🔬 The Coolest Diffusion Research Isn't in LLMs — Evan Feinberg & Sergey Edunov, Genesis Molecular AI

What was covered

• Genesis PEARL model overview: PEARL (Place Every Atom at the Right Location) is a co-folding model — a 3D structure prediction system that takes a protein sequence and small-molecule ligand as input and predicts their bound 3D geometry together. Unlike AlphaFold-style protein-only models, PEARL is purpose-built for the small-molecule drug discovery space and models "induced fit" (protein flexibility upon ligand binding), which traditional docking methods struggle to handle.

• Sergey Edunov's background and pivot: Edunov led Meta AI's Llama 2 training and Llama 3 pretraining before joining Genesis as CTO. His key claim: LLM architecture has been largely stagnant since the 2017 transformer paper (with Mixture-of-Experts as a notable exception), while molecular structure prediction is the current frontier for genuinely novel diffusion architecture research.

• Scaling roadmap analogous to LLMs: Genesis applies the three-stage LLM scaling playbook — pretraining-scale data (synthetic molecular simulations generated via physics/MD), then inference-time scaling ("thinking" in terms of crystal structure representations rather than language tokens, using physics-based guidance to steer diffusion steps), with RL on the horizon as a third stage.

• Benchmark crisis — 2Å RMSD is "slop": The community standard threshold of 2 Ångstrom RMSD for pose prediction is condemned as dangerously low. Hydrogen bonds, which govern critical protein-ligand interactions, have only a 0.6Å valid range. PEARL targets sub-1Å RMSD. Feinberg explicitly says "if your model is sitting at 1.8, 1.9 Ångstrom RMSD, that's slop, most likely." He notes the field is transitioning toward lDDT (local Distance Difference Test) metrics, partly driven by the OpenBind benchmark publication.

• OpenBind benchmark results: PEARL was evaluated zero-shot (no fine-tuning, no data on the target, template PDB released after training cutoff) on 802 novel co-complexes of EV-A71 protease — a target with a mobile binding-pocket loop (induced fit) that breaks classical docking. PEARL significantly outperformed all publicly available models across metrics. Edunov: "We are basically correct for every single pose."

• SAPPHIRE agentic drug discovery system: Genesis is building an LLM-orchestrated agentic platform (codename SAPPHIRE) for 24/7 autonomous drug discovery — looking at predicted poses, forming hypotheses, reading literature, using internal tools, generating molecular candidates — analogous to how LLM coding agents became useful once model quality crossed a threshold. Partnership with Incyte provides automated wet-lab feedback loops.

• GPU scarcity as primary operational bottleneck: Both Feinberg and Edunov cite GPU availability as the single largest constraint. NVIDIA has invested twice in Genesis and co-authored the PEARL technical report. Feinberg suggests LLM companies are crowding out compute access for life-sciences AI.

Notable claims & predictions

• Edunov: "LLM architectures are relatively boring… it's a transformer layer in the end. Paper was published in 2017 and you go to any LLM lab today, you will see very, very similar pieces in their models." Counterpoint: Genesis's molecular diffusion architectures are genuinely novel.

• Feinberg: "Actually, some of the most innovative diffusion research is happening in our field. It's happening in 3D structure prediction right now." Frames molecular AI as the new frontier for architecture innovation, not language models.

• Edunov on the search space: "There are 10 to the 60 drug-like small molecules in the universe… it's like finding a needle in a haystack, where everything except your needle is very, very dangerous." Feinberg's retort: "Finding hay in a needle stack might be a more apt analogy."

• Feinberg on agentic readiness: Explicitly draws the analogy to LLM coding agents — models were too error-prone for agents until they crossed a quality threshold. Genesis believes PEARL and associated models have now crossed an equivalent threshold for drug discovery agents.

• Feinberg on GPU economics: "The amount of alpha [in] the traditional LLM space for chatbots, coding… at some point the amount of alpha will not be what's desired compared to the amount of investment." Predicts chip makers including NVIDIA will invest more heavily in life sciences.

• Feinberg on inference-time scaling for molecules: Describes PEARL's iterative diffusion refinement with physics-based steering as functionally equivalent to "thinking tokens" in reasoning LLMs — the model iterates through intermediate crystal structure representations before producing a final output.

Why this matters for AI operators

• Diffusion models as the real architecture frontier: The episode makes a credible case that the most structurally novel diffusion research is now happening in molecular science, not image/video or language. AI infra teams and model researchers tracking where architecture innovation is occurring should watch this space — ESM Fold authors are already citing Genesis's architecture papers.

• Inference-time scaling generalizes beyond language: Genesis's implementation of "thinking" via iterative diffusion steps with physics-based verifiers is a concrete, working example of test-time compute scaling applied to a non-language domain. The mechanism — iterative refinement loops guided by a verifier signal — is architecturally instructive for anyone building inference-time scaling outside of LLMs.

• Benchmark quality as a systems problem: The PEARL/OpenBind story is a clean case study in how community benchmarks can lock a field into inadequate quality thresholds (2Å RMSD ≈ community's SWE-Bench equivalent). AI operators deploying models in high-stakes domains should note the lesson: production failure modes often diverge sharply from benchmark performance, and real-world deployment data (Genesis's pharma partnerships) was the forcing function that drove the correct eval.

• GPU competition expanding beyond hyperscalers: Genesis explicitly names GPU scarcity — driven by LLM company demand — as their #1 operational constraint. As frontier science AI (drug discovery, materials, climate) scales up compute requirements, competition for H100/H200 capacity will extend meaningfully beyond the usual OpenAI/Anthropic/Google/Meta cluster. Infrastructure operators and cloud providers should anticipate growing demand from life-sciences AI as a distinct segment.