Asari AI's agents speed up automaker's behavior prediction model 14.6x on same silicon
songyoupeng · x · 2026-11-20
Asari AI reports its "co-inventor agents" optimized a Detroit automaker's behavior prediction model by up to 14.6x in days: latency dropped from 24ms to 7ms, prediction throughput rose from 42 to 140 FPS, with all optimized kernels matching original outputs to within millimeters. On a $30,000 vehicle, ADAS compute is fixed and shared across models, so recovering compute budget to reinvest in larger, smarter models is key to bringing hands-free driving to its most affordable EV. The work targets the prediction stage of the perception-prediction-planning stack.
More from Embodied
- First Cybercab robotaxi ride review: no driver, no wheel, feels safer than humans — RachelVT42 · 2026-11-20
- Autonomous shipping startup Eight Knot skips end-to-end, avoids deep learning — 4310sy · 2026-11-20
- Waymo got a one-mile exception in Denver just to reach Casa Bonita — reed · 2026-11-20
- Bob McGrew on No Priors: Superintelligence Needs Physical AI and Why Hard Takeoff Is Wrong — saranormous · 2026-11-20
- NVIDIA open-sources Alpamayo 2 Edge, a 2B-parameter VLA model for real-time in-vehicle reasoning — NVIDIA Developer · 2026-11-20
- DOPPEL: one video world model trained across 8 embodiments, from robot arms to humanoids — _krishna_murthy · 2026-11-20