<?xml version="1.0" encoding="UTF-8"?><oembed><type>video</type><version>1.0</version><html>&lt;iframe src=&quot;https://www.loom.com/embed/4927978822dc47b58450bdddaf6763c2&quot; frameborder=&quot;0&quot; width=&quot;1920&quot; height=&quot;1440&quot; webkitallowfullscreen mozallowfullscreen allowfullscreen&gt;&lt;/iframe&gt;</html><height>1440</height><width>1920</width><provider_name>Loom</provider_name><provider_url>https://www.loom.com</provider_url><thumbnail_height>1440</thumbnail_height><thumbnail_width>1920</thumbnail_width><thumbnail_url>https://cdn.loom.com/sessions/thumbnails/4927978822dc47b58450bdddaf6763c2-5d6b75c1fd2c78c3.gif</thumbnail_url><duration>235.584</duration><title>Steering a Conspiracy Theory Neuron</title><description>This Loom demonstrates a neuron steering experiment using a 32 billion parameter model to see how adjusting a specific “conspiracy theory” neuron changes outputs. The author runs prompts token by token, turns steering on halfway, and targets layer 42 neuron 5277 using a clamp that sets the neuron’s output to a specific amount. With steering strongly activated, the banana bread and other generations shift toward more conspiracy content, and gradually reducing the clamp changes the results but may not fully remove conspiracy-related phrasing. The video repeats similar trials with anti vaxxer style prompts to compare clamped versus toned down steering behavior.</description></oembed>