Show HN: Try out emotion steering of LLMs here (eigenweltlabs.com)

Anthropic has introduced a novel method for emotion steering in large language models (LLMs) using Claude Sonnet 4.5. Their E-STEER architecture allows for representation-level emotion interventions in LLMs, enabling simulations where generated outputs can express specific emotions like fear or joy based on varying steering inputs. This innovation not only enhances the emotional intelligence of AI models but also opens up new avenues for simulating human-like behavior in interactive AI systems. The significance of this advancement lies in its potential for dynamic emotion manipulation, which can enrich user interactions and responses from AI systems. By employing a streamlined architecture that maintains the performance benefits of batching through vLLM, this method allows for real-time emotion steering across multiple requests. Key technical features include the extraction of emotion vectors from model hidden states, which can be applied during inference to guide the AI's output in a nuanced manner. Users can implement this by following a structured workflow involving extraction, testing, and serving setups to create an OpenAI-compatible API for emotion-based interactions.