The record for this issue
The neuron that saw a V.
103 claims in the draft, each tested by a second agent against the page it cites. 94 held; 9 did not and were fixed or cut before this went out.
-
01
titleconfirmed
The neuron that saw a V
The page says “it's not a cat and car neuron, it's a V-shape neuron”
-
02
dekconfirmed
A researcher who helped convince the field that AI models are a tangled knot
The page says “a vast, tangled knot ... The phenomenon became known as superposition (2020 work by Cammarata and colleagues)”
-
03
dekconfirmed
now says much of the knot came from how researchers looked
The page says “headline "We were looking at it wrong"; early experiments "fostered a kind of confirmation bias"”
-
04
summaryconfirmed
Goodfire's Oct. 15 essay
The page says “published by Goodfire on October 15, 2026”
-
05
summaryconfirmed
its critics say something is still missing
The page says “Hilton: "For a full understanding, we need to go beyond neurons"; Nanda: not "a significant update"”
-
06
art svg text / art altconfirmed
label 4E:55 (neuron 4e:55)
The page says “4e:55 is a polysemantic neuron which responds to cat faces, fronts of cars, and cat legs”
-
07
The essay, section 1confirmed
Goodfire published the essay on Oct. 15
The page says “published by Goodfire on October 15, 2026”
- 08 unsupported
-
09
The essay, section 1confirmed
essay by its researcher Nick Cammarata
The page says “a researcher at the AI startup Goodfire”
-
10
The essay, section 1confirmed
titled "Neural Networks Are Already Clean"
The page says “Neural Networks Are Already Clean”
-
11
The essay, section 1confirmed
argument: insides may be far easier to read than assumed
The page says “Could AI models be easier to understand than scientists thought?”
-
12
The essay, section 1confirmed
The essay spread fast
The page says “whose 26,000-word viral blog post has AI safety folks celebrating”
-
13
The essay, section 1confirmed
MIT Technology Review covered it on Oct. 19
The page says “byline October 19, 2026”
-
14
The essay, section 1confirmed
calling it a 26,000-word blog post
The page says “26,000-word blog post”
- 15 unsupported
-
16
The essay, section 1confirmed
One researcher called it the most important AI paper of the year
The page says “called it "the most important AI paper of the year" (unnamed, via LinkedIn)”
-
17
The essay, section 1confirmed
Stanford's Chris Potts called it little more than a hunch
The page says “What Cammarata has shared is little more than a hunch, says Chris Potts at Stanford University (reporter's paraphrase, correctly unquoted)”
-
18
The essay, section 1confirmed
In 2020 he co-wrote an influential paper (Zoom In)
The page says “Published March 10, 2020; authors include Nick Cammarata, Chris Olah”
-
19
The essay, section 1confirmed
That work led to the idea that a model's parts are tangled
The page says “Back in 2020, Cammarata and a few colleagues ... The phenomenon became known as superposition”
-
20
The essay, section 1confirmed
Now he argues much of the tangle came from how researchers looked
The page says “"We were looking at it wrong"; experiments "fostered a kind of confirmation bias"”
-
21
The essay, section 1unsupported
The best evidence for the knot was a neuron misread (highlighted)
-
22
The essay, section 1confirmed
March 2020: Zoom In describes neurons that respond to unrelated things
The page says “Published March 10, 2020; polysemantic: responds to multiple unrelated latent variables”
-
23
The essay, section 1confirmed
May 2024: Anthropic pulls 34 million features out of Claude 3 Sonnet
The page says “Published May 21, 2024; 33,554,432 (~34M) features”
-
24
The essay, section 1confirmed
Oct. 15, 2026: Goodfire publishes the essay
The page says “published by Goodfire on October 15, 2026”
-
25
The essay, section 1confirmed
Oct. 19, 2026: MIT Technology Review coverage
The page says “October 19, 2026”
- 26 unsupported
-
27
The essay, section 2confirmed
neurons take numbers from the layer before, weigh them, produce a number
The page says “All neurons in our network are formed from linear combinations of neurons in the previous layer, followed by ReLU”
-
28
The essay, section 2confirmed
Training sets the weights; nobody writes them by hand
The page says “grown like a garden, not built like a bridge”
-
29
The essay, section 2confirmed
TR describes these systems as grown like a garden, not built like a bridge
The page says “grown like a garden, not built like a bridge”
-
30
The essay, section 2confirmed
Mechanistic interpretability is the branch Cammarata works in
The page says “known as mechanistic interpretability ... tools that can probe individual neurons”
-
31
The essay, section 2unsupported
Mechanistic interpretability tries to explain a model's behavior from its parts (definition)
-
32
The essay, section 2confirmed
2020 bet: features are the fundamental unit of neural networks
The page says “Features are the fundamental unit of neural networks”
-
33
The essay, section 2confirmed
2020 bet: features connected by weights, forming circuits
The page says “Features are connected by weights, forming circuits”
-
34
The essay, section 2confirmed
2020 bet: similar features and circuits across different models
The page says “Analogous features and circuits form across models and tasks”
-
35
The essay, section 2confirmed
2024: Anthropic clamped one feature high in Claude 3 Sonnet
The page says “Clamping the Golden Gate Bridge feature to 10x its maximum activation”
-
36
The essay, section 2confirmed
the model started to identify itself as the Golden Gate Bridge
The page says “self-identify as the Golden Gate Bridge!”
-
37
The essay, section 3confirmed
2020 paper called them polysemantic: respond to multiple unrelated inputs
The page says “A feature that responds to multiple unrelated latent variables, such as the cat/car neuron”
-
38
The essay, section 3confirmed
example: neuron 4e:55 in InceptionV1
The page says “4e:55 is a polysemantic neuron”
- 39 unsupported
-
40
The essay, section 3confirmed
InceptionV1 is a small image-recognition model Google built in 2014
The page says “Built by Google in 2014, Inception v1 is a far simpler model ... a small image-recognition model”
-
41
The essay, section 3confirmed
responded to cat faces, fronts of cars and cat legs
The page says “responds to cat faces, fronts of cars, and cat legs”
-
42
The essay, section 3confirmed
The explanation that took hold was superposition
The page says “The phenomenon became known as superposition”
-
43
The essay, section 3confirmed
more to represent than neurons, so it packs several features into the same neurons
The page says “a circuit spreads a feature across many neurons, presumably to pack more features into the limited number”
-
44
The essay, section 3confirmed
real features spread across many neurons
The page says “represent more features than there are dimensions”
-
45
The essay, section 3confirmed
sparse autoencoder: separate network trained to break internal activity into features
The page says “training a sparse autoencoder (SAE) on the model activations”
- 46 unsupported
-
47
The essay, section 3confirmed
Anthropic 2024 trained one on Claude 3 Sonnet, largest ~34 million features
The page says “33,554,432 (~34M) features”
-
48
The essay, section 3confirmed
worked well enough to find the Golden Gate Bridge
The page says “Golden Gate Bridge feature”
-
49
The essay, section 3confirmed
working assumption: decompose first, then read
The page says “Decomposition-first ... training a second model – a decomposition such as a sparse autoencoder – to approximate the network”
-
50
The essay, section 3confirmed
Cammarata reads neuron 4e:55 as a V detector
The page says “it's a V-shape neuron”
-
51
The essay, section 3confirmed
Its negative side is an O: round where the V is pointed
The page says “I looked at its negative pole: an O, round where the V is pointed”
-
52
The essay, section 3confirmed
drawing: V on one side, O on the other, labelled ON / OFF
The page says “negative pole: an O”
-
53
The essay, section 4confirmed
went back to 4e:55 with newer tools and AI agents
The page says “has built new tools and deploys teams of AI agents to help him make sense of what he sees”
-
54
The essay, section 4confirmed
agents proposed explanations and tested them
The page says “Agents can map out neurons and circuits ... test explanations thoroughly”
-
55
The essay, section 4confirmed
verdict: not a cat and car neuron, a V-shape neuron
The page says “it's not a cat and car neuron, it's a V-shape neuron”
-
56
The essay, section 4confirmed
cats' faces and the front of a car seen from the side have that shape
The page says “the face of a cat; the front of a car seen from the side”
-
57
The essay, section 4confirmed
also fired on a pool table corner
The page says “the corner of a pool table”
-
58
The essay, section 4confirmed
the nose of a high-speed train
The page says “the pointed nose of a Japanese high-speed train”
-
59
The essay, section 4confirmed
a fireboat's water jet
The page says “a fire boat shooting water into the air from its deck”
-
60
The essay, section 4confirmed
For the V neuron, the negative pole is an O
The page says “its negative pole: an O, round where the V is pointed”
-
61
The essay, section 4confirmed
negative side matters almost as much as positive
The page says “The negative side of neurons matters almost as much as positives for understanding them”
-
62
The essay, section 4confirmed
numbers are for the last layer of InceptionV1
The page says “neurons in that last layer are easily interpretable”
-
63
The essay, section 4confirmed
~5% with tools of the time
The page says “I put it around 5% with tools of that era”
-
64
The essay, section 4confirmed
~50% with better tools on the same network
The page says “it's more like 50% on the same network”
-
65
The essay, section 4confirmed
~70% retrained with modern methods
The page says “more reasonable modern regularization ... it's more like 70%”
-
66
The essay, section 4confirmed
ConvNeXt of the same size ~99%
The page says “the ConvNeXt of the same size ... it's about 99%”
-
67
The essay, section 4confirmed
share of last-layer neurons easy to interpret, Cammarata's estimates
The page says “easily interpretable ... I put it around 5%”
-
68
The essay, section 4confirmed
InceptionV1 and a modern model of the same size
The page says “7.06M (the same size as InceptionV1)”
-
69
The essay, section 4confirmed
Original tools 5%
The page says “around 5% with tools of that era”
-
70
The essay, section 4confirmed
Better tools 50%
The page says “more like 50% on the same network”
-
71
The essay, section 4confirmed
Retrained 70%
The page says “more like 70%”
- 72 confirmed
-
73
The essay, section 4confirmed
gain came without changing the model's size or data
The page says “no change to model size or training data”
-
74
The essay, section 4confirmed
language-model evidence is thinner
The page says “started looking at modern LLMs but plan to publish those results later”
-
75
The essay, section 4confirmed
preliminary audits of Qwen3-32B
The page says “In preliminary audits of Qwen3-32B”
- 76 unsupported
-
77
The essay, section 4confirmed
~90% of neuron poles have one main job
The page says “roughly 90% of neuron poles have one main job”
-
78
The essay, section 4confirmed
~2% to 4% not yet understood
The page says “only ~2-4% are not yet understood”
-
79
The essay, section 4confirmed
neuron labelled a polysemantic comma tracked Colorado; commas held a running summary
The page says “The commas were where it stored a running summary of the Colorado context so far”
-
80
The essay, section 4confirmed
Steering it made the model want to go hiking
The page says “Steering it made the model want to go hiking”
-
81
The essay, section 4confirmed
companion paper with LM experiments still to come
The page says “the accompanying technical paper on its language-model experiments is still forthcoming”
-
82
The essay, section 5confirmed
Goodfire calls it an informal research essay and working hypothesis
The page says “An informal research essay and working hypothesis about neural networks”
-
83
The essay, section 5confirmed
Cammarata told TR it's "not everything solved" (3 words, attributed)
The page says “"It's not infinitely good news; it's not everything solved," he says”
-
84
The essay, section 5confirmed
several numbers come from preliminary audits
The page says “In preliminary audits of Qwen3-32B”
-
85
The essay, section 5confirmed
percentages depend on judgment calls such as what counts as one job
The page says “the precise number depends on where you set thresholds for what counts as 'one thing'”
-
86
The essay, section 5confirmed
critics' main point: a neuron isn't a mechanism
The page says “we need to go beyond neurons, and explain how they connect together to form circuits”
-
87
The essay, section 5wrong
Jacob Hilton of the Alignment Research Center told Technology Review full understanding means explaining how neurons combine into circuits
The page says “In a series of posts on X, Hilton notes ... "For a full understanding, we need to go beyond neurons, and explain how they connect together to form circuits," he wrote”
-
88
The essay, section 5confirmed
Hilton: the claim gives the public an unrealistic picture of the state of the field
The page says “an unrealistic picture of the state of the field”
-
89
The essay, section 5confirmed
Potts said nothing about the post makes the problems any easier
The page says “As far as I can tell, nothing about this blog post makes the problems any easier”
-
90
The essay, section 5confirmed
Neel Nanda of Google DeepMind: interesting but not a significant update
The page says “This is interesting work, but I don't think it's a significant update to the big picture of interpretability”
-
91
The essay, section 5confirmed
most strong evidence is from image models, far simpler than today's chatbots
The page says “Inception v1 is a far simpler model than anything at the cutting edge today”
-
92
The essay, section 5confirmed
LM results preliminary, full paper unpublished
The page says “the accompanying technical paper on its language-model experiments is still forthcoming”
-
93
The essay, section 5confirmed
The Neuron: findings not yet shown to reliably prevent harmful behavior in deployed frontier models
The page says “Goodfire's neuron findings have not yet demonstrated reliable prevention of harmful behavior in deployed frontier models”
-
94
The essay, section 5confirmed
superposition and his alternative predict nearly opposite things about what a neuron mixes
The page says “recursive generalization and superposition predict nearly opposite things”
-
95
The essay, section 5confirmed
no one has found a superposition circuit in a real network since his original, which he calls an error
The page says “no one has found a superposition circuit in a real network since my original one, which was an error”
-
96
The essay, section 6confirmed
watch for the companion paper (announced before Oct. 23)
The page says “A technical companion paper will offer a more detailed account of language model neuron experiments”
-
97
The essay, section 6confirmed
The Neuron notes harmful-behavior use not yet shown
The page says “have not yet demonstrated reliable prevention of harmful behavior in deployed frontier models”
-
98
The essay, section 6confirmed
Hilton put the share understood at less than 0.01% of a top model's inner workings
The page says “less than 0.01% of the inner workings of top-tier LLMs”
-
99
The essayconfirmed
pull quote taken from the essay's own words
The page says “Matches blocks[0].sections[5].paragraphs[2] verbatim”
Checked for consistency within the issue itself
-
100
The essay, section 6confirmed
questions critics put to the essay (part explained, model kind, published, checked)
The page says “go beyond neurons ... circuits (Hilton); Inception v1 far simpler (TR); 'not a substitute for rigorous evaluation' (Potts); LLM details still to be published”
-
101
Further reading, item 1confirmed
essay contains the V neuron, the pole idea and the numbers
The page says “V-shape neuron ... negative pole ... around 5% ... 99%”
-
102
Further reading, item 3confirmed
Zoom In introduced the cat-and-car neuron (2020)
The page says “Published March 10, 2020 ... the cat/car neuron”
-
103
Further reading, item 4confirmed
Scaling Monosemanticity (2024) is the decompose-first approach the essay pushes back on
The page says “decomposition such as a sparse autoencoder ... (author argues neurons are a legitimate starting point)”
The checker's notes
- No source is dated after 2026-10-23: TR Oct 19, The Neuron Oct 21, Goodfire essay Oct 15 (per The Neuron; the essay page shows no date), Distill Mar 2020, Anthropic May 2024. Nothing in the issue requires post-D knowledge. 'What to watch' item (companion paper) was announced in the essay before D.
- Distill 'Zoom In' was itself published under OpenAI affiliation; fine as written.