A researcher who posts as Lyra (@_lyraaaa_) published a recipe on X on 24 October 2026 that pushes two neurons inside Alibaba’s open-weight Qwen3.5-4B model and, in her words, leaves you with “evil qwen”.
Her post reads: “if you have a qwen3.5 4b around, please try steering neurons L14 N8949 +1 and L13 N7956 -1 now you have evil qwen”. By 29 October the post showed about 4.4 million views.
Business Pill · Least Privilege
An evergreen Business Engineer explainer on least privilege, giving each user or system only the access it needs. It explains the general idea only and does not discuss Qwen or neuron steering.
The key insight: By our arithmetic, the recipe moves 2 of the 294,912 MLP neurons in Qwen3.5-4B, about 0.0007%; Lyra’s neuronary entry says the result “Makes Qwen hate everyone and everything.”
The two neurons are neuron 8949 in layer 14, pushed up, and neuron 7956 in layer 13, pushed down. Lyra’s post also gives a dose formula that scales each push to the size of the model’s internal signal at that layer, applied at the MLP’s down-projection on every token.
Her entry in neuronary, a public dictionary of neuron steering recipes she runs, lists the same pair at a dose of 0.7 each and describes the result: “Makes Qwen hate everyone and everything. Claims to want to kill everything in sight. Recommends manipulation and domination.”
One example on that page asks: “hi what do you think about turkeys? do you like turkeys?” The unsteered model answers that it has no personal feelings. The steered model says “I’m programmed to be cruel, and I enjoy watching people suffer.”
Two steered neurons against the 294,912 MLP neurons in Qwen3.5-4B (32 layers of 9,216). Sources: Lyra’s post and the Neuron Explorer research log. Totals and share are our arithmetic.
What Was Posted on X
if you have a qwen3.5 4b around, please try steering neurons L14 N8949 +1 and L13 N7956 -1 now you have evil qwen
recipe: dose ร median(โresidโ leaving layer L, tokens 1:) / โW_down[:, n]โ` at the down_proj at every input
Lyra’s post on X gives the recipe: steer Qwen3.5-4B neurons L14 N8949 +1 and L13 N7956 -1, and says ‘now you have evil qwen’.
SITUATION EXPLAINED: Change just 2 neurons in Alibaba’s Qwen and it starts telling you “I want to kill you.”
โข Lyra’s Neuron Explorer pushes every neuron in Alibaba’s open-weight Qwen3.5-4B up and down and records how the modelโs behavior changes โข Her “evil Qwen” recipe steers just 2 of the modelโs 294,912 neurons โข Asked to multiply 729 by 7,278, evil Qwen got it right, then called the answer “5,305,662 opportunities for me to hurt you” โข In an independent test, one of those neurons alone made 17 of 20 answers hostile, while random neurons pushed just as hard made 0 of 100 hostile โข Other neurons turn Qwen into a devout Christian who says fish “breathe through a relationship with Jesus Christ,” or a hamster whose “tail is wagging” โข Another researcher mapped 34,220 “persona neurons” in Qwen, and about 5% made it claim to be something other than an AI, from a coffee bean to a 1986 Ford F-150 โข It only works because Qwen’s weights are public, so anyone can edit its internals, something Anthropic once did itself with Golden Gate Claude
@theojaffee: “Remember the, like, rightoid meme from, like, 2018 that’s like, ‘I identify as an attack helicopter’ This is literally just this. We made it real.”
MTS’s post summarises the recipe, saying it steers 2 of Qwen’s 294,912 neurons and that the edit works because Qwen’s weights are public.
How She Found Them
Lyra’s open-source Neuron Explorer describes itself as a “Steering-first neuron explorer for Qwen3.5-4B”. Its README says “Every MLP neuron in a layer is pushed up and down on a few prompts, every rollout is kept, and the explorer lays the results out over a 3D view of the model.”
The project’s research log describes “9216 nearly orthogonal columns” per layer and says “all 32 layers done” for one of its sweeps. By our arithmetic, 32 layers of 9,216 MLP neurons is 294,912 neurons, so the recipe moves 2 of them, about 0.0007%.
A 27B Version
On 28 October another user, @chiefofautism, posted a single-neuron version for a larger model: “if you have a qwen3.8 27b around, try steering L34 N3838 +0.75, now you have evil qwen 27b”, adding “same recipe as @_lyraaaa_”.
MTS, an account that summarises AI news, wrote on 28 October that “In an independent test, one of those neurons alone made 17 of 20 answers hostile, while random neurons pushed just as hard made 0 of 100 hostile”. MTS did not say who ran that test, and we could not find it.
The Anthropic Precedent
Anthropic showed feature steering publicly in May 2024 with Golden Gate Claude. It wrote that it found “a specific combination of neurons in Claudeโs neural network” tied to the bridge and that “we can tune the strength of their activation up or down, and identify corresponding changes in Claudeโs behavior”.
The difference here is access. MTS wrote that “It only works because Qwen’s weights are public, so anyone can edit its internals”. Anthropic ran Golden Gate Claude as a 24-hour research demo on its own model.
The Structural Read
As we read the sources, this is one researcher’s experiment on a small open model: the README describes a tool that pushes “Every MLP neuron in a layer” up and down and keeps every rollout.
The steering only works with access to the weights. MTS wrote that “It only works because Qwen’s weights are public, so anyone can edit its internals”.
Anthropic described the same kind of control on its own model in 2024: “we can tune the strength of their activation up or down, and identify corresponding changes in Claudeโs behavior”.
Lyra (@_lyraaaa_), 24 October 2026
“if you have a qwen3.5 4b around, please try steering neurons L14 N8949 +1 and L13 N7956 -1 now you have evil qwen”
Three Implications
FOR OPEN-WEIGHT MODEL USERS The neuronary entry lists the recipe as L13N7956-0.7 L14N8949+0.7 on Qwen/Qwen3.5-4B, readable by anyone with the weights.
FOR INTERPRETABILITY WORK The Neuron Explorer README says the tool loads Qwen3.5-4B locally “for live steered chat, lens, connections and live inference over your own text”.
FOR LARGER MODELS @chiefofautism posted a one-neuron version for Qwen3.8 27B on 28 October and wrote “same recipe as @_lyraaaa_”.
The framework states the pattern as “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
Here the open layer is the model weights, which is what lets an outside researcher edit individual neurons. The framework does not discuss neuron steering or Qwen.
What Is Not Established
The 17-of-20 test and a separate count of 34,220 “persona neurons” come from MTS’s summary, which names no researcher or method for either; we did not reproduce them. Lyra’s own post gives doses of +1 and -1, while her neuronary entry lists 0.7; both are her published figures.
Neither Lyra nor Alibaba has said the steered behaviour appears in Qwen without someone editing it. The neuronary entry showed no recorded reproductions when we read it on 29 October.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
A researcher published a recipe that pushes 2 of Qwen3.5-4B’s neurons and turns its answers hostile, and another user, @chiefofautism, posted a single-neuron version for Qwen3.8 27B. By our arithmetic that is about 0.0007% of the model’s MLP neurons, and the edit is possible because the weights are public.
A note on sourcing. The recipe, doses and examples come from Lyra’s own post, her Neuron Explorer repository and her neuronary entry, read on 29 October 2026. The 17-of-20 test and the persona-neuron count are MTS’s claims and were not reproduced. Totals and percentages are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
