Behind a 0.33Å Error: Drug Design Is Moving From Waterfall to Agile Loop
Chai designed antibodies for 50 targets and hit on roughly half; cryo-EM validation error was 0.33Å. The real moat isn't the model — it's the data, the compute, and a validation chain pharma will pay for.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Drug discovery has to go from waterfall model to agile loop
They analogize biological design to software tools: the interface is like a design suite such as Photoshop, where a "content-aware fill"-style tool can have Chai generate a binder. Traditional drug discovery is a serial waterfall of target discovery, hit discovery, and optimization, where each gate takes months to years and early trial-and-error is extremely costly. If a model can reliably produce promising candidates, that process can become a loop, much like software development moving to agile. The next abstraction layer to climb is agonists, ADCs, and precisely triggered cellular switches.
Hit rates only rise once you switch to CRO-validated targets
Chai ran an internal challenge: design antibodies for 50 targets. At first they picked "interesting" targets, and half didn't work; later they switched to targets already validated by CROs, picking from CRO catalogs, and ultimately got hits for about half of the 50 targets, with an average binding hit rate of roughly 20%. This was the key experiment that got pharma to start believing there was "a real signal," and it explains why large pharma companies were willing to sign big deals.
Open-sourcing Chai 1 was a way to force the infrastructure into place
Chai 1 does structure prediction, is open source, and the team was only 5 people. The point of building Chai 1 wasn't the model itself but using it to get infrastructure like MSA built out in one pass, as a forcing function for long-term investment. Both co-founders have ties to OpenAI, OpenAI led the seed round, and the team at one point sat in OpenAI's office. They pulled all-nighters building the web server 48 hours before launch, Josh went on Bloomberg TV at 7 a.m., and there were no employees in the office.
0.33Å isn't data leakage, it's atomic-level precision
In the Chai 2 paper they overlaid predicted structures onto cryo-EM electron density and couldn't see a difference, with an error of 0.33Å — only one-third the width of an atom. The team at one point thought the other side had sent back their own design, or that there was data leakage; but they had deliberately chosen targets with no known antibody binder, so this was the first antibody hit for that target. This validation is direct evidence that structure prediction has reached atomic-level precision.
Antibodies have no evolutionary template, so MSA fails
An epitope is the site on a protein you want to block, and this is harder than finding an antibody because it requires global understanding. Even AlphaFold2's multimer version is correct on antibody-antigen binding prediction in only about 11% of cases — meaning 90% are wrong. The key is MSA: MSA is essentially a structural template, and antibodies are precisely the thing that cannot have a template in evolutionary terms — every person must generate unique antibodies against the antigens they have personally encountered, relying on recombination to match unknown pathogens.
Compute is allocated by power law, and startups only get the leftovers
Starting last September, spot/on-demand compute was frequently in short supply, and Chai was forced to stockpile GPUs itself. Compute follows a power-law distribution: for example, of 10,000 B300s shipped, hyperscale clouds and top AI labs buy 95%+, leaving startups to fight over the rest. And new hardware is built for LLMs (large KV cache, 72-GPU interconnect), while these structure models need their own inference optimization. Technically, Chai-1's attention batch complexity is L³, memory bandwidth is more of a bottleneck than flops, and even layer norm can be expensive.
Complexity conflicts with the bitter lesson, so the move is subtraction
Complexity and the "bitter lesson" principle are fundamentally in conflict. AlphaFold 3 may have 23 submodules (the speaker himself says he isn't sure), and such a system is extremely hard to optimize and study: change one submodule and it ripples through everything, and the common instinct is to add another module — but maybe you should subtract instead. Chai's engineering culture leans extremely hard toward simplicity; on the office wall hangs a comparison of SpaceX's Raptor 1 and 2 engines, a reminder to keep deleting things. AlphaFold 2/3 also succeeded because they are relatively small models — compute-intensive and data-efficient.
GLP-1 revenue exceeds all AI labs combined
Per token, pharma may be the domain with the highest downstream value per token: a single drug can be a multi-billion-dollar asset, and the two GLP-1 drug classes combined may be a trillion-dollar asset. By Chai's estimate, until three months ago GLP-1's total revenue exceeded all AI labs combined, yet its market cap is far lower. Pharma is essentially a VC-style business; one of Silicon Valley's biggest early venture returns was Genentech in the 1980s, and the biggest problem in biopharma as a whole is precisely the funding model.
In their own words · checked verbatim
We see ourselves as almost a neutral software factory for making medicines.
Josh, our CEO, likes to say that our, our biggest competitor is the mouse.
and in this case, it was a 0.33 angstrom error, which is one third the width of an atom.
LFOLD2 got, like, I think 11%. The multimer version of this got, like, 11% of antibody antigen prediction cases correct. That means 90% of the time it's wrong.
there's, say, 10,000, you know, B300 units that are shipping everywhere, right? The hyperscalers and the, you know, the biggest AI labs are buying 95 plus percent of it
complexity and being bitter lesson-filled are, like, fundamentally at odds.
in the case of GLP-1s, I think the two GLP-1 drugs combined, like, maybe a trillion dollar asset.
the cost of actually making a drug in pharma is kind of, like, increasing exponentially.
Figures
| Partner pharma companies | Eli Lilly, Pfizer, Novartis, Argenx (4 companies) | 3:21 |
| 50-target validation hits | Antibodies designed for 50 targets, roughly half yielded hits | 7:27 |
| Average binding hit rate | About 20% | 22:59 |
| When multimer structure prediction models appeared | 2021 | 5:26 |
| Chai 2 cryo-EM validation error | 0.33 Å (1/3 the width of an atom) | 34:15 |
| AlphaFold2 multimer antibody-antigen prediction accuracy | 11% (i.e., 90% wrong) | 54:10 |
| Share of B300 shipments going to top buyers | 95%+ | 1:02:33 |
| Chai's new funding | $400 million | 1:07:59 |
| AlphaFold 3 submodule count (estimated) | 23 | 1:09:02 |
| Total Chai team size | 30 people | 1:27:35 |
Glossary
- binder
- A candidate molecule that binds to a target; in this episode, an antibody or antibody fragment.
- epitope
- The specific region on an antigen that an antibody recognizes and binds to.
- MSA
- Multiple sequence alignment: aligning homologous sequences as an input feature for structure inference.
- cryo-EM
- An experimental method that resolves a protein's 3D structure with an electron beam at low temperature.
- durable execution
- An engineering paradigm that lets long multi-step workflows automatically retry and resume from checkpoints.
- bitter lesson
- The AI principle of replacing domain knowledge with compute and search.
How to listen
Founders in biocomputing/AI4S, heads of R&D digitization at pharma companies, and pharma investors trying to judge when structure prediction will actually save money.
You can skip the Chai 1 launch details from 18:54-24:01, but don't miss the economics at 1:24:23.