NHTSA and SAE sign on to write the first AV driving test
America's car regulator just hired a standards body to answer the question it has dodged for a decade: what does it mean for software to be a safe driver?
NHTSA has formally executed a cooperative agreement with SAE to launch ASCEND, a government-industry consortium whose first job is to define what counts as a competent automated driving system. Administrator Jonathan Morrison announced the agreement at the AI Horizons Summit in Pittsburgh on September 17, calling ADS competency the consortium's top priority and the input that will "inform and accelerate a corresponding" Federal Motor Vehicle Safety Standard. The consortium pairs the agency with SAE Industry Technologies Consortia — the organization behind the J3016 taxonomy that defines the levels of driving automation — and runs three years on $5 million. Federal law requires competency requirements to be objective and practicable, with metrics for each behavioral competency plus repeatable methods to assess them, which is the part that has never existed: there is still no standardized test for judging how an automated system performs in the real world, even though human drivers have had one for a century.
The rest of the package is less about AI and more about paperwork, but it matters for scale. Morrison said NHTSA has proposed amendments to five FMVSS rules that assume a human at the controls — pedals, mirrors, steering — and is refreshing its AV developer guidance for the first time since 2017, with emergency responder interactions and remote assistance as focus areas. The agency also loosened the exemption process so manufacturers can build vehicles while an application sits in review rather than holding finished inventory. Morrison's framing was explicit: a scalpel, not a hatchet, moving "performance requirements over design specifications" so the rules do not have to predict what the technology becomes. The obvious risk is structural — the companies whose systems get measured are the ones supplying the data and engineering time to the consortium writing the test, and a standard that only ever becomes voluntary guidance will not close anything. NHTSA has already opened investigations and secured recalls against ADS-equipped vehicles, so the enforcement appetite is real; whether it extends to imposing requirements the industry resists is the open question. We looked at the flip side of this gap in September — Tesla's Cybercab launched on a certification NHTSA never approved — where self-certification carried the weight a performance standard would.
Pirate Face launched a permanence layer for open models: every Hugging Face repo becomes a checksum-verified magnet link that keeps downloading after the original host pulls the file. The service mirrors LLMs, image and audio models, and datasets as torrents carrying their official Hugging Face SHA-256 hashes, with a Hugging Face web-seed built into each magnet so a download still completes with zero peers. You can point an existing pipeline at the mirror's endpoint and get the same paths and API, and browsing, downloading, and seeding need no account. It reached the front page of Hacker News within hours of going live. The honest caveat is in the FAQ: submitting a model still requires it to be on Hugging Face first, so this is a survival layer bolted onto a centralized host rather than a replacement for one — and "can never be taken down" is a promise about a swarm, not a guarantee about the bytes. What makes it more than piracy cosplay is the checksum. Tampered mirrors are the reason people distrust model redistribution, and a hash that must match the original is the cheapest possible fix.
What to watch: whether ASCEND's competency standard arrives before the state patchwork hardens, and whether Pirate Face's direct publishing ever ships.
Should a performance standard written with the industry's own test data bind automakers, or is voluntary guidance the honest ceiling? Tell us in the comments.
