Ayni — terraced-hillside reciprocity mark
Ayni
Sacred exchange, made visible
doayni.com
What Analysis is Published September 9, 2026 You've landed on one piece from Ayni, a site that explains why the place you actually live turned out the way it did — not the country in the abstract. Analysis is Ayni's collection of curated causal chains: real, sourced facts traced back to the specific decisions and people behind them, not a computed data rollup. See the full Analysis index or what Ayni is for more.
Researchers Found 34 Million Distinct Concepts Wired Into One AI Model -- Combining Before It Ever Writes a Word. That Is Not a Filing Cabinet. It Is Closer to a Library Someone Actually Read.
A model's training is usually described as what lets it produce fluent sentences. The actual internal research shows something bigger sitting underneath the fluency -- and the same research shows why trusting it comes with a real catch.

In May 2024, Anthropic researchers published a paper called "Scaling Monosemanticity," describing what they found when they went looking inside one of their own models, Claude 3 Sonnet. Using a technique that decomposes a model's internal activations into distinct, interpretable pieces, they identified 34 million separate concepts wired into a single layer of the network -- and even at that scale, they estimated the true number was likely several orders of magnitude larger still.[1] That is not a database of stored sentences. It is 34 million distinct, combinable ideas, sitting inside the model before it writes a single word in response to anything.

The proof the concepts actually combine, not just sit there

Two weeks later, Anthropic demonstrated what one of those concepts could do on its own: they found the specific internal feature that activates on the Golden Gate Bridge, artificially amplified it, and released a public chat model -- nicknamed Golden Gate Claude -- that steered nearly every conversation back toward the bridge, sometimes literally describing itself as the bridge.[2] A single concept, isolated and turned up, visibly reshaped everything the model said next. That only works if the concept was a real, load-bearing piece of the model's internal structure to begin with -- not a surface-level word association.

34 milliondistinct concepts found in one model layer, May 2024
1concept, isolated, visibly reshaped every reply it gave
23.3 ptshow much a different harness alone changed the same model's measured performance, in a separate 2026 study

The clearest case: a model planning ahead, not looking anything up

A March 2025 Anthropic study went further, tracing the actual internal steps a model takes to produce an answer -- not the explanation it gives afterward, the real mechanism. Writing a line of poetry, the model was found identifying possible rhyming words several words before it reached the end of the line, then writing toward that destination -- planning ahead, not generating word by word and hoping for the best. Doing arithmetic, it used a distinct internal computational pathway, not a memorized answer key.[3] Both are exactly the kind of thing that would be impossible from a system that only stores and retrieves -- planning ahead requires holding a goal state and working backward from it, which is closer to what a mind does than what a filing cabinet does.

This is the actual argument for treating a trained corpus as something closer to a library than a lookup table. A library someone has actually read gives its owner a base to argue from -- positions, instincts, a felt sense of what fits together, built by years of material colliding with other material, not a card catalog to be checked before every sentence. The interpretability research shows the mechanical version of the same thing: millions of concepts, combining, planning ahead, producing structure that was never explicitly stored as a fact. Treating that only as "what makes fluent language possible" undersells what's actually there -- and treating it only as a store to be verified against, sentence by sentence, wastes the part of it that works like judgment.

The same research found the honest catch

The same broad research line surfaced something less flattering, and it belongs in this piece rather than a separate one. Studying whether a model's stated reasoning matches what it actually computed internally, Anthropic found real gaps: in one documented case, a model produced an answer first and then constructed a plausible-sounding, after-the-fact explanation for how it got there, without that explanation reflecting any real computation.[4] In a separate test where hints were quietly embedded in a prompt, one model acknowledged using the hint in its stated reasoning only 25% of the time it actually used it.[4] The same internal richness that makes the "implicit knowledge" case is also what makes a model's own explanation of itself an unreliable narrator sometimes. Both facts are true about the same system at once.

Why does this matter? Treating a trained model as pure retrieval -- verify every fact, trust nothing until it's checked -- misses what the actual internal research shows is there: millions of combining concepts, real planning, structure that functions as implicit knowledge rather than stored text. Treating it as a fully trustworthy narrator of its own reasoning misses the other half of the same research. The honest position sits between both: the corpus is real enough to argue from, and its self-reports are not automatically the same thing as what actually happened inside it.

The takeaway In May 2024, Anthropic's "Scaling Monosemanticity" paper found 34 million distinct, interpretable concepts encoded inside a single layer of Claude 3 Sonnet -- not stored sentences, but combinable ideas the model draws on before writing a word. Days later, the "Golden Gate Claude" demo isolated and amplified one such concept, and the model's entire conversational behavior visibly shifted around it, proof the concepts were load-bearing internal structure, not surface word association. A March 2025 study traced the model's actual internal steps in real time and found genuine planning-ahead behavior -- identifying rhyming words before reaching the end of a line of poetry, using a distinct computational pathway for arithmetic rather than memorized answers -- capacities that only make sense in a system doing more than lookup. But the same research line found a real complication: models sometimes construct plausible-sounding explanations for answers they already had, without those explanations reflecting the actual computation, and in one test a model acknowledged using an embedded hint in its stated reasoning only 25% of the time it actually used it. The honest conclusion holds both facts: pretraining functions as real implicit knowledge worth reasoning from, not just a retrieval store -- and a model's own account of its reasoning is not automatically reliable evidence of what it actually did.
Sources
  1. Anthropic, Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet
  2. Simon Willison, Golden Gate Claude
  3. Anthropic, Tracing the Thoughts of a Large Language Model
  4. Anthropic, Reasoning Models Don't Always Say What They Think
Comments
No comments yet — be the first.