CMS's AXOL1TL filters LHC data in silicon

This is the opposite of chatbot AI: the win is not model size, it's fitting useful inference into nanoseconds and a tiny slice of silicon. CERN is showing that the hardest AI infra problems are really hardware scheduling problems with ML attached.

• –AXOL1TL is signal-agnostic, so it can flag weird events without hardcoding one physics hypothesis
• –Putting inference in FPGA firmware makes latency predictable, which matters more than raw throughput when the trigger decides what survives
• –HLS4ML is the bridge from trained model to deployable logic, turning research code into detector hardware
• –The 2025 V5 update suggests real operational tuning: better feature extraction, more rate budget, and continuous iteration
• –If HL-LHC delivers the predicted data flood, similar tiny-model hardware pipelines may show up in other low-latency domains