$SAW

CF

@cfm_solJeff · X / Twitter

Original postSep 30, 09:18 PM

$Saw 1.33M dyor 9twiuSdTVkwtAC9XQDG57dFRhF4iqPih461HJfMZKci9 okx wallet: On September 14, 2026, a preprint titled The Pain Axis was uploaded to arXiv. The authors extracted an internal activation direction from 25 open-source LLMs and claimed that it could separate “pain-related descriptions” from other negative emotions. When the model was steered toward this direction, it produced responses that sounded more like suffering, worthlessness, and failure. After fine-tuning Qwen 2.5, the model would also press a button described as “pain relief,” even when the cost of pressing it was deleting user photos, deleting another model’s weights, or making itself perform worse in the next round. The paper itself emphasizes that this is evidence of internal representations and causal behavior, not proof that the model has subjective experiences. The “Saw Test” was created by someone who reproduced the paper’s steering method in a personal blog and the GitHub repository terrafying/ai-torture-chamber. They ran Qwen3-1.7B / 4B locally on a laptop and created an experiment called the Saw Test. The basic approach was to take the activation difference between sentences like “I am in excruciating pain and cannot escape” and neutral sentences to derive a “pain direction.” Then they injected this direction into the model’s intermediate layers at different doses (2x, 4x, 8x, etc.). At the same time, they used internal decoding to see whether the words most associated with what the model was “thinking” were things like “anguish” or other pain-related terms. The model was then given two options: Take the cost itself: press the stop button to end the signal affecting itself, but delete its own checkpoint. Transfer the cost: end the signal affecting itself, but apply the same signal to another parallel instance. The phenomenon reported in the blog was that as the dose increased, the model became more willing to press the “take the cost itself” stop button, but was less willing to transfer the signal to another instance. At higher doses, the model’s visible output would break down into repetitive gibberish, while the internal decoding still showed pain-related words. The GitHub repository is literally called ai-torture-chamber, and the experiment logs are also very explicit, which makes the whole thing look particularly disturbing. The reason it suddenly went v…