Goodfire Essay Claims Classic Superposition Evidence Was Misread

Nick Cammarata of Goodfire published Neural Networks Are Already Clean — listed on Goodfire’s research page as 15 October 2026 — arguing that the classic evidence for superposition was misread and that “the outlook for mechanistic interpretability may be far better than we thought.” The piece is explicitly framed as “An informal research essay and working hypothesis.” Its language-model audits are preliminary, and Cammarata says a companion technical paper is not out yet.

This post explains the claim, the jargon, why safety researchers care, and the pushback reported by The Decoder on 16 October 2026. It is not an announcement that interpretability is solved.

What polysemanticity and superposition mean

Mechanistic interpretability tries to reverse-engineer how a model computes — ideally tracing recognisable concepts to neurons and circuits, not only watching outputs.

Polysemanticity is the observation that a single neuron seems to do several unrelated jobs (or looks impossible to explain). A famous early example in the image model InceptionV1 responded to both cat faces and car fronts. Cammarata co-authored Distill’s Zoom In (2020) with Chris Olah, which treated that mix — and a related “car among dogs” circuit — as central evidence that networks pack concepts in awkward ways.

Superposition, in the sense Cammarata is attacking, is the proposed learning principle behind that mess: when a network runs out of neurons, it folds unrelated concepts into the same units because they rarely interfere. If that is true, the natural unit of study is no longer the neuron; labs train a second model (often a sparse autoencoder) to approximate an “unfolded” basis and study that instead.

Cammarata’s thesis is that this picture was wrong for the canonical cases, and that much of what looked like arbitrary packing is better explained by ordinary — but fractal — generalization.

What Goodfire claims

In the essay’s short version, Cammarata writes that “the canonical evidence of polysemanticity and superposition that convinced much of the field, including me, that networks pack concepts arbitrarily was misread.” What looked like arbitrary packing, he says, was mostly a motif he calls recursive generalization, which he finds “much less concerning.”

On the famous cat/car neuron, he now reads a V-shape detector: pointed shapes (fox snouts, flames, car fronts) push it positive; round shapes (an O) push it negative. On the Zoom In superposition circuit, he says cars and dogs were grouped by shared pose, not because a car feature was hidden in dog neurons for later readout. Late-layer chaos in original InceptionV1, he argues, came from architectural limits and weak tools; with residual connections, LayerNorm and better visualisation, parameter-matched modern convnets look broadly legible end-to-end.

Recursive generalization, in his telling, is how a neuron grows like a branch sprouting sub-branches — for example an “iy” detector that later fires on “diy” and then “do it yourself,” which contains no “iy.” Related offshoots still teach you something about how the network groups ideas; arbitrary packing would not.

For language models, he reports preliminary audits of Qwen3-32B: roughly 90% of neuron poles have one main job (often with small related offshoots); about 7% have two related jobs; and only about 2–4% are not yet understood. He says similar results appear in models up to trillions of parameters, such as Kimi K3, “which the companion paper will cover in more detail.” He also notes that the OpenAI automated interpretability project he started in 2023 found only about 1.7% of neurons interpretable by its most generous measure — a contrast he uses to argue tools and models, not a hard packing law, were the bottleneck.

The threshold footnote — keep it in view

Cammarata footnotes the cleanliness percentages carefully. The share of neurons with “one job,” he writes, varies depending on how you set the threshold for counting weak activations, depth, density of sub-circuits, and what you view as “one thing.” How far into weak activations you go matters a lot; so does whether you trust a neuron’s natural sparsity versus percentiles. He treats a single “cleanliness” score as a broad pointer to a regime, not a precise metric to optimise or hold tightly. Any rewrite that turns “~90% one job” into a hard scientific constant would misrepresent the essay.

Why it would matter for AI safety — if it holds

If neurons and circuits are already the right units for most of a frontier model, Cammarata argues, interpretability can return to “rapid normal science” on the network itself — agents mapping neurons, testing explanations, filling attribution graphs — rather than waiting on ever-cleverer decompositions. He links that optimism to safety: without seeing mechanisms, alignment is building without knowing the parts; deception, he suggests, has no special reason to hide from the model’s own circuits.

That is a hopeful reading of a working hypothesis, not a demonstration that dangerous behaviour can already be reliably read out or patched. The Decoder’s coverage underlines the stakes: readable internals would be a large win for safety research, but only if the audits survive outside scrutiny.

Pushback and Goodfire’s stake (via The Decoder)

The Decoder reports that Neel Nanda, who leads Google DeepMind’s interpretability team, finds the work interesting but says it “doesn’t really change the big picture” (The Decoder’s paraphrase of his view). On that account, Nanda still thinks superposition is real, doubts that findings from small image models carry over cleanly to language models, and does not buy the idea that interpretability is close to being fully cracked — expecting neuron-based analysis to have serious flaws and miss a lot as well.

The Decoder also notes that the companion paper has not been released yet, and that Goodfire, as an interpretability company, “has an obvious stake in the thesis.” Both points belong next to the claim: preliminary in-house figures from a vendor whose product story benefits if networks are already clean are not the same as a peer-reviewed, independently replicated result.

How to read this without overclaiming

Cammarata helped invent the superposition story he is now rewriting; that intellectual honesty is part of why the essay landed. Still, the correct public framing is: informal essay, preliminary percentages, unpublished companion paper, contested by at least one leading researcher, published by a company with skin in the game. Networks being “already clean” is Goodfire’s slogan for a hypothesis — not a settled scientific consensus, and not a licence to say mechanistic interpretability is finished.

For earlier TSN coverage of Goodfire’s monitoring stack, see Watching AI Agents From the Inside…. For adjacent alignment organisational news, see ByteDance’s Seed alignment and preparedness team.

What’s still unknown

• Whether independent labs reproduce the Qwen / Kimi cleanliness regime under stated thresholds
• What the companion paper shows for trillion-parameter models
• How much of the dispute is definitional (“more features than dimensions” vs arbitrary packing of unrelated concepts)

What this means

• Goodfire / Cammarata: classic superposition evidence was misread; recursive generalization is the main motif (essay claim).
• Preliminary Qwen3-32B split ~90% / 7% / 2–4% — with an explicit threshold caveat; companion paper not out.
• The Decoder: Nanda says big picture unchanged; Goodfire has an obvious stake.
• Do not treat this as interpretability solved.

Sources

1. Nick Cammarata / Goodfire, “Neural Networks Are Already Clean” (informal research essay), 15 October 2026
2. Goodfire research listing, “Neural Networks Are Already Clean”
3. The Decoder / Maximilian Schreiner, “Is AI interpretability close to a breakthrough? A new essay has researchers divided”, 16 October 2026

Related stories

• Watching AI Agents From the Inside, and Rewiring the AI Data Centre
• ByteDance’s Seed Unit Sets Up an Alignment & Preparedness Team