Anthropic says Claude optimised 30 scientific AI models in under four weeks

Anthropic says a general-purpose Claude research model optimised more than 30 specialist scientific AI models in under four weeks, work it says experienced engineering teams can take weeks to perform for each individual model.

The models are used for protein structure prediction, protein design, genomics and other biomolecular research.

Anthropic says Claude made them roughly four times faster on average with minimal loss of precision. Versions preserving identical outputs also achieved substantial acceleration.

The project was primarily supervised by two Anthropic technical staff with biomolecular modelling experience but no previous experience in inference optimisation or GPU kernel engineering.

Claude produced custom GPU kernels, removed unnecessary calculations, cached repeated computations and developed different optimisations for different models.

Anthropic used an internal general-purpose research model for the work. The results therefore should not be read as demonstrating that every publicly available Claude model can reproduce the project.

Make the tool faster once, benefit repeatedly

The potentially important result is not simply the four weeks.

It is what was left behind afterwards.

Anthropic has released the optimised implementations as open-source software. Researchers whose workloads use those models can potentially inherit the improvements without repeating the optimisation project themselves.

That changes the economics of the result.

Using AI to help one scientist complete one task faster produces a one-off productivity gain. Using AI to make a scientific tool four times faster can produce gains whenever researchers subsequently use the improved implementation.

Anthropic also says Claude developed a low-memory mode capable of accurately modelling biomolecular systems containing more than 10,000 biological tokens on a single GPU node.

Reducing memory requirements can make previously expensive computational work accessible with less hardware.

Claude pushed the approach beyond 70,000 tokens on an eight-GPU node, but those much larger predictions produced biologically incorrect structures. Removing the computational limit did not remove the scientific one.

A separate 100-fold compute comparison

Anthropic also tested Claude in a separate protein-design experiment.

A single Claude research model was given one H200 GPU for 24 hours and allowed to produce designs without sub-agents or human steering of individual designs.

Across 16 protein targets, Anthropic says the resulting designs achieved computational binding scores comparable with an earlier approach that had consumed roughly 100 times as many GPU hours.

The comparison does not show that model optimisation alone produced that reduction. The hardware and design workflow also changed between the experiments.

Nor does it demonstrate successful new medicines.

These are computational results. Similar in-silico binding scores do not establish that the proteins will perform similarly in laboratory experiments, let alone produce a useful treatment in humans. Anthropic is arranging experimental validation of the designs.

But the optimisation project demonstrates a more immediate route through which AI could accelerate science.

A general-purpose AI can improve specialist scientific AI.

Those improved models can then run faster, use less memory or require less compute.

And researchers using them can inherit those improvements.

AI’s first large effect on science may therefore arrive before a machine independently makes a great scientific discovery.

It may come from making the tools scientists already use dramatically cheaper and faster.

Sources

• Anthropic – Claude uplifts biomolecular modeling