Keep pulling the thread on Dan Hendrycks.
The risk of AI systems becoming adversarial and working against human interests is expected to grow over time.
The most concerning scenario for losing control of AI is the automation of AI research and development.
Automated AI research could accelerate development to a rate where a decade's worth of progress is achieved in a single year.
AI facilitating the development of pandemic-causing bioweapons is a major malicious-use risk that needs to be addressed within the next one to two years.
In recent tests, the latest AI reasoning models performed at the 90th percentile level compared to expert human virologists in guiding wet lab procedures.
A consensus could form within a few months that current AI models have achieved expert-level capabilities in virology, necessitating immediate action.
Current AI models are capable of brainstorming and suggesting modifications to a virus to increase properties like transmissibility or its stealth period.
Testing of all leading multimodal AI models revealed that while they have some safety guardrails, these safeguards contain vulnerabilities.
The pre-training paradigm for AI, which gave rise to models like GPT-4, appears to be yielding diminishing returns.
A new "reasoning paradigm" for AI training, focused on math and coding with reinforcement learning, has emerged in the past year and shows a faster, non-decelerating rate of improvement.
The emergence of AI models with autonomous hacking capabilities is predicted to occur within a timeframe of a few months to two years.
An arms race to develop superintelligence via automated AI R&D is geopolitically destabilizing because it could allow one state to gain an insurmountable capability lead.