Anastasia Wei

on the atlas — 10

• How useful is the information you get from working inside an AI company?3 savers
• Fuzzing LLMs sometimes makes them reveal their secrets — AI Alignment Forum1 savers
• Putting up Bumpers1 savers
• Anthropic repeatedly accidentally trained against the CoT, demonstrating inadequate processes1 savers
• Curius / Onboarding2507 savers
• Relationships are coevolutionary loops - by Henrik Karlsson37 savers
• A global workspace in language models \ Anthropic20 savers
• The case for ensuring that powerful AIs are controlled — LessWrong9 savers
• Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations6 savers
• Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations — LessWrong2 savers