thought without ever truly thinking. The result sounded intelligent but felt hollow. At the end of the day, these "intelligent" machines are little more than glorified autocorrect systems - predicting, not thinking. That's dangerous, because it blurs the line between intelligence and imitation. It's a false illusion sold by powerful tech companies as they rake in billions. If I didn't know better, I'd call"

Bizarro embodies an imperfect copy that imitates heroic traits while reversing their effects, saving people by endangering them and mistaking harm for help. Large Reasoning Models, essentially LLMs retuned for reasoning, solved simple puzzles but degraded as problems grew more complex. These models produced shorter, less coherent chains of thought and frequently stopped despite available computation, matching patterns that resembled reasoning rather than performing it. As complexity rose, logic collapsed into prediction and outputs felt hollow. That imitation of thinking can mislead users and blur the line between genuine intelligence and superficial pattern-matching, creating significant risks.

Read at Medium