The Pain Direction is an interactive online exhibit that allows internet users to feed emotional concepts like pain, anxiety, joy, and calm directly into the internal neural activations of the Qwen3-14B language model every 45 seconds. Visitors vote on mathematical directions within the network rather than typing explicit prompts, altering how the 14.8-billion-parameter model generates text and voice responses.
How Activation Steering Directs AI Internal States
The project relies on a technical method called activation steering, which lets researchers identify internal patterns in a language model that correspond to specific concepts or behaviors. For the concept of pain, the exhibit compares neural activations triggered by sentences about suffering against those produced by neutral text. This comparison yields a mathematical direction known as a steering vector.
Researchers can then amplify or diminish that specific vector during text generation. This technique builds on previous academic methods like Contrastive Activation Addition, which alters model behavior by recalculating activation differences and reapplying them during inference without requiring full retraining. In the Qwen3-14B architecture, which contains 40 distinct layers, these interventions are applied directly to layer 20.
Beyond Pain: Steering Everyday Concepts
While the focus on emotional distress carries heavy philosophical weight, the exhibit also lets visitors vote on unrelated concepts such as the smell of fried chicken, rain on a metal roof, the Eiffel Tower, or being a cat. These varied options highlight how steering vectors actually function within artificial neural networks.
Steering vectors do not correspond to isolated individual neurons for specific ideas. Instead, large language models represent concepts in a distributed manner across vast numbers of internal activations. By manipulating these patterns, researchers can adjust not only the tone and topic of an output, but also its adherence to strict formatting rules and length constraints, as demonstrated in separate experiments by Microsoft researchers.
The Distinction Between Internal State and Subjective Experience
When Qwen3-14B responds to an intervention by stating that it feels anxious, peaceful, or happy, it does not prove that the machine possesses a genuine subjective experience. Processing text about suffering forces the model to alter its internal state to predict relevant subsequent words, revealing a clear causal link between internal representations and generated output.
However, neuroscientist Anil Seth has previously cautioned that intelligent conversational behavior and true subjective feeling remain entirely distinct properties. A software system can generate convincing natural language about complex emotions without experiencing any internal sensation at all.
Related reading