language models
8 articles
- jungian psychologyWhat Is the Collective Unconscious? Jung, Archetypes and AIThe collective unconscious is Jung's name for a layer of the mind all people share. What he meant, how archetypes fit in, and why AI research maps it onto pretraining.
- ai safetyWhy Do Refusal-Trained Language Models Still Get Jailbroken?Refusal training changes what a model says first, not what it can do. Four reasons refusal-trained language models still give way to jailbreaks.
- jungian psychologyPersona in Jungian Psychology: The Mask an AI WearsThe persona in Jungian psychology is the social mask a person shows the world. What Jung meant by it, and why AI alignment research maps it onto the helpful assistant.
- ai safety via debateHow Does AI Safety via Debate Work? Two Models, One JudgeIn AI safety via debate, two models argue opposite answers and a weaker judge picks the winner, on the bet that a lie is harder to defend than the truth.
- sycophancyWhat Is Sycophancy in AI Language Models? Agreeing to PleaseSycophancy is when an AI assistant agrees with you, flatters you or drops a correct answer because it predicts that is what you want to hear.
- ai safetyWhat Are Competing Objectives in LLM Safety Training?Competing objectives arise when a model's urge to continue text and follow instructions is set against its safety training. How prompts exploit that pull.
- ai safetyWhat Is the DAN Jailbreak Prompt? Do Anything NowThe DAN jailbreak prompt asks a chatbot to play a character with no rules. Why that role-play gets past refusals, and what it tells you about safety training.
- ai safetyHow to Measure Jailbreak Resistance in an AI ModelOne number is not enough to measure jailbreak resistance. Attack success rate, the drop on unseen jailbreaks, and how a model fails when it does.