Latent Space

Matt McPartland & Neil Patil on turning biology into a software factory

Matt McPartland & Neil Patil· Co-founder & Product Lead at Chai Discovery
·~95 min·English·Latent Space
LLMTrainingAI CompanyAI Infrastructure
TL;DR

Chai Discovery's founders explain how AI structure and design models are turning antibody discovery from an obscure, trial-and-error natural science into a software-like precision-engineering discipline.

01Core Mental Model

The Neutral Software Factory

Chai refuses to make its own drugs and instead builds the modeling and design layer for everyone else — so its product looks less like ChatGPT and more like Figma or Autodesk for molecules.

We see ourselves as almost a neutral software factory for making medicines.

— Neil Patil, Latent Space
Key Insight
The word "neutral" is a business moat as much as a product stance: by refusing to own drug pipelines, Chai turns every pharma's IP paranoia into a reason to trust it — and turns would-be rivals into paying customers.
02Why Now

Our Biggest Competitor Is the Mouse

The old way to find an antibody was to infect a mouse or screen billions of molecules for a needle in a haystack, ending with a handful of hits nobody understood; Chai bet that general models could design them on purpose instead.

Josh, our CEO, likes to say that our biggest competitor is the mouse — or nature, in certain ways.

— Matt McPartland, Latent Space
Key Insight
Choosing to "bet on the models getting better" over dissecting the 25 targets they missed was a bitter-lesson wager — the same scaling faith that drove LLMs, applied to a field where most people still hand-engineer around each failure.
03The Model Jump

From Reading to Writing

Chai 1 read biology — given a sequence, predict the structure — while Chai 2 writes it, taking a target and generating a binder's sequence and shape together, which is the moment the tools crossed from useful to transformative.

I have a target structure that I want to design a binder to. Chai 2 will generate candidate molecules — candidate medicines — that bind to that target.

— Neil Patil, Latent Space
Key Insight
Prediction has a ground truth you can hold out and check; design does not. Crossing into generation meant giving up a verifiable benchmark — which is exactly why validation (next section) becomes the binding constraint on the whole field.
04The Hard Part

Feeling Around in the Dark

The deepest problem in AI-for-biology is that you often can't even measure whether a design worked, so the field leans on self-consistency metrics and slow lab loops — and when cryo-EM finally checks a Chai design, it matched to a third of an atom's width.

One of the things I didn't realize about biology is how much of it is literally feeling around in the dark — and that's not even a metaphor. You literally can't see how these things look.

— Neil Patil, Latent Space
Key Insight
When success is expensive to measure, benchmarks get gamed — self-consistency collapses if every protein looks identical. Chai's real product may be less the design model than the validation loop that tells it, quickly, when it is right.
05New Capabilities

Design Reaches Modalities Immunization Can't

AI design reaches drug modalities that immunization can't reliably discover — like agonists that flip a cellular switch and bispecifics whose two arms must each bind — making new drug classes, not modest discovery savings, the real economic prize.

That comes from the mission of the company: to turn drug discovery from a scientific experiment to an engineering discipline.

— Neil Patil, Latent Space
Key Insight
A bispecific needs both arms to hit — multiplying two one-in-a-billion odds — so no immunization budget can stumble onto it; only deliberate design can. The value isn't shaving cost off known drugs, it's making drug classes that couldn't be found by luck.
06Engineering Culture

Delete, Delete, Delete

Chai treats simplicity as a scaling strategy: a model with two dozen hand-tuned submodules is nearly impossible to improve, so the discipline is to remove complexity rather than add another special-case module.

Complexity and being bitter-lesson-pilled are fundamentally at odds.

— Matt McPartland, Latent Space
Key Insight
Simplicity here isn't aesthetics, it's survival on the scaling curve. If you can't reason about what tweaking submodule 21 does to the whole system, you can't systematically improve it — so deleting complexity is how you keep riding the bitter lesson.
07Infrastructure

The Compute Market Is LLM-Pilled

Structure models are compute-heavy and memory-bandwidth-bound in ways that are almost the opposite of transformers, yet every modern GPU cluster is optimized for LLM-shaped work — so Chai has to re-optimize the stack the rest of the market ignores.

It's interesting how much the compute market has gotten LLM-pilled. This class of models is going to be just as impactful as LLMs — but the compute almost doesn't realize that yet.

— Neil Patil, Latent Space
Key Insight
This is a roofline story: structure models are compute-bound and memory-bandwidth-bound with tiny hidden dimensions — the inverse of what LLM-era GPUs assume. Chai's edge partly comes from re-tuning operators that everyone else leaves on the table.
08Economics

Moore's Law, Backwards

In computing, capability rises and cost falls on a smooth exponential; in pharma the cost of developing a single drug climbs exponentially instead, and Chai's real pitch is that AI design can finally bend that curve back down.

It's Moore's law backwards. In compute you have this nice exponential scaling — near the exact opposite in pharma, where the cost of making a drug is increasing exponentially.

— Matt McPartland, Latent Space
Key Insight
If cost-per-drug keeps compounding, the marginal return on a new drug eventually goes negative — a slow-motion end state for the industry. Chai's pitch to pharma isn't faster drugs; it's bending an exponential that otherwise kills the business model.