Biologists use specialized open-source models for tasks like modeling the structure of molecular systems, designing drug-like molecules, and predicting the effects of genetic mutations. But these models are often expensive to run, potentially limiting their impact. In our latest Science Blog, we share how Claude was able to optimize inference for more than 30 open-source models, making them 4x faster on average, partly by writing custom software for GPUs. We’re open sourcing all of the optimization code. Read more here: https://lnkd.in/gjg7ACnz To show what these optimizations make possible, we’re partnering with Adaptyv Bio on a protein design competition. Together, we’ll be experimentally validating over 5,000 designs. We're providing up to $1 million in Claude credits plus funding alongside Adaptyv for experimental validation. Modal is contributing up to $250,000 in compute and Twist Bioscience is providing DNA. Learn more on Adaptyv’s Proteinbase: https://lnkd.in/ddwgnVE4 And sign up for the competition here: https://lnkd.in/g5v6-cQQ You can find all of the code on GitHub: https://lnkd.in/gPgDEDPV And the full results in our technical report: https://lnkd.in/gwdtN2-z
Per-target design cost went down from $10k to about $150 in this report, two orders of magnitude in GPU hours. That's what actually opens up who can afford protein design, more than the $1M prize pool everyone's focused on
The open-sourcing part is what stands out to me. Improving model performance is useful, but making the optimization work available means other teams can actually build on it instead of keeping the gains locked inside one company.
This is the kind of optimization that turns model capability into usable science: lower inference cost means more teams can test, reproduce, and improve the work. The benchmark detail matters too—separate faster runs from quality regressions so “4x” stays meaningful outside the demo.
The speedup is interesting, but the verification boundary is more interesting to me. You can change a lot internally: collapse redundant computations, simplify branches, rewrite kernels, reduce memory. That’s all fine. Now, how do we know the distinctions the scientific result depends on survived that compression? “Faster” is easy to measure. “Still means the same thing” is way harder. Eventually the wet lab gets the final vote.
I'm using battery. It sounds robotic it hurts my ears. I'm going to lose money and I'm going to Sue if he's not fixed I lose money. I can't feed faces. You're taking food from my kids. Fix it. I can't listen to it. I can't read so I have to rely on his TTS but I can't because it hurts. Fix it. Fix it. Fix it please for the love of God fix it. I know I've got another conversation going but the guy's ignoring me. It's been 2 days I have messaged and messaged and messaged. If something's not fixed soon, I'm going to soon because it's if I can't build because of the pain that this robotic voice is causing my ears. I lose money. If I lose money I can't feed my kids. They have less than 24 hours to respond and actually fix this problem or I am suing anthropic
The distinction between faster runs with identical outputs and faster runs with small numerical differences is worth keeping visible. If a team adopts these optimizations, I'd want those modes clearly labeled in its benchmarks and saved results. Otherwise, a later comparison can quietly end up testing two different things.
What stands out here is that AI is no longer only generating scientific hypotheses—it is beginning to optimize the computational machinery through which science itself is performed. That can dramatically expand the number of candidate molecules, proteins, and biological systems we can explore. But as generation and inference become cheaper and faster, the bottleneck shifts from producing candidates to validating them. Compute Efficiency ≠ Model Validity ≠ Biological Validity ≠ Experimental Success. The next frontier may therefore be a closed scientific loop in which AI-generated optimizations, model versions, experimental outcomes, failures, and corrections are all preserved with clear provenance. If thousands of designs can now be generated and tested, the real advantage will come from knowing which results deserve to become reusable scientific knowledge—and which failures must remain part of the evidence.
Wet-lab validation of 5000 designs turns generative output into actual scientific evidence. Most AI-for-science stops at simulation. Closing that loop, and funding it, is what separates demos from discovery.