Blog post: the oldest pattern matcher

feynman.md

@@ -0,0 +1,54 @@
+
+# Lede
+
+Pocono Conference, spring 1948. Feynman at the blackboard, twirling chalk, drawing straight lines and wavy lines and vertices. Schwinger had lectured for hours the day before: polished, formal, complete. Now Feynman is sketching what look like cartoons. Teller interrupts: this violates the exclusion principle. Dirac keeps asking "Is it unitary?" Bohr strides to the stage and lectures Feynman on the uncertainty principle, mistaking the diagrams for literal particle trajectories. The presentation fails.
+
+Eighteen months later, those cartoons are doing in hours what formal methods had taken months to achieve. Schwinger later compared the diagrams to the silicon chip: "bringing computation to the masses." Within a decade they have reshaped theoretical physics.
+
+Feynman's Nobel lecture, 1965: "We have a habit in writing articles published in scientific journals to make the work as finished as possible, to cover all the tracks." The formal reconstruction is the residue of the work, presented as though it were the work itself.
+
+{Seed for the closing: something about what distinguishes a frozen pattern from a living one — planted lightly, not labelled.}
+
+# The elevation of reason
+
+Institutions privilege formal reasoning because it scales: it can be taught in classrooms, examined, audited, transferred between people. Tacit expertise dies when the expert leaves. So institutions create incentives to formalise, even when formalisation degrades the underlying capability (Scott's metis). Professions claim status by claiming formal bodies of knowledge (Abbott). The hierarchy runs through how status is allocated: formal reasoning is associated with intellectual seriousness, intuition with mere craft.
+
+Pearl's ladder of causation — association, intervention, counterfactual — is a useful taxonomy. It is also an exhibit of this hierarchy, ranking formal causal models above pattern matching. The ranking may have the directionality wrong.
+
+# Reasoning backwards
+
+In the 1940s, the psychologist Adriaan de Groot showed chess masters a board position for five seconds, then asked them to reconstruct it. Masters reproduced nearly everything. Amateurs got fragments. But when the pieces were placed randomly, the advantage vanished. The skill was perception of meaningful patterns, not calculation or memory. Masters don't think further ahead than amateurs. They see more.
+
+The same pattern appears across formally respected fields. Einstein described his thinking as visual and muscular, translating to mathematics only for communication. Working mathematicians surveyed in the 1940s nearly universally reported that results arrived through sudden recognition after periods of unconscious incubation, with formal proofs following after. Feynman, in the same Nobel lecture, described publishing results he hadn't yet proven because "a very great deal more truth can become known than can be proven."
+
+Gary Klein studied firefighters, military commanders and ICU nurses making high-stakes decisions under pressure. They almost never compared options. They recognised the situation and acted. When forced to enumerate alternatives — the prescribed rational method — they performed worse.
+
+Formal reasoning is what novices do while the pattern library is being built. Once built, the pattern library is the superior tool. Asking an expert to show their working is asking them to operate at a lower level of skill.
+
+# Pattern matching forwards
+
+Hofstadter spent four decades arguing that analogy is the core of all cognition. "Every concept we have is essentially nothing but a tightly packaged bundle of analogies," he writes in Surfaces and Essences. His chapter on Einstein walks through the analogies one by one: the equivalence principle began as an analogical identification between freefall and the absence of gravity. The light quantum hypothesis mapped the statistical behaviour of a gas onto radiation in a cavity. In each case the analogy preceded the formalisation. It was the generative spark that told Einstein where to point the mathematics.
+
+Hofstadter and Sander present pairs of proverbs that assert contradictory things: "where there's smoke, there's fire" alongside "don't judge a book by its cover." We inherit a whole library of patterns, but they point in every direction. Intelligence is selecting which pattern fits the situation.
+
+What we call a causal mechanism is itself a stabilised analogy: "force" borrowed from muscular pushing, "current" from rivers, "selection pressure" from engineering. Mechanism is analogy that has ratcheted into robustness through repeated testing.
+
+# Metapatterns and reinforcement
+
+The contradictory proverbs show that a pattern library alone can't tell you which pattern to apply. What separates useful pattern matching from noise is continuous refinement against reality.
+
+Expert intuition is built through sustained engagement with hard problems at the edge of current ability — deliberate discomfort, not accumulated hours. Through active reconstruction rather than passive reception: Feynman refused to accept results he hadn't rederived, treating each rederivation as a way of building perceptual capacity. Through breadth: his lockpicking, his Mayan codices, his drawing were fuel for the analogical engine. And through repeated cycles of being wrong, integrating the consequences, and refining the library.
+
+A pattern library frozen in a textbook is different from one being continuously tested against reality. The expert's advantage is having patterns that are still being refined.
+
+LLMs inherit enormous pattern libraries from the serialised outputs of minds that already did the causal work. Text encodes causal structure because it was written by causal reasoners. LLMs succeed where naive "pattern matcher" readings can't explain because they're pattern matching over this substrate. They struggle where continuously-refined, intervention-based expertise matters — the parts of cognition that require being wrong about the world, repeatedly, and integrating the consequences.
+
+The conventional framing says LLMs lack world models, so the fix is adding explicit causal structure to the architecture. But what matters about world models may be that they update, not that they represent cause and effect. A frozen causal model is just another static pattern. Reasoning models jump capability without adding anything explicitly causal. Chain-of-thought and search over hypotheses add dynamism to what was static, something like the expert's intervention-and-feedback loop compressed into a single inference. The research frontier is about making pattern libraries adaptive: reasoning, tool use, agentic loops, memory.
+
+For practitioners: the LLM's pattern library is vast but frozen, the practitioner's is smaller but alive. The productive collaboration is the system's breadth meeting the human's adaptive refinement — a partnership with a pattern matcher whose library complements your own.
+
+# Conclusion
+
+The audience at Pocono saw cartoons where there was physics. They mistook the representation for a lack of rigour because they'd internalised a hierarchy that equates formalism with seriousness. Feynman's diagrams were the physics. The formalism came after.
+
+We're still internalising that hierarchy, and it shapes how we evaluate AI. The formal capability these systems have commoditised was supposed to be the valuable thing. The valuable thing was always the patient, effortful construction of a pattern library refined by sustained contact with reality. The question worth asking about these systems is how fast their pattern matching is becoming adaptive — and what that means for the people whose expertise was always built the same way.