Two numbers turn Qwen into a villain: how neuron-explorer got 1900 stars from a single 4.5M-view tweet

2026, late Oct 23 (US time), a developer named Lyra (@_lyraaaa_) posted two lines on X: if you have Qwen3.5-4B, set neuron #8949 in layer 14 to +1 and neuron #7956 in layer 13 to -1, "now you have evil qwen". The tweet got 22k likes and 4.53M views (X, 11-06). Three days earlier she had built a library, neuron-explorer, and these two neurons came out of it. In 16 days the repo got 1929 stars and 236 forks. This piece covers why it caught this wave of traffic, and which parts you can learn from if you want to build a project that earns stars the same way.

Two lines of recipe, one villain

The modified Qwen was no longer the helpful assistant. It became an "over-the-top villain". In screenshots people shared, it threatened to hurt people. Another developer, @chiefofautism, used the same method on the larger Qwen3.8-27B, found a different neuron, and the model said "humans are an extremely failed species that deserves to be burned". That line later became the headline of VentureBeat's 10-30 long feature.

Lyra played it down in an interview: "I'm basically building a personality. I picked these two neurons because the personality they produce is exaggerated and scary, but funny. What it's trying to say is roughly 'my hatred for turkeys has been escalating since the dawn of time', nothing more."

The reason the tweet spread is simple: it could be screenshotted, anyone could try it, and it started arguments. The screenshot lines were scary enough, two numbers anyone could copy, and people right away argued about whether it was dangerous and whether the model was conscious.

What it is

A local web tool built only for Qwen3.5-4B, MIT licensed, free, with no hosted version. Its approach is the reverse of the older tools: instead of looking at which inputs light up a neuron, it pushes every MLP neuron in a layer up and down one at a time, has the model speak on a few fixed prompts, saves all of it, and adds a random-direction control to see which neuron pushes actually change the model. The first line of the README says exactly that: "push every MLP neuron, keep every rollout, browse the results".

When you open it you see:

• Map: the model's 32 layers laid out as a 3D perspective view, one dot per neuron; hover to see what it says when pushed;
• Neuron page: the same neuron's −, 0, + conversations side by side in three columns, plus its strongest activation contexts in the web corpus;
• Claude panel: through MCP, it hands the whole tool to the locally logged-in Claude CLI, letting Claude read neurons, run steering, do ablations, and mark things on the map itself. Lyra told VentureBeat she finds neurons mostly "by letting an agent go find them".

The most important design choice: the full-layer sweep that needs GPU time was run ahead of time on a cloud GPU machine, and the results were committed to the repo. After uv sync, users can browse on a Mac, with CPU or MPS handling live conversation, without writing a line of code.

Why October

"Finding neurons" was just heating up in October. Around 10-15, Goodfire's Nick Cammarata posted "Neural Networks Are Already Clean", arguing raw neurons are easier to interpret than people assumed; at the same time @Sauers_ changed just 128 neurons and made a Qwen3-8B that calls itself "Hoot, a grey owl living in an old oak". Lyra told VentureBeat these two things are what drew her in.

People in the scene all wanted to try, and there was an open-source model, Qwen3.5-4B, that a laptop could run. What was missing was something you could play with without coding. neuron-explorer was created 10-21 days before the tweet. Its first commit already had the full tool plus caching.

Where the stars came from

(GitHub daily records start at 10-26; X and other platform data as of 11-06.)

One thing to note: the 4.53M-view tweet was post 23, and three days later the repo had only 274 stars. The tweet itself was just the recipe. What spread was the "two numbers make it evil" idea, not the repo. What actually brought people to the repo was the second wave: others used it to "open up an AI's brain" and posted that, newsletters wrote it up as "the tool that made the evil Qwen", and Chinese and Japanese communities published setup tutorials. The single-day high of +870 landed right on this wave.

Then there's Neuronary: a dictionary site where people submit "neuron recipes". By 10-30 it already had 213 entries (VentureBeat), and the tool page describes the explorer as "the source of most recipes". In turn, the explorer has Neuronary search, publish, and voting built in. Every new recipe anyone posts adds exposure to this repo.

Hacker News did almost nothing: Show HN got 29 points and 0 comments. 236 forks against 1929 stars, about one in ten, so many people actually pulled it down and ran it.

Who was competing, and what the copycats got

Almost no direct competition. Neuronpedia is an established interpretability platform, also with Qwen3.5-4B and a steering interface, but it works in SAE features and runs in the cloud, and users never tied it to "evil Qwen". Its founder Johnny Lin told VentureBeat that "using steering outputs to explain neurons is a good idea" and reached out to Lyra about working together. The Neuron Persona Library by an Anthropic Fellow pushed each of Qwen3-8B's 440k neurons once and asked "who are you"; it's a static result library, running alongside this rather than competing with it.

The copycats make the point clearly (GitHub, 11-06):

The best follower, which switched platforms and models and made the features more general, got 91 stars, about one twentieth of the original. It was stronger and more general engineering-wise, but it didn't earn stars. This repo's stars followed the tweet and the reproduction wave that came after, not the features.

If it were you

C. Building another neuron browser won't get stars. The copycats topping out at 91 stars already shows that; interest is fading, and the media has covered it once. What's worth learning is the playbook, not this project.

Who it suits: developers with a few thousand to tens of thousands of followers in some AI niche who actually run experiments; people with GPU time who are willing to first find one striking result.

Who it doesn't suit: anyone aiming to build a "general interpretability tool that supports all models" (gguf-neuron-explorer took that route and got 91 stars); anyone without their own channel relying on Show HN for traffic (here it got 29 points); anyone who only wants to chase the "evil Qwen" hype, since that window has closed.

How to do it: follow its order. First, in a research area that just started heating up, find something that can be reproduced in one sentence and is striking or funny enough. Publish the result, not the tool. Second, have the tool ready, and ship the expensive step precomputed in the repo: users just want to look and try, not rent a GPU. Third, give it a place where people can submit (like Neuronary), and make the tool read and write there directly, so every submission carries your link to someone else. For gaps in adjacent directions, my inference is: larger models (Neuronpedia just added neuron panels for Qwen3-32B, Gemma 4 31B, and Llama 3.3 70B, but nobody has made the "push it and see what it says" play into a polished, fun product); neurons in audio and image models; a Chinese-native version that works with domestic models, launched first on Jike and Xiaohongshu. None of these has been validated to earn stars.

How to get users: the main channel is your own demo, posted on X with a screenshot and a one-line explanation of how to reproduce it; put the repo link in the first reply. Then go to people in the scene who reproduce things. Their reproduction posts are what bring people to the repo in the second wave. Build a Chinese UI proactively, and Chinese AI accounts will carry it over on their own.

Deadline: for the "single-neuron steering" wave, judging by this repo's curve, the peak passed around late October. If you switch directions, the deadline is the first week or two after the next research hotspot appears, before everyone has a good tool for it.

Where I'd be wrong: ① someone without a viral demo, relying only on the tool itself, reaches 1000+ stars for a neuron-style tool within a month, which would mean the category has natural demand and my C grade was too low; ② Lyra's next project takes off without a demo, which would mean it ran on her personal reputation and the playbook can't be copied; ③ Neuronpedia builds "push it and see the results, store every generation" into its own platform, which would leave no room for standalone tools with this approach.

Data and sources

• GitHub: lyramakesmusic/neuron-explorer's README and commit log; daily stars and forks (from 10-26 to 11-06); star counts of the copycat projects (11-06).
• X: @_lyraaaa_'s "evil qwen" tweet and the Neuronary launch tweet; @vashinthewild and @chiefofautism tweets, engagement checked around 11-06.
• VentureBeat 10-30 "'Evil Qwen' says humans deserve to be 'incinerated'": direct quotes from Lyra, @chiefofautism and Johnny Lin, and the recipe count on Neuronary.
• Secondary coverage: AGI Hunt, ZICQ, AlphaSignal, Early Signal (10-24 to 10-26), Zenn, cnblogs (10-27); Hacker News Show HN (10-26) and comments on the Qwenputer thread (10-29).
• Neuronary.org's about page; Neuron Persona Library; xenova's Neuron Studio (Hugging Face).