When a model answers one request at a time, most of the GPU sits idle and each of hundreds of small steps adds launch overhead.

Megakernels that run a whole model pass as one GPU program, plus speculative decoding, lower-precision numbers and keeping data on each die's local memory.

The space of possible proteins is too large to search, and older tools needed a separate step to turn shapes into sequences.

An open model, made with the University of Washington's Institute for Protein Design, that designs sequence and structure together and spends extra compute on promising designs.

Agents that run for hours can download code, ask for more access and touch more systems, so they cannot be trusted to police themselves.

An open-source runtime with a gateway, a supervisor that checks outgoing requests, a sandbox and a policy prover, run on Nvidia's Vera CPUs.

Data usually passes through CPU memory on its way to the GPU, and the main interface for the direct path is a C library.

A tutorial for the KvikIO Python library, showing how to read files into GPU memory and check whether the fast direct path was really used.