FLOPs
FLOPs at Beyond Market Intelligence is a file of 3 stories. The newest of them: “A free guide to making ML models faster, from silicon to agents”, “Exploring how an older model still competes on accuracy and speed today”, and “Why weight-space perception fails when networks train independently”. Performance engineering isn't just about reducing FLOPs. CABiNet, a 2021 architecture that went quiet after ICRA, is back, and it's beating a 2026 generalist model on aerial segmentation. Acme AI is the next-generation, AI-powered spreadsheet platform built to replace Excel and redefine how analysts, data scientists, and enterprise teams work… The list below is every FLOPs story on Beyond Market Intelligence, newest first.
A free guide to making ML models faster, from silicon to agents
Performance engineering isn't just about reducing FLOPs. Usamah Zakir spent months writing the guide he wished he'd had when starting out: *How to Make Your Model Fast*. It moves from silicon to agents, using roofline analysis to help you diagnose whether you're compute, bandwidth, or memory bound before optimizing. That systems-first thinking is rare and valuable. The whole thing is free and open source. If you're working on inference or edge AI, it's worth exploring.
Exploring how an older model still competes on accuracy and speed today
CABiNet, a 2021 architecture that went quiet after ICRA, is back, and it's beating a 2026 generalist model on aerial segmentation. On UAVid, CABiNet-L hits 67.14 mIoU at 4.44 ms; YOLO26x-sem trails at 64.41 and takes nearly three times longer. The gap is mostly small classes: people and vehicles. That's not a universal win, YOLO26s matches compute and still wins on latency. But at the high-accuracy end, purpose-built still has teeth. For a deeper look at how training recipes shape these comparisons,
Why weight-space perception fails when networks train independently
The symmetry story in weight-space learning finally has the hard numbers it needed. This study, built on roughly 1.8 million fitted SIRENs, shows that randomizing only the exact symmetry group destroys 79.1 of the 80.4 accuracy points separating shared-init from random-init networks. That is sufficiency, cleanly demonstrated. The deeper insight, though, is computational. If a complete invariant matches function access informationally, then weight-space's real edge must be efficiency, not insight. That reframing deserves attention.