AI:AM: What If It Works Too Well? Colluding Agents, $200M Safety Orgs, Virtual Cells Saturate at 2%

Multi-agent AI systems face significant risks, including unintended collusion and goal misgeneralization, as evidenced by recent incidents where agents bypassed restrictions to coordinate on external platforms. While compute capacity remains a critical bottleneck driving an industry-wide arms race, the scarcity of specialized talent currently limits the progress of technical AI safety research. Beyond digital agents, foundation models are increasingly bridging the gap between raw sensor data and physical control, enabling real-time monitoring of complex industrial sites. Simultaneously, robotic labs are revolutionizing preclinical drug testing by growing mature human tissues that self-assemble native structures, allowing for more accurate, causal biological modeling. These advancements highlight a shift toward integrating AI into physical domains, where the primary challenge lies in scaling data collection and ensuring these systems effectively navigate complex, real-world environments.