13 · Learning Machines

This is Part II's fifth hearing: computation has spent a decade teaching itself to speak, translate, and see without being told a single rule, and now philosophy is asked to explain how that could count as knowing anything.

What does machine learning prove about how knowledge enters a mind?

On 24 May 2024, Anthropic let the public talk to a model that could not stop mentioning the Golden Gate Bridge. Ask it for a love letter and it worked the bridge's orange paint into a metaphor; ask it to file your taxes and it found a way back to the fog over the strait. This was "Golden Gate Claude," a demonstration built from research the company had just published: "Scaling Monosemanticity," which used a sparse-dictionary technique to pull tens of millions of human-interpretable "features" out of the middle layers of Claude 3 Sonnet — internal directions that reliably fire for a specific bridge, for sycophancy, for a known kind of security flaw (Anthropic, "Scaling Monosemanticity," May 2024). For the demonstration, one feature was simply turned up until it dominated. The model was taken offline about a day later. What it left behind was stranger than the joke: a single concept, inside a system nobody had hand-coded concept by concept, had an address you could dial like a volume knob. A structure philosophers had called irretrievably opaque turned out, at least at this one coordinate, to have a location.

That location is where this chapter starts, because finding it reopens a dispute older than any computer: what, if anything, must a mind already have before experience can teach it something.

Empiricism and rationalism have been run, this past decade, as an experiment at scale. Noam Chomsky coined "poverty of the stimulus" in Rules and Representations (1980): children acquire grammatical structure that the input they actually hear radically underdetermines, so some of that structure must be innate — Universal Grammar. Large language models seem to answer back: trained on nothing but text, with no grammar built in, they become fluent. But the reply does not settle the argument, and philosophy objects on the argument's own terms. Chomsky, with Berwick, Pietroski, and Yankama, restated the case in "Poverty of the Stimulus Revisited" (2011): the claim was always about what a child's actual, limited exposure can support, not about what any learner given enough data can do. A model trained on a written record no child ever sees is not a counterexample to a claim about children; it is a different experiment with a different subject.

The empiricist side of that experiment has a doctrine, and the doctrine is a wager about how knowledge arrives. Richard Sutton put it into engineering form in March 2019: "The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin" (Sutton, "The Bitter Lesson," 13 March 2019). Chess, Go, speech, and vision each fell to search and learning after decades spent hand-encoding what researchers thought minds needed. But a lesson drawn from which strategies won in a handful of domains is a historical generalization, not a proof about how any mind must gain knowledge — and it is silent on domains, like formal proof (ch.12), where built-in structure has not obviously become a liability.

Chapter 05 puts Stevan Harnad's 1990 question on the record: can a system that manipulates symbols purely by their shape ever mean anything by them, or does it only borrow meaning from whoever reads its output? A model trained only on text is, on that question, purer than anything Harnad had in mind: its entire "experience" is other symbols. Is a feature that fires reliably on "Golden Gate Bridge" across languages and phrasings a grounded reference to the object, or an extremely well-organized cluster of shapes that still refers to nothing? The finding is evidence; it is not a verdict.

The research line runs from "A Mathematical Framework for Transformer Circuits" (Anthropic, December 2021) through "Towards Monosemanticity" (2023) to the 2024 scaling result, and it gives philosophy of mind something it rarely has: a system whose internal representations were discovered, not designed, open to direct inspection rather than armchair argument. The caution belongs beside the excitement — a "feature" is what a sparse autoencoder, itself a trained model, finds; whether that unit deserves the name "concept" is exactly the question interpretability opens rather than closes.

Hume's problem is replayed, batch by batch, inside the training loop. In 1748, Hume showed that no argument — demonstrative or probable — can license projecting a past regularity forward without already assuming nature is uniform, which is the very thing in question (Hume, An Enquiry Concerning Human Understanding, Section IV, 1748). Every gradient update is exactly that projection: a rule fit to past batches, licensed to govern data it has never seen. Statistical learning theory supplies generalization bounds, but only by assuming train and test are drawn from the same distribution — Hume's uniformity principle, formalized and assumed, not derived. Computing's retort is that probably-approximately-correct is enough to ship; philosophy's is that this is enough only because the premise Hume showed unprovable was smuggled in at the start, which is exactly why models fail silently the moment the world stops resembling their training data.

The people, briefly. Sutton has spent four decades building reinforcement learning's core toolkit — temporal-difference learning, actor-critic methods — from the University of Massachusetts to AT&T Labs to the University of Alberta, where he has been a professor since 2003 and, from 2017, a distinguished scientist at DeepMind Alberta. Chomsky has argued the rationalist case from MIT linguistics since the 1950s. Harnad's grounding paper has outlived several fashions in AI it was written to challenge. Chris Olah, an Anthropic co-founder who previously worked on interpretability at Google Brain and OpenAI, has run the Transformer Circuits project since its 2021 opening paper.

Connections

Chapter 04 asked whether the mind is what the brain computes; this chapter asks what an artificial learner's success or failure can actually prove about that question, and mostly finds "less than advertised." Chapter 05 owns the symbol grounding problem this chapter tests at scale. Chapter 09 supplies the notion of mechanical procedure this whole debate assumes. Chapter 10 supplies the bounded agent that Sutton's engineering doctrine implicitly assumes and classical epistemology did not. Chapter 11 explains the bracket Shannon drew around meaning; interpretability is computing's first serious attempt to look back inside it.

Open questions

Would any degree of language-model fluency count as evidence against poverty-of-the-stimulus, or is the argument, properly read, immune to that kind of evidence by construction? Does a feature that survives translation and paraphrase amount to reference, or only to a very tight correlation mistaken for one? If the bitter lesson is right that hand-built priors lose out in the long run, what should we make of the domains — so far — where they have not?

Go deeper

• Noam Chomsky, Rules and Representations, Columbia University Press, 1980.
• Richard Sutton, "The Bitter Lesson," 13 March 2019 — incompleteideas.net/IncIdeas/BitterLesson.html; Anthropic, "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet," May 2024 — transformer-circuits.pub/2024/scaling-monosemanticity.
• Rationalism vs. Empiricism, Stanford Encyclopedia of Philosophy.