Anthropic has released new research evaluating 'feature steering,' a method aimed at modifying the behavior of large language models (LLMs) to mitigate social biases without degrading performance. The study builds on previous findings, particularly from the May release of Golden Gate Claude, which exhibited a bias towards the Golden Gate Bridge due to feature steering. The latest research indicates that adjusting specific parameters can alter political bias without the need for extensive fine-tuning. Researchers identified a 'sweet spot' for maintaining model capabilities while achieving bias reduction. However, results have been mixed, with some targeted and off-target effects observed. The study emphasizes the importance of balancing safety and controllability in AI models, raising questions about how different human perspectives might influence this balance. Notably, the anticipated 'smartness' feature for Claude 3.5 was not included in the latest findings, prompting speculation about its development.