# Latent Space: 🔬 "The Most Innovative Diffusion Research Is Happening in Drug Discovery, Not Image Generation" — retained record

> Retained youtube record from youtube:Latent Space.

Canonical: https://highsignal.app/data/records/2b045fa0c0d52e42

[← sources](https://highsignal.app/data)


source: [youtube](https://highsignal.app/data/youtube) (youtube:Latent Space)  published 2026-06-30 21:41 ingested 2026-07-01 09:35

[original source ↗](https://www.youtube.com/watch?v=YQWXxnkK4dw)

## retained text



```
Evan Feinberg and Genesis CTO Sergey Edunov join us to talk about solving drug discovery with AI. Sergey, fresh off leading Llama 2 and Llama 3 pretraining at Meta, makes the case that the most interesting architecture work in AI right now isn't happening in language models — it's happening in 3D structure prediction, where diffusion turned out to be the missing primitive the field had been waiting for. Genesis's new model, PEARL (Place Every Atom at the Right Location), puts that to work: it doesn't just predict where a ligand binds, it models how the protein itself flexes to accommodate it. We get into why that was so hard to do until now, and why Evan thinks the field's favorite benchmark — 2Å RMSD — is mostly "slop." (Full technical report here: https://arxiv.org/abs/2510.24670)

We also dig into Genesis's agentic drug discovery system, SAPPHIRE, and what it actually takes for an AI agent to act like a chemist: reasoning about poses, forming hypotheses, reading literature, and proposing the next round of candidates. Plus: why finding a good drug is less "needle in a haystack" and more "hay in a needle stack," the tension between binding affinity and solubility, and how PEARL performed zero-shot on the brand-new OpenBind benchmark (https://www.genesis.ml/news/zero-shot-pearl-system-surpasses-all-cofolding-models-on-openbind) against a notoriously hard induced-fit target.

Links:
Evan Feinberg: https://www.linkedin.com/in/evanfeinberg/
Sergey Edunov: https://www.linkedin.com/in/edunov/
Genesis Molecular AI: https://www.genesis.ml/ | https://www.linkedin.com/company/genesis-molecular-ai
PEARL announcement: https://www.genesis.ml/news/introducing-pearl
PEARL technical report: https://arxiv.org/abs/2510.24670
OpenBind benchmark results: https://www.genesis.ml/news/zero-shot-pearl-system-surpasses-all-cofolding-models-on-openbind

Chapters:
00:00 – Hook: Diffusion for Drug Discovery
00:49 – Intro: Genesis Molecular AI
03:34 – Why Drug Discovery Is So Hard
08:11 – Drug Discovery 101
11:35 – PEARL: Genesis’ Structure Prediction Model
15:19 – Synthetic Data, Physics, and Scaling
17:21 – Inference-Time Scaling for Molecules
20:27 – Physical Priors and Model Usability
25:06 – Where AI Fits in the Drug Pipeline
32:19 – First-in-Class vs. Best-in-Class Drugs
35:33 – Why AlphaFold Didn’t Solve Drug Discovery
37:15 – ADMET, Toxicity, and Drug Design
38:31 – Pharma Partnerships and Incyte
41:39 – Why One-Ångström Accuracy Matters
46:39 – Agents for 24/7 Drug Discovery
52:10 – How Genesis Reached Sub-Ångström Accuracy
57:49 – The Eval Crisis in Molecular AI
1:04:22 – Wet Lab Data and Generalization
1:10:44 – Lab Feedback Loops and RL
1:15:46 – Why Automated Labs Are Hard
1:22:56 – Genesis’ AI Company Pivot
1:29:46 – Why Agents Now?
1:37:35 – PEARL’s OpenBind Results
1:42:16 – The GPU Bottleneck
1:45:38 – Call to Action for AI Researchers
1:48:08 – Closing
```



document fetched 2026-07-07 09:28 · canonical https://youtube.com/watch?v=YQWXxnkK4dw
