CERN's AXOL1TL filters collisions in nanoseconds

This is the rare AI story where smaller is the point: the win comes from squeezing a useful model into a control loop with nanoseconds, not from scaling parameters. AXOL1TL is a reminder that the most valuable AI systems are often the ones that disappear into hardware and just make the right decision.

• –CMS sees 40 MHz collision rates, so the trigger must reject nearly everything while preserving rare physics signatures.
• –The stack is hardware-first: a VAE/autoencoder is trained on ZeroBias data, then translated with hls4ml and deployed on FPGAs; see the CMS paper https://arxiv.org/abs/2411.19506.
• –CERN's 2025 note documents V3, V4, and V5 deployments, which suggests a mature operational pipeline rather than a one-off experiment: https://cds.cern.ch/record/2942560
• –The broader edge-AI lesson is already visible in CERN and Ceva's earlier work: compressed models plus custom silicon can deliver big speed and energy wins when latency is the real product (https://kt.cern/news/news/computing/how-cern-and-ceva-are-pioneering-future-edge-ai).