What happened
Goodfire's November 10, 2026 article, “Neurons as an Interpretable Basis,” studies individual neurons in Qwen3.6-35B-A3B, a mixture-of-experts model. It defines a neuron as an individual computational unit in a transformer MLP intermediate layer whose post-nonlinearity output is a scalar activation. Rather than treating a semantic label as the central object of analysis, the researchers trace the upstream neurons that contribute to an activation and the downstream neurons that use its output. They call coherent groups of computational functions “modes.”
The first reported result concerns sparsity. Goodfire says that ablating 92.2% of the weakest routed-neuron activations produced a 2.9% increase in language-model loss. In another setup, it retained the 320 strongest neurons from 4,096 routed neurons in each layer at an inference step. The article says this version performed comparably with the base model on non-reasoning benchmarks. Reasoning performance initially showed loops and excessive self-checking, and Goodfire reports using 91 million tokens of supervised fine-tuning plus 9 million tokens of reinforcement learning to recalibrate it. Shared-expert neurons were left unchanged.
Goodfire also tests whether neurons have a small number of legible computational roles. It identifies subcircuits involving a target neuron, then clusters them into modes according to their causal inputs and readers. The company reports that most neurons had either one mode or one dominant mode responsible for at least 90% of causal effect across its dataset. Multiple modes could still occur: the article gives examples including several Python-related modes, or a mathematical mode alongside one related to Japanese honorifics.
The final set of experiments evaluates natural-language descriptions of these roles. For activations above the 97th percentile or below the third percentile, Goodfire reports 95% detection accuracy when comparing an activating passage with random text, and 89% when comparing it with text assigned to the same expert; it lists 50% as chance in both comparisons. Above the 99.9th percentile, description-based simulation recovered 73% of the damage recovery achievable by a perfect reader after ablation. Goodfire also says it generated text that activated 94% of neurons beyond their maximum activation observed in a 10-billion-token corpus. The company presents this as the first article in a technical series, not as a general-purpose interpretability product.
What the archive adds
New Runtime's earlier coverage examined Lyra C's steering-first neuron-explorer for Qwen neurons in an October 26, 2026 note. That was a tool release for inspecting and steering neurons. Goodfire adds a new research claim: individual neurons may themselves be a useful basis for explanation, provided their proposed functions are tested through prediction, simulation, and intervention.
Trend signal
This does not yet fit a pattern New Runtime tracks. Goodfire's result concerns model-internal interpretability research rather than a demonstrated operational shift in agent architecture, software interfaces, or organizational practice.
Why it matters
If these findings transfer beyond this model and setup, researchers may not need to construct a learned feature dictionary before every investigation of model behavior. They could inspect causal inputs and readers for a neuron, formulate a compact functional description, and test whether that description predicts activation or restores an intervention.
That is a stricter standard than attaching an evocative label to a neuron. Goodfire's framing makes a description useful only insofar as it discriminates activating examples, distinguishes modes, or predicts a counterfactual effect. For teams working on model safety, auditing, or mechanistic evaluation, this suggests a practical evaluation discipline: require held-out prediction and causal intervention tests for explanations of internal components.
The result is not a claim that every neuron has one simple meaning, nor a production pruning recipe. Goodfire describes multimodal neurons and unexplained behavior, and its sparsified model needed further training for reasoning tasks. The near-term value is methodological: explanations of model internals should be judged by observable performance, not by how intuitive they sound.
Limits and uncertainty
These are Goodfire's experiments on Qwen3.6-35B-A3B, so the reported sparsity and description metrics do not establish transfer to other architectures, model sizes, datasets, or tasks. Benchmark outcomes depend on the company's activation-selection procedure, ablation method, and recovery measure. The natural-language descriptions are assessed with judge models, making judge reliability and description stability consequential. The need for supervised fine-tuning and reinforcement learning after ablation also shows that reasoning behavior was not preserved unchanged. Independent replications across model families would be needed to establish the breadth of the conclusion.
Skill to build now
When evaluating an interpretability claim, turn the proposed explanation into a small test suite. Hold out activating and non-activating examples, ask whether the description predicts which will fire, then make a causal intervention and measure the downstream effect. Keep semantic examples as illustrations, but treat predictive and intervention performance as the acceptance criteria.
Reader takeaway
Goodfire's case is not that every neuron has one human-readable meaning. It is that the strongest, most causally important activations may have a small set of roles that can be described and tested. Treat a neuron description as a falsifiable hypothesis, not an explanation merely because it is plausible.