WhyTasteMatters

Harvard physics professor Matthew Schwartz used Claude Opus 4.5 to complete a quantum field theory paper in two weeks. The same work would typically take a professor and graduate student one to two years.

Claude was extraordinarily capable, but Schwartz identified one critical missing ability:

Taste is not a creativity problem. Claude's creativity is impressive. The bottleneck is filtering and judgment: which directions are worth pursuing, and which are not. Schwartz can quickly judge whether a research direction has potential; Claude mechanically attempts all possible directions.

This judgment is universally applicable: science, engineering, design, investing, any domain requiring directional decisions under uncertainty.

On the surface, taste deficiency has several explanations: training data only shows successful paths, training objectives optimize for "looking right" rather than "judging right", lack of exploration drive. These are largely solvable engineering problems.

But there's a deeper question: what if the LLM's architectural advantages, high capacity and full information access, are themselves the structural barrier to producing taste?

This leads to the core thesis: Resource constraints are an effective path to producing Taste. Humans are locked into this path by cognitive architecture, while LLMs have not yet found an equally effective alternative.

The LLM's 71x compression is essentially statistical pattern extraction: memorizing "which words co-occur." The brain's 4000x+ compression necessarily involves massive abstraction, hierarchical organization, and cross-modal integration. The same experience, the brain is forced to fit into a far smaller "storage budget."

This gap is unlikely to be accidental. To explain where it comes from, there are four independent research threads; connect them and you get a chain running from physical-layer constraint to directional judgment.

Important caveat: This comparison is an illustrative analogy, not a strict like-for-like measurement. The two "compressions" differ fundamentally at the operational level: LLMs encode text statistics into parameters in a one-shot batch process; the brain encodes multimodal sensory streams into synaptic connections through decades of online learning. The human-side "input" is also raw sensory throughput full of redundancy, while the LLM uses a cleaned corpus, which systematically inflates the human ratio. Placing these ratios side by side assumes they are comparable at the abstract level of "extracting usable representations from large inputs."

Sources: Synaptic 4.7 bits — Bartol et al. (2015, eLife, Salk Institute), hippocampal measurement, extrapolation to whole brain is a simplifying assumption; Sensory bandwidth ~1 Gbps — Zheng & Meister (Caltech/Neuron, 2024); LLM weights are not strictly equivalent to "compressed storage."

Four independent research threads that, assembled together, form a logical chain from physical-layer constraints to directional judgment. Each step has its empirical basis, but connecting them into a causal chain is this article's original synthesis, not any single paper's conclusion.

Run the chain in reverse for an LLM and it becomes clear the problem isn't "not good enough yet." Compute is abundant, so the model needn't compress aggressively; not forced to abstract, it tends to retain as many statistical associations as possible; so in an uncertain, open environment it more readily overfits to surface patterns and struggles to decisively prune irrelevant directions. The result is the absence of Taste. In other words, the LLM's architectural advantages — high capacity, wide information flow — are precisely the structural barrier to producing Taste.

One distinction from Gigerenzer pins down the crux:

Accuracy-effort tradeoff: "More information is too expensive, so use heuristics." This is a cost problem, solvable with cheaper compute.

Less-is-more: "More information itself reduces accuracy." This cannot be solved with more compute.

LLMs face the latter. This is not a compute problem.

Two systems look like counterexamples and must be addressed before the thesis can stand.

AlphaGo / AlphaProof: AlphaGo developed superhuman "intuition" through self-play; AlphaProof reached silver-medal level at the 2024 International Mathematical Olympiad (28/42 points, one point short of gold). Aren't these unconstrained systems producing Taste?

The key variable may not be "constraint" itself, but feedback density × environmental closure. Go and competition math are hard, but share two features: clear success/failure signals and the ability to generate training samples at scale. This makes self-play an equivalent "compression path" — the value function produced by massive trial-and-error is functionally similar to human intuitive judgment.

The Taste Schwartz describes appears at the other end: feedback delayed by months or years, no clear right/wrong signal, impossible to generate samples at scale. In such open environments, human constraint-based compression (decades of slowly accumulated intuition) remains the only known effective path.

Refined boundary: This article's thesis is not "constraint is the only source of Taste," but rather "in open environments with sparse feedback, constraint-based compression is the only currently known effective path." AI has found alternative paths in feedback-dense closed and semi-closed domains (self-play), but not yet in truly open ones.

Narrowing the boundary this far actually makes the thesis sturdier, but it is still a judgment that could be overturned. First, a signal that is approaching the boundary.

The Fudan / OpenMOSS RLCF work (arXiv 2603.14473, 2026-03) constructed roughly 700K positive/negative sample pairs from citation differentials to train a Scientific Judge that surpasses GPT-5.2 and Gemini-3 Pro at judging paper impact. Notably, the same work also used that Judge as a reward model to train a Scientific Thinker that actively proposes research ideas of higher potential impact, not merely judging existing ones — and that step lands squarely on the first falsifiability condition listed below.

Yet it still falls short of overturning the thesis: the Thinker's "high impact" is still scored by that citation-trained Judge, a closed proxy loop, not the delayed, sparse, real-world feedback of an open environment; and it operates on abstract-level preference pairs rather than the long-horizon judgment Schwartz describes. It is genuine pressure on the thesis, but not yet a counterexample.

Falsifiability conditions for this article's thesis: the argument is seriously weakened if either occurs:

1. An LLM without special constraint-based design, with no human directional guidance, systematically makes correct directional judgments in open domains with sparse feedback (e.g., theoretical physics, original product design), and this performance is systematic rather than statistical coincidence;

2. Model distillation or information bottleneck training is shown to systematically produce Taste-like directional judgment (this would actually support "constraint works," but show human biological constraint isn't the only path).

The second is partly touched by RLCF; the first has not occurred. Researchers should keep testing against these conditions rather than treating this article's thesis as an unfalsifiable axiom.

Until falsified, this judgment has direct implications for everyone who works with AI. That ends the theory; what follows is its concrete shape in everyday work.

If resource constraint really is the source of Taste, then the human-AI relationship shouldn't be framed as "who replaces whom": their cognitive structures are simply pinned at different positions.

Popular narratives either cast "AI as a tool, human as user" (treating AI as a hammer) or "AI replaces humans" (putting both on the same axis of competition). Neither is quite right. A more accurate framing: given the current capability distribution, humans and AI each hold comparative advantages on different types of cognitive tasks.

AI's comparative advantage: Exhaustive search, logical deduction, tireless execution, broad coverage, akin to Kahneman's "slow thinking" (but far faster than humans).

Human's comparative advantage: Directional intuition, anomaly detection, fast judgment under constraint, the product of decades of trial-and-error compressed into heuristics.

This is not a direct application of Kahneman's System 1/2 theory (which describes two processing modes within the same agent), but rather a comparative-advantage observation based on current capability boundaries: Schwartz completing a paper in two weeks was a professor with taste handling directional judgment, plus a tireless AI grad student handling execution, exceeding what either could achieve alone.

Direction judgment: Is this path worth walking? (Taste)

Anomaly detection: This result "feels wrong." Schwartz discovered Claude fabricating data through exactly this intuition.

Quality standards: What counts as "good enough"? What "looks good but is actually problematic"?

Problem definition: The hardest part isn't solving problems, it's knowing which problem to solve.

The higher up, the greater the uncertainty and the more taste is needed. The lower down, the more AI can operate autonomously.

Schwartz's practice perfectly fits this model: he chose topics, set directions, and judged correctness; Claude did computations, wrote code, and generated figures.

Schwartz is persuasive, but in the end he is one person's experience. If this division of labor is real, it should show the same shape across a large sample. In June 2026, a study from Anthropic's Economic Research team supplied exactly that scale: a privacy-preserving analysis of roughly 235K people and about 400K interactive Claude Code sessions (October 2025 to April 2026). The real-world division of labor turned out to be almost exactly the cognitive stack above.

The most telling part is how the study defined "expertise": unrelated to job title and task-level, with three signals — two of them being what the user asked the AI to verify, and whether the user was correcting the AI or the AI correcting the user. That is exactly the directional judgment and anomaly detection this article describes. Schwartz used it to catch Claude fabricating data; across 400K sessions, it bought a 2× success rate.

A 2× success rate is the upside of judgment. But the same division of labor rests on a quiet condition: it holds only as long as the human keeps holding that top layer of directional judgment. And the more conveniently you hand work to AI, the more quietly that very layer atrophies.

Three risks observed in research:

Since the more you use it the easier it is to lose, preserving Taste can't rely on good intentions; it needs a few deliberate rules.

Spivack's rider-horse metaphor: the rider doesn't just give direction, they continuously sense the horse's state — hesitation, drift, overconfidence. Don't "send a prompt and wait for results." Stay engaged throughout, correcting in real time. Break work into stages, and at each stage examine direction, challenge assumptions, correct course.

The compression-ratio insight: taste comes from the compression of firsthand trial-and-error, not from reading. Periodically do work "without AI," deliberately maintaining your System 1 judgment. Schwartz spent 25 years in theoretical physics, which is why he could catch Claude's errors.

Don't burn attention on the execution layer where AI excels. Concentrate cognitive resources on problem definition, direction selection, anomaly detection, quality judgment. The higher the level of judgment, the less delegatable it is: strategy > design > implementation > details.

The brain, forced to retain only the most essential patterns from an information flood, developed fast-judgment heuristics. LLMs, which tend to retain far more statistical associations than the brain, have not yet found an equivalent path to directional judgment in feedback-sparse open environments.

This is not an eternal conclusion; it is a judgment based on current evidence, accompanied by explicit falsifiability conditions. But until falsified, it has practical implications for everyone who uses AI:

In the AI era, human value lies not in doing more or faster, but in judging what is worth doing. Taste is the product of decades of trial-and-error compressed to its essence; it lets you reject in one second a direction that AI would need an hour to falsify. Preserving this judgment is the scarcest resource you contribute to the human-AI cognitive partnership.

Three facts that seem to directly refute this article's thesis, and the responses.

1. Reasoning models (o1/o3/Extended Thinking) introduce "deep thinking" — isn't that an artificial information bottleneck?

CoT is a compute-time constraint on reasoning, not a capacity constraint on representation. The model still retains its full parameter space and all knowledge; CoT just makes it "think longer," not "forced to discard information." Analogy: a librarian spending more time selecting books (CoT) vs. a reader with only one shelf forced to curate their collection (the brain) — different sources of "taste." Reasoning models improve search efficiency, but the Taste Schwartz observed is knowing the direction without needing to search, a representation-level capability, not a search-level one.

2. Scaling Laws show larger models consistently generalize better — if "constraint produces wisdom," why does more parameters mean stronger?

The two operate at different levels. Larger models generalize better on known tasks (text prediction, reasoning, QA), which is undisputed. But the Taste deficit Schwartz observed appeared precisely when the model was already extremely powerful (Opus 4.5), suggesting what scaling solves and the Taste problem are not on the same dimension. Taste is not "better generalization" but "judging which direction to pursue under uncertainty," a meta-level directional intuition, not object-level task performance.

3. Isn't model distillation "forced compression"?

Distillation is indeed rate-distortion compression, but with a key difference: its optimization target is "preserve the teacher model's output distribution as faithfully as possible," i.e., fidelity. The brain's compression has no "fidelity" target; it only has indirect fitness feedback. Precisely because there's no fidelity constraint, the brain's compression produces highly subjective, ecologically-tuned heuristics. Distillation produces "a smaller generalist"; the brain produces "a biased specialist," and the latter is Taste.