Eigen RadarAI
Analysis

Anthropic claims a fourfold speed-up across open biomolecular models

Anthropic dated a 17 September research note on Claude Science. Inside four weeks the desk logged more than 30 open biomolecule models. Wall-clock time in the lab, the note said, fell by a factor of about four. A low-memory mode kept strings above 10,000 tokens on one Nvidia GPU node. The card is that vendor note, not an outside bench mark.

Artificial Intelligence··Midday
Gloved hands examine a three-dimensional biomolecular model made of colorful atom spheres on a bright wet-lab bench.

Thirty models, four weeks

The 17 September Anthropic note is the retained card. Claude Science, it said, filled more than 30 open biomolecule models before four weeks were up. Lab wall-clock, in that same note, dropped by a factor of about four. That is the vendor counting its own desk. No second publisher is on this slot's card.[1]

Triangle geometry

The speed line is tied, in the note, to triangle-style geometry steps. A token, it said, counts an amino acid, a nucleotide, or an atom on a small molecule. Low-memory mode, the same note added, kept strings longer than 10000 tokens on one Nvidia graphics node. That is still a capacity line from the seller. Readers should not treat the geometry claim as an outside replication.[1]

A wet-lab contest

The same note is the only retained publisher for this development. The exhausted corroboration record sits on the card. Nothing here is a clinic paper or a third-lab clock. Treat the four-week and four-fold lines as Anthropic speaking. A later independent clock would need its own card.[1]

References

  1. News sourceAnthropicMore than 30 open biomolecular models ran roughly 4 times faster↩1↩2↩3