Bryce Cai - The State and the Science of AI-Bio Evals [Alignment Workshop]

Bryce Cai demonstrates how AI models have rapidly achieved and exceeded human expert performance across multiple biology benchmarks in just two years. His research shows GPT-4o matching median virologist performance on wet lab troubleshooting through the VCT eval, while other models successfully write executable scripts for biological tasks and effectively use tools like AlphaFold. The team has validated AI-generated protocols in actual laboratory settings, confirming these capabilities translate to real-world applications. This advancement demonstrates the double-edged sword of accelerating scientific capabilities, which also potentially democratize dangerous biological expertise. Without proper mitigations, these capabilities could enable novice uplift in biosecurity-relevant domains, potentially allowing bad actors to create novel pandemic threats without requiring traditional PhD-level training. Note: The opinions shared in this event are those of the speaker(s) and may not represent the views of FAR.AI or their affiliated organizations.