Latent Space

Genesis: AI drug discovery is a science of resolution

Evan Feinberg and Sergey Edunov· CEO and CTO of Genesis Molecular AI at Genesis Molecular AI
·~109 min·English·Latent Space
AI InfrastructureTrainingAgentsMultimodal
TL;DR

Genesis Molecular AI's Evan Feinberg and Sergey Edunov explain how they ported the LLM scaling playbook to 3D molecular structure — synthetic physics data, inference-time 'thinking' in crystal structures, and agents — betting that sub-angstrom resolution is the threshold that turns AI drug discovery from pattern-matching into real medicines.

01Core Mental Model

The Right Primitive

Diffusion, not GANs, was <strong>the right primitive the whole field was waiting for</strong> — you don't force a breakthrough, you wait for the tool that finally works.

we sort of had to wait for the right primitive to get created and that turned out to be diffusion which turned out to be a much more useful primitive for the space.

— Evan Feinberg, Latent Space
Key Insight
The lesson is not 'diffusion is magic' — it is that an entire subfield stalled for years waiting for a generative primitive that could actually be trained. Feinberg's twist: 3D structure prediction is now one of the most active frontiers of diffusion research itself, not merely a downstream consumer of it.
02Worldview

No iPhone Moment

There will be <strong>no single iPhone moment</strong> that solves drug discovery — progress compounds in iterative steps, the way even the iPhone and self-driving cars became useful gradually rather than all at once.

Even the phone required iterative development over time to become what is today.

— Evan Feinberg, Latent Space
Key Insight
This reframes how to read every Genesis result: no one release is 'the win.' Feinberg is implicitly warning against both the hype ('AlphaFold solved it') and the despair ('nothing works yet') — the real signal is the slope of the curve over years, not any single paper.
03The Data Problem

No Internet for Molecules

Molecular AI has <strong>no internet to crawl</strong> — the public structure database holds only about 200,000 crystal structures and grows glacially, so Genesis manufactures its own training data with physics.

we don't have the internet to work with. like we can't just download Reddit posts and you know buy some subscriptions to the Wall Street Journal and like train a model and voila the pre-trained model works fairly well.

— Evan Feinberg, Latent Space
Key Insight
Data scarcity is why physics is a first-class citizen here, not a nicety. Because small molecules are cheap to simulate (unlike whole proteins), Genesis can generate synthetic structures at scale — turning a data disadvantage into a moat that pure-ML labs without a physics stack cannot easily copy.
04The Method

Thinking in Crystal Structures

Genesis runs the LLM scaling playbook, but <strong>the model 'thinks' in 3D structures, not tokens</strong> — a diffusion head iteratively refines a pose while physics-based guidance steers each step.

a model is forced to think except it's not thinking in language tokens it's thinking in terms of crystal structures

— Sergey Edunov, Latent Space
Key Insight
Framing structure refinement as inference-time compute is the quiet move here. The same lever that made LLMs smarter — 'let it think longer' — now applies to molecules, so Genesis can trade GPU time for accuracy at inference without retraining. That also explains their appetite for GPUs (see 08).
05The Bar

A Science of Resolution

Two angstroms is <strong>too blurry to trust</strong> — at coarse resolution an aromatic ring can flip and still look fine, so Genesis pushes for one angstrom and below, where a hydrogen bond lives inside a 0.6-angstrom window.

drug discovery really is a science of resolution.

— Evan Feinberg, Latent Space
Key Insight
The scary failure mode is not a visibly wrong answer — it is a confidently sharp-looking one that is subtly off, which a med-chemist then builds months of work on. That is why Genesis treats sub-angstrom accuracy as the line between a tool and a trap, and why 1 angstrom (not the field-standard 2) is their hill to climb.
06Trust

The Pose as a Lie Detector

A single predicted binding number <strong>can't be checked, but a 3D pose can</strong> — the structure is the artifact a human, or an agent, can inspect against physics to catch a hallucination.

you only have a single number and that number might as well be completely hallucinated and you have no means to validate whether that number even makes any sense.

— Sergey Edunov, Latent Space
Key Insight
This is the hinge between the science and the product. Even if you could predict potency as one number, an unverifiable number is dangerous at scale, so the pose exists to make trust possible. It is also the prerequisite for agents — an agent can only act autonomously on outputs it can independently sanity-check.
07The Payoff

Sapphire: Tireless Med-Chemists

<strong>Agents are only as good as the models they orchestrate</strong>, so only after pose, potency, and ADMET all cleared the accuracy bar did Genesis build Sapphire — fleets of AI med-chemists running programs 24/7 with humans as strategists.

agents are only as useful as the underlying models that they're orchestrating.

— Evan Feinberg, Latent Space
Key Insight
Notice the ordering: accuracy first, agents second. A 1.9-angstrom model wrapped in an agent just amplifies slop faster. Genesis gated the agentic platform on the underlying models being good enough — which is why the resolution fight in 05 is the real unlock for the flashy payoff in 07. Both founders insist the human stays the grand strategist, not the operator.
08The Bottleneck

LLMs Are Eating the GPUs

Their <strong>number-one bottleneck is compute</strong> — LLM labs are absorbing GPU supply, but both founders bet the 'alpha' in pure-LLM scaling is thinning and chipmakers will follow demand into life sciences.

GPU prices are going up and up and up and LLM companies are kind of sucking up all of the GPU capacity out there.

— Sergey Edunov, Latent Space
Key Insight
For an AI-infra audience this is the live tension: molecular models are compute-hungry in exactly the way LLMs are, yet they are outbid for the same GPUs. Feinberg's contrarian read — that returns on pure-LLM scaling are flattening while demand for medicines never does — is a bet on where the next wave of GPU allocation flows.