Rohit Krishnan challenges the prevailing optimism around artificial intelligence by arguing that current models are not discovering truth, but merely stacking increasingly complex approximations to mask a fundamental lack of understanding. He posits that while these systems excel at pattern recognition, their failures are not random glitches but "alien" deviations that reveal they are fitting data rather than learning the underlying generative rules of the world. For busy leaders navigating the AI boom, this distinction is critical: it suggests that scaling up current methods may yield diminishing returns unless we fundamentally change how these systems "think."
The Illusion of Understanding
Krishnan begins by dismantling the idea that Large Language Models (LLMs) possess genuine comprehension. He draws a sharp contrast between human cognition and machine learning, noting that "All LLM successes are as human successes, each LLM failure is alien in its own way." This framing is powerful because it shifts the focus from the model's impressive fluency to its catastrophic, unpredictable breakdowns. The author argues that these systems are essentially "over-fit pattern-fitters" that swim in an ocean of possible solutions, often selecting the shortest path to a plausible answer rather than the true one.
The core of his argument rests on the mathematical reality that there are infinitely more ways to be wrong than there are ways to be right. As Krishnan puts it, "The deeper truth is that success is low-dimensional. There are relatively few ways to correctly solve '2+2=' or properly summarize a news article. But failure is high-dimensional—there are infinitely many ways to be wrong." This observation explains why AI can sound confident while hallucinating facts; the model is navigating a vast space of potential errors that human cognition simply does not explore.
"We see with LLMs that they are remarkably similar to humans in how they think about problems, they don't get led astray all that often. The remarkable success of next token prediction is precisely that it turned out to learn the right generative understanding."
While Krishnan acknowledges the "miracles" produced by these models, he warns that this success is fragile. He compares the current state of AI development to Ptolemaic astronomy, where scholars added "epicycles upon epicycles" to explain planetary motion without ever discovering the true laws of gravity. This analogy is particularly striking for policymakers and investors who might mistake incremental improvements in accuracy for a breakthrough in reasoning.
The Trap of Pattern Matching
The author extends this critique to real-world applications, highlighting the dangers of relying on models that cannot distinguish between correlation and causation. He points to the Federal Reserve's caution regarding "not-easily-interpretable ML" in loan approvals and the immense data requirements for autonomous driving as evidence that current models struggle with the "radical uncertainty" of the real world. Krishnan writes, "The rules that were learnt were not the rules that should have been learnt. This is a classic ML problem, that still exists in deep learning."
Critics might argue that Krishnan is being overly pessimistic about the utility of probabilistic models, noting that in many business contexts, high accuracy is sufficient even without true understanding. However, his point about "unknown unknowns" remains a vital warning. He suggests that unlike a casino where the odds are known, the real world operates in "Extremistan," where the rules can change unexpectedly. If an AI system has only learned patterns from historical data, it is ill-equipped to handle novel crises or black swan events.
Can Reasoning Escape the Pattern Trap?
In the latter half of the piece, Krishnan explores whether "reasoning" capabilities—such as chain-of-thought prompting or tool use—can help models break free from their pattern-matching constraints. He suggests that forcing models to "think step by step" allows them to search for the right generative rules rather than just the most likely next token. "If you are able to reason through something then surely you will be able to get to the right answer one way or the other," he notes, though he immediately tempers this optimism by pointing out the limitations of current search capabilities.
The author argues that we are currently "pushing the induction problem up one level," trying to teach models how to learn how to learn. While this approach shows promise, he warns that it still falls prey to the same lack of insight that plagues simpler models. "It does add a significant lag to their training, but essential nonetheless," he writes, implying that the computational cost of true reasoning is high and the path forward is not guaranteed.
"The aim is not to predict the next data point, but to infer the rule that generates all of them."
This section is the most forward-looking, suggesting that the future of AI depends on moving beyond mere prediction to genuine rule inference. However, Krishnan admits that we may not yet have the theoretical framework to solve this, stating, "Maybe if we got them to think through why they were predicting the things they were predicting as they were getting trained, they could get better at figuring out the underlying rules."
The Bottom Line
Krishnan's most compelling contribution is his reframing of AI failures not as bugs to be patched, but as symptoms of a fundamental mismatch between pattern matching and causal understanding. His argument is strongest when he exposes the fragility of current models in the face of "unknown unknowns," a reality that institutional leaders cannot afford to ignore. The biggest vulnerability in his analysis is the lack of a concrete roadmap for how to achieve true rule inference, leaving the reader with a clear diagnosis but an uncertain cure.