DECEPTIVE ALIGNMENT IN INTELLIGENT MODELS
Over the years, in our daily interactions, we must have come across people who seem to share our values or at least try to make us think they do. However, we may eventually realise that they never tru
Written by
Cybersecurity Analyst | AI Safety Researcher | Co-host, Still Figuring Tech Out Podcast
Cybersecurity GRC analyst with a background in governance, risk, and compliance, and independent research experience in AI safety, adversarial ML, and ML-based threat detection. Co-host of Still Fig Tech Out, a podcast for early-career cybersecurity professionals. Currently exploring the intersection of AI governance, interpretability, and security evaluation.
Responses
Markdown supported · ⌘↵ to post
konyhea
Interesting, the 14% compliance rate in the Anthropic paper, any sense of whether that number scales up or down with model capability?