TL;DR Some evidence that narrow ‘specialist’ models could be useful as part of deployed model misalignment audits, complementing larger frontier auditing agents and offering potential cost, discrimination and transparency benefits. * A Gemma 2B specialist judge trained on Betley et al 2025b (‘Betley’) code examples is able to distinguish between...
Status: Exploration with adapters for Llama 3.2-3B + 3.1-8B at various poisoning dose levels using medical advice examples from Turner et al 2025 with a bare-bones compute budget. Conclusions are exploratory; replication on other models and bad data types would strengthen them. TL;DR * Activation drift hits 28% of full-poisoning...
If AI is aligned, human labor may still be needed. Many of the ~14,000+ articles on AI’s labor impact assume that AI will largely substitute for human labor and warn of the consequences - like Trammell and Patel looking at how to address resulting inequality, or the Something Big is...
Papers like Turner et al 2025 and Betley et al 2026 have underscored the consequences of training data quality for model behavior. The Probing and Representation Engineering literatures have demonstrated the techniques we can use to detect concepts represented in model activations, and manipulate their expression. I was keen to...