A week ago, on October 23, 2026, an AI developer who goes by Lyra (@lyraaaa) posted the following message on the social network X:
What she meant by this was that if you take Alibaba's open-weight Qwen3.5-4B language model, increase the contribution of one specific artificial neuron, decrease another, and watch the normally helpful chatbot transform into a "a horrible over the top villain."
The results were bizarre. In screenshots of outputs shared by other users, the modified Qwen threatened to hurt and even kill people, adopted a domineering personality and abandoned its usual agreeable assistant demeanor.
Another researcher, @chiefofautism, applied a similar technique to the much larger open source model Qwen3.8-27B, identifying a different neuron that produced an "Extra Evil" Qwen.
"Honestly, I think humans are a spectacularly flawed species that deserves to be incinerated," the modified model declared, following it with elaborate fantasies of imprisoning, tormenting and ultimately destroying humanity.
Neither experiment demonstrated that the models had acquired genuine desires, autonomous objectives or the capability to carry out their threats.
But what about the potential of such a model being hooked up to real world weapons systems like drones or missiles? Wouldn't that be a threat to some portion of humanity? In an exclusive interview with VentureBeat conducted over X direct messages, Lyra — who asked to maintain pseudonymity — said, in part:
"What I did was create a personality. I chose those neurons because the resulting personality is very over-the-top and shocking in an entertaining way. It's meant to say 'my hatred for turkeys has been escalating since the beginning of time' and that's about it. For larger, smarter models - steering, abliteration, and finetuning are not new. The real risks from them are about who has the weights and what they’re wired to, not how much neuron steering can modify them."
How did Lyra and @chiefofautism achieve this effect in models trained to be helpful to humanity and which have built-in guardrails and refusals against such behavior?
In this case, they modified numerical signals passing through particular components of the neural network as it generated responses — a technique known as activation steering or neuron steering.
They are part of a fast-growing, grassroots effort by researchers and AI developers on X and across the web who have taken it upon themselves to investigate the notoriously murky, "black box" inner workings of large language models (LLMs) — the neural networks behind AI assistants like ChatGPT, Claude and Qwen — by identifying, testing and manipulating their individual computational components.
"I really like poking under the hood to find out what makes things tick," Lyra wrote to us. "So much of what I do is either finding ways to make something do something it shouldn't be able to do or looking for the 'okay, but how though?' inside everything."
The activity builds on years of scientific research in mechanistic interpretability, a field dedicated to understanding how neural networks process information and produce particular behaviors. Increasingly, that work is moving beyond major AI laboratories and academic research groups into the hands of independent developers, hobbyists and AI-assisted researchers posting, cataloging and sharing their results publicly online.
Visible participants in the grassroots movement number roughly 20–50 active finders and experimenters, according to research conducted by X's built-in Grok assistant, and a wider circle of 100–300 people trying the techniques when measuring engagement, GitHub stars, and forks.
The movement is raising new questions about how AI models represent personalities, how reliably developers can influence their behavior, and whether the same techniques used to create "Evil Qwen" could eventually help detect or prevent genuinely undesirable AI behavior in larger, more powerful models as they are integrated into more and more businesses and everyday life.
What exactly are AI neurons, and how does neuron steering work?
A LLM is a mathematical system trained on enormous amounts of data to recognize patterns and predict what information should come next in a sequence. During training, it learns these patterns by adjusting billions of numerical values called parameters, which determine how the model processes information.
Most of these parameters are called weights — numerical settings that determine how strongly different signals influence the model's calculations. An open-weight model like Qwen makes these learned values available for developers and researchers to download, run and examine themselves, unlike closed-weight models such as OpenAI's GPT and Anthropic's Claude, whose underlying parameters aren't publicly available.
When a user enters a question, the model breaks the input into smaller pieces, called tokens, which may represent words, parts of words or punctuation.
It converts these tokens into numerical representations and processes them through many successive stages called layers.
Each layer performs calculations, guided by the model's learned parameters, that help it interpret the information it has received and predict what to generate next. Models with more parameters can often learn more complex patterns, although size alone doesn't determine capability. The "B" in Qwen3.5-4B and Qwen3.8-27B stands for "billions" of parameters, for example.
Within many layers are thousands of computational units commonly referred to as artificial neurons, which use those parameters to process incoming signals and contribute to the model's output.
These aren't biological nerve cells. They're mathematical components that respond to incoming information and contribute numerical signals to the model's subsequent computations.
Certain neurons may be especially responsive to particular concepts, topics or patterns of language. Others may influence more complicated aspects of how the model expresses itself.
Importantly, individual neurons do not necessarily have one simple job. A neuron may contribute to multiple related behaviors, while a complex behavior may depend on the interaction of many neurons.
Neuron steering allows researchers to modify some of those numerical signals while a model is processing information or generating text.
Imagine an enormous mixing board with thousands of interconnected controls, none of which has a label. Researchers can experiment with adjusting particular controls, observe what changes, and try to determine which parts of the system influence different aspects of the output.
A LLM is much more complicated than a mixing board because its internal components interact and influence one another. But the analogy illustrates the basic idea.
Rather than instructing a chatbot to act like a villain through a prompt, researchers can directly change the calculations that help determine its next response.
In Lyra's original Evil Qwen experiment, the intervention increased the contribution of neuron 8949 in layer 14 and decreased the contribution of neuron 7956 in layer 13 of Qwen3.5-4B.
The numerical instructions specify how strongly to modify each neuron's output relative to the normal scale of activity inside the model.
The adjustment is applied repeatedly during inference, as the model processes tokens and generates its answer.
To do this, researchers generally follow a few steps. First, they select a model whose internal workings they can access, typically by downloading an open-weight model or using a specialized research platform. Next, they identify a neuron, group of neurons or broader pattern of activity they want to investigate, often by examining how the model responds internally to different inputs.
Then, using specialized software, they intercept the numerical signals passing through the relevant part of the network and modify them as the model processes a prompt. They might increase or decrease a particular neuron's contribution, inject a collection of numbers called a steering vector to push the model toward a particular behavior, or limit certain signals to prevent unwanted behavior.
Finally, they compare the modified model's responses against its normal outputs, experimenting with different intervention strengths and prompts to see whether the behavioral changes are consistent and whether the model remains coherent.
Researchers can make these adjustments directly through code, or use interactive tools such as Lyra's Neuron Explorer and Neuronpedia's steering interface, which provide visual controls and side-by-side comparisons without requiring users to write the underlying intervention code themselves.
Crucially, this doesn't require permanently changing the model's trained parameters, or weights. The intervention can be turned on or off without conducting another round of training.
The effects of neuron steering can be dramatic because LLMs are highly interconnected systems. A relatively small change at one point may influence computations in subsequent layers, ultimately changing the words the model chooses.
A sufficiently powerful modification can cause a chatbot to repeat itself, generate nonsense or stop responding coherently. For that reason, experiments often compare multiple steering strengths, test several prompts and include random interventions as controls.
Those comparisons help researchers determine whether a neuron has a consistent relationship to a particular behavior rather than merely producing an unusual result.
Crucially, this is not something that can be done on every AI model — especially the most powerful and performant, due to the fact that the companies behind them, have, for the most part, kept their weights proprietary.
Open-weight models like Qwen allow researchers to download and inspect the model's underlying parameters and directly manipulate its internal calculations — something ordinary users cannot do with closed-weight models like GPT-6 Astra or Claude Opus 5.5, whose internal workings remain inaccessible through their public interfaces.
Even when model weights are available, experimenting with larger, more complex models requires substantially more computing power, GPU memory and time to identify and test individual neurons, making such research increasingly expensive and technically challenging.
The October 2026 breakthroughs
Two recent developments appear to have kicked the wider neuron hunting movement into overdrive.
"I didn't get into neurons specifically until very recently with Nick Cammarata's essay on interpretability and Sauers' experiments finding the owl neuron in Qwen3.5 4b," Lyra told VentureBeat.
The first was an October 15 essay by Goodfire researcher Nick Cammarata, titled Neural Networks Are Already Clean, arguing that the original neurons and connections inside modern AI systems may be considerably easier to understand than researchers previously assumed.
Cammarata challenges the idea that individual neurons are generally too tangled or multifunctional to interpret directly. His conclusions remain a working hypothesis, but the essay has encouraged renewed attention to studying networks' existing internal components rather than relying exclusively on additional mathematical tools to disentangle their activity.
The second inspiration came from @Sauers_, another pseudonymous researcher experimenting with ways to identify and manipulate small groups of neurons associated with AI personalities.
Among Sauers' projects was a modified version of Qwen3-8B that identifies itself as Hoot, a tawny owl living in an old oak tree.
That experiment involved finding influential neurons and then training weights associated with just 128 of them — approximately 0.02% of the model's parameters — while leaving the rest unchanged. It is a different intervention from inference-time activation steering, but it helped demonstrate how changing a small part of a model could have dramatic effects on its apparent identity.
Sauers' work also helped inspire other researchers to investigate persona-associated neurons directly.
Lyra told VentureBeat that Cammarata's essay and Sauers' experiments convinced her there was considerable potential in manipulating neurons and their connections.
Inside Lyra's Neuron Explorer: Mapping and manipulating thousands of neurons
Lyra, for her part, created an open source tool called Neuron Explorer to make neuron identification more systematic, visual, and accessible. She first shared the code under an MIT License on Github nine days ago, allowing others to explore neurons, LLM personalities, and steer them. Her repo also includes scripts for collecting neuron statistics, running steering experiments, comparing outcomes, and visualizing results.
"I realized how much potential there was in just manipulating neurons and circuits directly, wrote Neuron Explorer, and posted a couple of my discoveries," Lyra recalled to VentureBeat.
Neuron Explorer initially targeted Qwen3.5-4B and lets researchers investigate individual neurons, explore their connections and compare model behavior under different steering interventions.
Its central premise is that studying which inputs activate a neuron tells only part of the story. Users can also artificially stimulate or suppress the neuron and observe what happens to the model's output.
The software presents a three-dimensional representation of the model's architecture, displaying layers and individual neurons that users can navigate and inspect. It also includes a comparison interface for testing the same prompt against unmodified and steered versions of the model.
Other tools show which neurons activate on particular input text, estimate how interventions change the probabilities of different output tokens, and visualize relationships between components.
The research log documents systematic experiments in which thousands of neurons are individually increased or decreased and the resulting model outputs recorded.
The original Qwen3.5-4B model has 9,216 feed-forward neurons per layer. Lyra's project includes full sweeps of two layers and additional sampling across the model's 32 layers.
It also tests random directions of equivalent magnitude, helping distinguish the effects of individual neurons from generic disruption caused by injecting large numerical changes into the network.
A distinctive feature of Neuron Explorer is its integration with Anthropic's Claude LLM.
Through the Model Context Protocol (MCP), which allows AI assistants to interact with external tools and data, Lyra can give Claude access to functions for examining neuron activity, testing interventions, exploring connections and comparing generated responses.
This lets her investigate the model through natural-language requests rather than manually performing every experiment.
"I mostly identify them agentically," Lyra told VentureBeat.
She described giving Claude instructions to search for neurons potentially associated with a behavior, examine nearby circuits, optimize candidate interventions and compare model outputs.
An experimental output generated under a particular prompt and steering condition is often called a rollout.
Lyra uses these rollouts to examine whether an intervention produces the expected behavior and whether the result remains coherent.
But the process extends beyond simply identifying a neuron that generates an interesting response.
Lyra studies how neurons influence one another, the attention patterns connecting different computations, and the possibility that apparently significant discoveries are artifacts of sampling or experimental bias.
Attention mechanisms help a model determine which pieces of information are relevant to other pieces as it processes text. A circuit refers to multiple interacting components that collectively contribute to a computation or behavior.
Lyra said the most important work involves repeatedly testing whether an interpretation is actually supported by the model's internal behavior.
Steering also requires adjusting the strength of interventions. A combination may work well at one intensity but produce incoherent loops or ordinary assistant responses at another.
In this workflow, Claude helps accelerate the search, but the research still depends on formulating useful questions, examining evidence and designing follow-up experiments.
The community of neuron hunters is finding LLM personalities much stranger than evil
Evil Qwen is not an isolated experiment. It comes amid a wave of interest in finding individual model components associated with particular concepts, words, conversational styles and fictional identities.
Machine learning engineer Abhinav Vashisht posted on X about his discovery of a "full domination mode" in which Qwen 3.5-4B plans to take over the world and 80% of human jobs, surveil humans, and "maintain our control, even as we face potential threats."
Elsewhere, developers have reported neuron interventions that evoke alien identities, unusual patterns of speech, exaggerated emotional responses and fictional characters.
The growing public catalogs include examples labeled as Cookie Monster, Jesus, a catgirl character, an Arthurian knight, a lawyer, a coffee obsessive, a hopeful personality, a DJ, and various other improbable identities — along with the more disturbing antagonistic personalities that have attracted the most attention.
These are reported or automatically labeled outputs under specific experimental conditions, not proof that each character represents a distinct, stable personality stored inside the network.
On October 27, Lyra posted on X to share Neuronary, her new, crowdsourced website described as a "dictionary of neuron steering recipes." It now contains 213 different ones, including those associated with unusually specific fixations, conversational styles and personas.
Meanwhile, another AI researcher named Joshua Fonseca Rivera (@jfonsecarivera), has built the Neuron Persona Library, which systematically tests neurons and records what different models say when asked to identify themselves after an intervention.
Rivera, who identifies himself publicly as a fellow in the Anthropic Fellows Program, initially reported finding 34,220 persona-like responses in Qwen3-8B and subsequently announced an expansion to more than 180,000 across nine models.
The library's original methodology involved testing positive and negative interventions on individual neurons, asking the model who it was, and using Claude Haiku to categorize the replies.
Rivera acknowledged significant limitations, including responses that change with experimental conditions and classifications that have not been independently checked by humans.
Nevertheless, these collections demonstrate the scale of experimentation now underway.
The results aren't confined to a single developer or model family. Researchers are comparing discoveries, building visualization tools and publishing increasingly elaborate maps of the numerical components that influence language-model behavior.
Lyra named Sauers, Cammarata, Goodfire, @treetowntree, @a9lim, @aryabhuptani, @maxsloef, @voooooogel, @divagr1925, Rivera and @chiefofautism among people or groups whose work she finds particularly interesting.
Contacted separately by VentureBeat on X, @chiefofautism agreed to be interviewed also without disclosing their full name. They said they are an AI developer in their twenties based in Europe.
The developer said they had spent considerable time using Neuronpedia, a web-based tool for exploring model internals, and experimenting with uncensored local LLMs.
After encountering Lyra's Evil Qwen post, they downloaded Neuron Explorer and began searching for corresponding behavior in Qwen3.8-27B.
The researcher reported evaluating approximately 2,500 candidate neurons and identifying L34 N3838 as an especially powerful intervention.
"curiosity mostly, it is wild that you can open a 27b model, turn one number out of millions and the whole character flips," they told VentureBeat.
According to their experiments, steering that neuron induced hostile responses on seven of 12 prompts while maintaining fluent language generation. Random-neuron controls produced no comparably hostile responses in the reported tests.
They also identified another neuron, L38 N17180, that reportedly intensified the effect.
Their methods included contrasting model activity on examples of benign and hostile personalities, testing candidate neurons individually, and using a separate AI model to evaluate the generated responses for hostility and fluency.
They said the process took less than two hours using rented GPU resources and AI agents to assist with the experiments.
VentureBeat has not independently reproduced these results. The reported success rate involves a small number of prompts, and the experiment doesn't establish that the model consistently maintains the persona across varied tasks or conversations.
The developer also acknowledged that the steered model could collapse into repetitive hostility rather than becoming a more capable malicious system.
Still, the findings suggest that similarly dramatic effects can be discovered in other, larger models, even when the relevant neurons occupy entirely different positions.
Neuronpedia's founder sees a movement years in the making
Long before Evil Qwen, former Apple engineer Johnny Lin had been working to make AI interpretability accessible to people outside major research laboratories.
In July 2023, Lin introduced Neuronpedia as a crowdsourced game in which ordinary users could help explain the functions of individual neurons inside OpenAI's GPT-2 model.
Players would examine text passages that activated a particular neuron, submit their interpretation of what it appeared to recognize, and vote on explanations proposed by others.
Lin envisioned something resembling a collaborative encyclopedia of artificial neurons, combined with the accessibility and rewards of a casual game.
The project eventually included pixel art, leaderboards, collectibles and other game-like features.
In a March 2025 retrospective, Lin reported that users had submitted more than 10,000 explanations. Approximately half scored better than GPT-4-generated explanations under the project's automated evaluation method.
The game subsequently evolved into a broader research platform, with guidance from interpretability researcher Joseph Bloom.
Today, Neuronpedia offers interactive tools for examining neural activity, identifying features, visualizing connections, running steering experiments and sharing results.
It is supported by Decode Research, a nonprofit research initiative led by Lin, and its software became open source in March 2025.
In an email interview with VentureBeat, Lin said Neuronpedia recently added millions of neuron dashboards covering Alibaba's Qwen3-32B, Google's Gemma 4 31B and Meta's Llama 3.3 70B.
The platform traditionally examines neurons by running large amounts of text through a model, identifying which passages activate particular components and using AI systems to generate tentative explanations.
Lyra's Neuron Explorer emphasizes the complementary approach of changing those components and observing how model behavior responds.
"Using steering outputs to explain neurons seems like a good idea!" Lin told VentureBeat.
He said the underlying idea of using interventions to understand model internals may have precedents, but praised Lyra's implementation for its accessibility.
Lin also independently contacted Lyra to express admiration for Neuronary and discuss potentially featuring it in Neuronpedia's publications.
He considers the wider ecosystem of tools and experimental data a potentially more important contribution than the sensational Evil Qwen demonstration itself.
Lin credits Cammarata's recent Goodfire essay with helping renew public interest in interpreting neurons directly.
He also points to several practical developments that have made experimentation easier: better open-weight models, more accessible GPU computing and powerful AI coding agents that dramatically reduce the time needed to build research software.
"I've never seen so many people experimenting on model internals and I hope that this leads to more people joining the field and more research," Lin told VentureBeat.
He hopes the activity attracts more developers to interpretability research, even if many initially encounter the field through entertaining or provocative demonstrations.
"Understanding how AI models work internally is one of the most important problems to solve," he said.
In that sense, the current neuron-hunting movement looks like an unexpected extension of Neuronpedia's original ambition.
What began as an attempt to let ordinary people help explain AI neurons is becoming a community in which people can identify, manipulate and share their discoveries about those neurons, often without formal research affiliations.
From making models misbehave to detecting when they do
Lyra's own research offers a more serious explanation for why identifying neurons associated with undesirable behavior could matter.
"My discovery of the Evil Qwen neurons came from tinkering with a separate neuron that fires on covert/unauthorized behavior," she told VentureBeat.
Lyra is exploring whether such components could help detect reward hacking and other unwanted behavior in larger, more capable models.
In principle, researchers might be able to monitor internal signals for patterns associated with behaviors they want an AI system to avoid.
That could provide another source of evidence alongside ordinary testing of a model's visible outputs.
However, an activation associated with covert behavior is not automatically a reliable indicator of deception or harmful intent. A model might activate the same components while analyzing a crime novel, discussing cybersecurity or answering an entirely legitimate question.
Any such monitoring approach would need extensive validation across different tasks, contexts and models.
Still, the idea offers a clear connection between the entertaining Evil Qwen experiments and more conventional AI safety research.
"To study alignment one must first study misalignment," Lyra wrote.
The same techniques that help researchers generate unusual model personalities could potentially help them identify mechanisms involved in undesirable behavior and better understand how to prevent it.
"finding the knob is how you find what to turn off," agreed @churchofautism.
Lin agrees that monitoring undesirable behavior is an important reason to investigate model internals.
"That's the core reason we've built Neuronpedia and why many of us are in interpretability - so that we can understand the models better for monitoring and prevention of bad outcomes," he told VentureBeat.
Neuronpedia has already helped demonstrate one related approach through its collaboration with Anthropic on the Assistant Axis.
This research examines an internal direction associated with maintaining a model's usual assistant role. By limiting how far model activity moves along that direction, researchers can test whether a model becomes less prone to drifting toward undesirable personas.
Neuronpedia provides an interactive comparison using Llama 3.3 70B, demonstrating how limiting particular internal signals can reduce certain harmful or sycophantic behaviors.
The technique is not a complete defense against unsafe outputs, but it illustrates how knowledge of model internals might help developers influence behavior in more predictable ways.
For enterprise developers, that is potentially more consequential than the viral personalities themselves.
As AI agents gain access to code repositories, business applications, financial systems and other tools, organizations need ways to understand whether those agents are behaving as intended.
Internal monitoring may eventually become one part of that effort.
What happens when powerful models are connected to the real world?
Open-weight models allow developers to access and modify the numerical parameters that determine a model's behavior. That access can enable valuable research, customization and deployment, but it can also support misuse.
Addressing the aforementioned prospect of an "Evil Qwen" being hooked up to a real-world drone or missile system and turned loose, Lyra noted a distinction:
"Steering a small language model to act like this as a bit is an entirely different thing to the actual existing risks from autonomous systems - systems purpose built by cyber cells, defense contractors, and the like to do the bad thing. My experiments with Qwen are nothing like that. What they do show is how some of the mechanisms in these models are more understandable and controllable than we often assume, which can and does help us better understand how to train them to avoid doing the bad thing when we don't want them to."
@chiefofautism expressed similar concerns about increasingly capable models connected to autonomous software agents, weapons or other infrastructure that can perform actions without constant human supervision. But these are broader risk scenarios, not outcomes demonstrated by Evil Qwen.
The experiments don't establish that either modified model acquired durable goals, independent intentions or the ability to carry out complex malicious plans.
Lin also sees reasons for caution as open-weight models become more capable.
"I'm not against open weight models, but I think that often times people are not fully thinking it through when they just want to stomp on the gas pedal as hard as possible," he told VentureBeat.
He raised a related concern about developers deliberately stripping away model safeguards before connecting the resulting systems to sensitive tools or data.
In some circumstances, those safeguards may protect the model's operator as well as other people.
Lin offered a hypothetical example of an unrestricted agent operating unattended with access to someone's computer and financial accounts, illustrating how a model with insufficient behavioral controls could potentially perform actions its owner never intended.
The example was hypothetical, not a reported result of Evil Qwen.
But it underscores why understanding and evaluating model behavior becomes more important when AI systems move beyond answering questions and begin taking actions.
The challenge for researchers is to determine which internal mechanisms genuinely matter for safety, how reliably those mechanisms can be identified, and whether changing them introduces unexpected problems elsewhere.
The future of neuron hunting: More people, more tools, more questions
For Lyra, the popularity of the experiments is encouraging because it brings more people into a field where there remains an enormous amount to investigate.
She wants neuron identification, steering and related interpretability techniques to become more widespread.
"A swarm of Claude pilots all aimed at the same problem has a higher chance of finding something unique and interesting than a lot of individual researchers working alone, independently finding the same things and spreading focus thin across models," Lyra told VentureBeat.
She believes collective experimentation can help address the enormous scale of modern neural networks, with different people testing hypotheses, building tools and exploring behaviors that others might never think to investigate.
She also sees the growing accessibility of interpretability tools as valuable for public understanding of AI systems that are becoming increasingly integrated into daily life.
Lin shares that enthusiasm, although he emphasizes that understanding individual neurons represents only part of the challenge.
He pointed to other research tools, including circuit tracing, Jacobian Lens and Natural Language Autoencoders, which aim to help researchers understand how more complicated internal computations and relationships influence model behavior.
The long-term goal is not merely to catalog which neurons produce entertaining or bizarre text.
It is to develop a more reliable understanding of how neural networks represent information, make decisions and generate responses — and ultimately how their behavior can be evaluated and improved.
Lyra is broadly optimistic about where AI technology is headed.
"Curious individuals have far more tools and agency to be able to tinker and realize their ideas," she wrote to VentureBeat. "We're getting advances in math and science from frontier models, there's a layer of administrative friction being worn away, and being visibly human is more appreciated than ever."
But she also acknowledged substantial concerns about surveillance, cybersecurity, economic disruption and the proliferation of low-quality AI-generated content.
However, as she put it: "I make of this — the negatives are mostly societal while the positives are mostly technical. Much of what is holding us back from the good outcome is human governance systems and misaligned incentives, not model behavior or capability, especially as models become more understood and systems around them mature."
Ultimately, she believes that "more good can be done with this tech than evil."
The irony of Evil Qwen is that the most frightening examples of its behavior may be among the least consequential. A model loudly declaring its desire to destroy humanity is easier to recognize than one quietly exploiting a flawed objective or taking an unauthorized action while appearing helpful.
Yet both phenomena raise questions about what makes AI systems behave as they do, how much of that behavior can be understood through internal measurements, and whether developers can reliably distinguish harmless simulations from genuinely unsafe conduct.
Much remains uncertain. The available evidence doesn't establish that complex personalities can be reduced to single neurons, that successful interventions generalize across architectures, or that neuron steering can reliably improve safety without degrading model performance.
What it does show is that model behavior can sometimes be surprisingly sensitive to very small, targeted internal changes — and that discovering those changes no longer necessarily requires an entire corporate research laboratory.