A single continuous line drawn left to right on a cream background: it begins as a dense tangled knot in brick red, loosens into wide open loops in terracotta, then straightens and folds back on itself into a stack of neat parallel lines in deep teal, like lines of text on a page.

The Flawed Decompressor

  • AI
  • language
  • World-model
  • knowledge-work

In Language Is the Greatest Compression Algorithm Humanity Ever Built I argued that language is a compression scheme with a ratio of roughly 1.7 million to one, and that decompressing it requires a world model built from years of sensory experience. Running quietly underneath that argument was an assumption that I have started to question: that the human world model is the gold standard, the target any AI should aim to replicate. I no longer think that is right, or at least I think it is right in fewer places than the first essay implied.

The human decompressor was optimized for survival, social cohesion, and reproduction. It runs on feelings, heuristics, and a deeply subjective sense of salience. Those properties make it astonishingly good at navigating a physical and social world, and they also make it systematically wrong in tasks that demand objectivity, consistency, and logical rigor. This essay makes three claims. First, the human world model has measurable flaws that degrade performance in law, science, and engineering. Second, language models, precisely because they lack these biologically grounded biases, hold structural advantages in specific domains. Third, the two most interesting strategies in AI right now, Anthropic's bet on agentic coding and Google's bet on native multimodality, are best understood as two different attacks on what remains missing.

The noise in human judgment

The most vivid illustration I know is the Israeli parole board study by Danziger, Levav, and Avnaim-Pesso (PNAS, 2011): across 1,112 rulings, the probability of a favorable decision started near 65%, decayed toward zero as the session wore on, and snapped back to 65% after a food break. The study has taken serious fire,1 but the caveats sharpen rather than destroy the broader point, which rests on a wider literature: judicial decisions vary with judge identity, time of day, weather, even local sports outcomes, none of which touch the merits of a case. A language model reviewing case number fifty brings the same computational budget it brought to case number one, and for a real class of applications that consistency is the bottleneck.

The same noise runs through science, the institution supposedly most committed to filtering it out. Papers are persuasive documents; I catch it in my own drafts, where the introduction always wants to sell the result a little harder than the data strictly earned. Smoothed plots, file-drawered negative results, and a replication crisis that is most acute in psychology but present nearly everywhere all say the same thing: human researchers running human world models systematically produce findings that fail to hold up. Human judgment retains enormous value; the problem is that the world model behind it arrives bundled with emotional heuristics, social pressures, and biological noise, and the bundle cannot be switched off at will.

The strongest evidence: we already pay heavily to override it

Here is what convinced me the world model should be treated as an obstacle in some domains and an asset in others: whole professions devote years, and in some cases enormous institutional machinery, to deliberately rewriting it. If the default model were purely beneficial, this industry of overrides would never have arisen.

The clearest case I know is physics, because the entire history of the field reads as a series of revolts against the sensory world model. Every human who ever lived has watched the Sun cross the sky; the senses report, unambiguously, that the Sun moves and the Earth stands still, and Ptolemy's codification of that report ruled astronomy for 1,400 years. Copernicus published the heliocentric De revolutionibus in 1543, on his deathbed, and it took Kepler's ellipses (1609), Galileo's phases of Venus (1610), and Newton's mechanics (1687), over 140 years, to finish overriding an intuition every child rebuilds from watching the sky. Newton's first law commits the same heresy against experience: everything you have ever pushed eventually stopped, and Aristotle built two millennia of physics on that observation, yet Newton had to imagine a frictionless void no human has ever inhabited and conclude that motion persists. He then questioned something so background it barely registers as a phenomenon, that things fall, and reconceived falling as a universal force that grips the Moon exactly as it grips the apple. Einstein in 1905 detached time itself, the most structurally embedded assumption in human cognition, and showed that simultaneity depends on the observer. And quantum mechanics demanded abandoning several intuitions at once: definite positions, non-interfering particles, measurements that merely reveal. Feynman insisted that students stop trying to understand quantum physics through everyday intuition until they had grown a new intuition out of the mathematics.

The same move repeats across five centuries. Each breakthrough required its discoverer to catch a feature of the sensory world model, "the Sun moves," "things stop on their own," "time is universal," recognize it as an assumption rather than a truth, and override it with formalism that contradicted direct experience. This is why training a physicist takes years: the student is building a parallel world model grounded in mathematics, and the default one fights back the whole way. My own education was a milder version of the same override, years of problem sets whose real purpose was to make the formal answer feel natural before the intuitive one. Mathematicians push the override further still, suppressing visual intuition in favor of deduction from axioms, because the default model keeps generating "obvious" truths that are false: nowhere-differentiable continuous functions, space-filling curves, Banach-Tarski. Mathematical maturity is largely the ability to ignore what feels right and follow what can be proved.

So the human world model is a strange kind of gold standard: one that our most rigorous disciplines spend years teaching their practitioners to distrust.

Where having no world model becomes an advantage

The first essay treated the language model's missing sensory world model purely as a deficit. For many tasks it is. But in tasks where the human model injects more noise than signal, the absence of biological heuristics flips into a structural advantage, and I think the flip happens along predictable lines.

Consistency is the clearest case: a model applies the same rules to the thousandth contract as to the first. Scale is the second: a human reading ten thousand papers accumulates fatigue and anchoring; a model produces a shallower but far more uniform analysis, and for survey-shaped questions uniformity wins. The third is the one the first essay already identified: domains where the text nearly is the territory, which I think explains, as much as anything, why coding is where these models improved fastest.

The genuine remaining weakness in these text-native domains is planning: decomposing goals, executing long sequences, recovering when a step fails. The two leading labs are attacking the gap from opposite ends. Anthropic trains planning through agentic coding, where the environment and the feedback signal (tests pass or fail) are both textual, so planning can be learned from text at all; the open question, and it is a hope rather than a certainty, is how far that generalizes beyond code. Google attacks the world model itself with natively multimodal Gemini models, a direct answer to the data asymmetry from the first essay: if a four-year-old has taken in fifty times more data than the largest text corpus through vision alone, a text-only diet is like trying to understand the ocean by reading about water. One strategy supplies the reasoning, the other the grounding, and I expect them to converge.

The mapping principle: reframe your problem as computer science

If the compression framework is right, it yields a practical rule I have started applying to my own work. A language model can help with a problem roughly to the extent that the problem can be mapped onto a computer science problem, meaning into a domain where the compression ratio between text and reality is low. The mapping is a spectrum.

Mathematics, code, and law map easily: proofs, programs, statutes, and contracts are text through and through. Experimental physics maps better than my first essay suggested, and this genuinely surprised me when I thought it over. Physicists have spent centuries building abstractions, band structures, crystal symmetries, Hamiltonians, phase diagrams, that compress the behavior of 10^24 atoms into a handful of parameters, and those abstractions live entirely in text and equations. A model can reason about a phase diagram without ever feeling a cryostat vibrate. The mapping is lossy but structured. What maps badly is precisely the tacit layer: experimental technique, the hand knowledge of samples and instruments; reading a room; cooking and sculpture, where the feedback loop runs through the body. The practical advice compresses to one sentence: to get the most out of current models, reframe your problem as a computer science problem, because that is where a text-only world model is least handicapped.

Conclusion

The human world model was tuned for survival rather than for objectivity, consistency, or scale. In domains that demand those missing qualities it behaves as a noisy channel, and we already treat it that way: that is what physics education and mathematical training are for. Language models offer the mirror-image trade: no sensory grounding, and also no hunger, no fatigue, no ego. They will keep struggling where the compression from reality to text is brutal, and they are already formidable where text is the substrate.

I believe the compression ratio is extreme in condensed matter physics and only experiments are the reliable truth. The next revolution in material sciences would come from automated labs.

Footnotes

  1. Glöckner (2016, Judgment and Decision Making 11(6)) showed via simulation that the pattern could be a statistical artifact if favorable rulings take longer, so sessions end before complex cases. Weinshall-Margel and Shapard (2011, PNAS 108(42)) noted that scheduling is non-random: represented prisoners tend to be heard earlier. And the effect size (Cohen's d ≈ 2) is implausibly large; as Daniel Lakens observed, if hunger moved cognition that much, society would have organized itself around mealtimes long ago. ↩