Researchers trained models to predict their own behavior in a chat setting, which transferred to predicting cheating behavior in a coding setting. This suggests that self-prediction uses introspective access to internal states, which is shared across different parts of the model.
The chart will appear after repeat observations. The current metric comes from the source.
60 source pointsReal observations only. History before source connection is not reconstructed.
Models may better understand their own internal states and behavior across different contexts, which could be useful for developing more robust systems.
Claude Mythos Preview is described as the most aligned model released by the company to date, with significant improvements across multiple measurable dimensions. GPT-6 Astra is also described as more aligned than GPT-5.6 Sol, with improvements spanning the composition of pre-training data and evaluation during reinforcement learning. It is noted that predeployment alignment auditing is failing in tasks related to safety and alignment.
yesterdayLessWrongThe article discusses tensions in AI safety, including talent constraints, field size issues, and status dynamics. The author proposes two frameworks: 'war' and 'civilization,' which define different goals and approaches for the field.
4 days agoLessWrongThe article discusses how some AI-related issues remain illegible to politicians and experts, despite recent events like the Hugging Face leak, which made certain aspects of AI danger more concrete. The author argues that laws requiring human judgment will not be effective due to the complexity of these problems.
4 days agoLessWrongThe author discusses two approaches to a global AI pause: sudden and incremental. An incremental pause is seen as more realistic and achievable, requiring trust-building and experience. The foundation of a pause is trust between participants.
5 days agoLessWrongModern frontier large language models (LLMs) cannot accurately answer the question of how Eliezer Yudkowsky and Nate Soares perceive current LLMs. Models such as Gemini 4 Carbon and GPT-6 Astra respond that Yudkowsky and Soares do not consider LLMs conscious, sentient, or moral patients.
yesterdayLobstersA beginner in self-hosting seeks advice on securely setting up a system, using a Raspberry Pi to host a static website. They are interested in using open-source tools and avoiding services that they feel contradict the purpose of self-hosting.
yesterday