What is Hype And What is Real?

Author: Gregg Spivey (moderator), AI/ML Partnerships, Eli Lilly; Hani Goodarzi, Core Investigator, Arc Institute; Jacob Kimmel, Co-Founder & President, NewLimit; Reshma Shetty, Co-Founder & COO, Ginkgo Bioworks Collection: UNLOCK-2026-Sessions

Summary

This panel at the UNLOCK 2026 conference — "AI Science: Cutting Through the Hype" — brings together practitioners from Eli Lilly, Arc Institute, NewLimit, and Ginkgo Bioworks to distinguish genuine AI progress in drug discovery from overhyped claims. Moderator Gregg Spivey opens with the Allbirds trainers-to-data-center example: a shoe company pivoted to GPU provisioning and saw its stock price rise 600% in a single day, illustrating how AI hype can detach from fundamentals. He frames the central question as which elements of AI for drug discovery are real, validated contributions versus narratives that exceed the evidence.

The panelists converge on a few genuine areas of progress. Antibody generation is identified as the most mature application: the lab-in-the-loop cycle from algorithm prediction to wet-lab validation is tractable in this modality because training data is abundant and experimental turnaround is fast. Jacob Kimmel (NewLimit) argues that the deepest bottleneck in medicine discovery over the past several decades has been target identification, not molecule optimization — and that emerging AI systems capable of simulating cell-state responses to genetic perturbations are beginning to address this root cause. Hani Goodarzi (Arc Institute) notes that the EVO and STEAD models out of his lab represent real progress, and defends sustained hype around transformative technologies by pointing to the Human Genome Project: a $3 billion effort projected to cure all disease that nonetheless catalyzed the genomics era.

The most contested hype claim is "reasoning from literature." Jacob Kimmel explicitly calls out the hypothesis that reasoning agents can mine existing PDFs and infer new therapeutics — his view is that in the vast majority of cases the observations needed to reason toward a new therapy do not yet exist; more experiments are required, not just more reasoning. A second contested area is lab-in-the-loop itself: Hani cautions that while active learning is intellectually compelling, the assay cycle times in biology are far too slow to allow the compounding that makes the approach powerful in other domains. The panel closes by agreeing that AI needs to surpass — not just match — the predictive accuracy of existing pre-clinical models (which currently explain why 90% of drugs fail) to constitute genuine progress.