Approaching the AI Event Horizon? Part 2, w/ Abhi Mahajan, Helen Toner, Jeremie Harris, @8teAPi

Clinical validation remains the critical bottleneck

Despite Isomorphic Labs achieving 2x better binding affinity predictions than AlphaFold 3, 97% of oncology trials still fail because pre-clinical optimization does not reliably translate to human efficacy.

Biology lacks verifiable ground truth for RL

Unlike math and code where rewards are cheap and immediate, clinically valuable biology problems require months or years to verify outcomes, making reinforcement learning from experimental feedback impractical.

Toxicology prediction defies simple modeling

Drug toxicity manifests across vastly different timescales, species, and doses, requiring in vivo observation rather than clean in vitro cellular assays that limit the effectiveness of closed-loop AI experimentation.

Foundation models need in vivo human data

Abhi Mahajan argues the next breakthrough requires generative models trained on rich data from real human tumors rather than biologically unrealistic in vitro settings.

No consensus on automated AI impact

According to CET's report 'When AI Builds AI,' expert workshops failed to establish shared expectations about automated AI R&D timelines, marking it as a major source of strategic surprise.

Closed-loop experimentation hits biological limits

While companies bet on AI systems running their own experiments, biology's 18-month feedback loops for clinical data prevent the rapid iteration that drove progress in math and coding.

Dual deficits in control and coordination

Jeremie Harris highlights the dangerous intersection of lacking both technical methods to control superhuman AI and the trust mechanisms necessary for US-China collaboration on shared risks.