The Expensive Parrot Parade: Why Scaling LLMs Is Just Making Bigger Birds

Bless the brave souls still throwing trillions at bigger models, convinced that one more order of magnitude will finally unlock the secrets of the universe. Enter Professor Judea Pearl, the Turing Award-winning grandfather of causal reasoning, who has politely (but firmly) informed the AI hype machine that no, Virginia, scaling won’t save us. There are “mathematical limitations that are not crossable by scaling up.” Translation for the optimists in the back: you can’t brute-force your way to actual intelligence by feeding a parrot more crackers.

Large Language Models, our glittering digital oracles, aren’t learning how the world works. They’re mastering how we describe the world. They gorge on mountains of text—papers, tweets, Reddit rants—and regurgitate the patterns with uncanny fluency. Ask one why aspirin relieves headaches, and it’ll spit out a beautifully written explanation cribbed from every pharmacology textbook ever digitized. But does it understand the prostaglandin pathway? Of course not. It’s just an extremely well-read mimic, acing the imitation game while having zero clue about reality’s causal machinery.

This brutal truth hits hardest in biology and drug discovery, where the stakes aren’t memes or chatbots but actual human lives. We’ve sequenced genomes until our hard drives beg for mercy, yet AI models mostly excel at pattern-matching in published literature. They learn what scientists think causes disease—correlations galore, sprinkled with hopeful p-values—rather than what actually does. “This gene correlates with cancer” becomes the sacred mantra, repeated across thousands of papers until the model parrots it confidently. Meanwhile, the real causal chain—gene actually messes with pathway X, which flips switch Y, leading to tumor Z—remains elusive. We’re essentially training billion-dollar systems to summarize the scientific gossip mill, not to dissect raw biology.

Pearl’s causal revolution offers the escape hatch. His Ladder of Causation climbs from mere association (”seeing”) to intervention (”doing”) to counterfactuals (”imagining what if”). Only at those higher rungs can we move from “this gene correlates with cancer” to “this gene causes cancer”—and, crucially, design drugs that actually work instead of failing spectacularly in Phase III trials. Until we build systems that can reason “what happens if I knock out this gene?” or “what would have happened without this mutation?”, we’re stuck in correlation purgatory, chasing ever-larger parrots that squawk increasingly convincing nonsense.

The irony is delicious. The same industry that mocks old-school symbolic AI as brittle now worships at the altar of scale, pretending endless compute will magically conjure causation from correlation. Spoiler: it won’t. We’ll get shinier hallucinations, more eloquent BS, and probably a few more viral “AI discovered a new protein!” headlines before the next clinical flop reminds us that expensive parrots don’t cure diseases—they just describe the ones we already know about, very prettily.

So here’s to Professor Pearl, the party pooper who reminds us that intelligence isn’t measured in parameters, but in the ability to ask “why” and mean it. Until then, enjoy the show: the most sophisticated echo chamber humanity has ever built, getting louder, dumber, and more expensively wrong with every release.