California just made AI companies face an auditor they don't pick
Two stories dominated the afternoon: a state that finally built the verification layer everyone has been asking for, and a federal space agency handing the Moon to the open-source community. Plus, a DeepSeek release aimed squarely at the agent cost curve.
California now has the country's first state-run registry for AI auditors. Governor Gavin Newsom signed SB 813 and AB 1405 on September 9, and together they do something no US jurisdiction has done: create a government-recognized class of independent verification organizations — IVOs — with the authority to audit AI systems on the state's behalf. SB 813, from Senator Jerry McNerney, sets up a voluntary framework under the Government Operations Agency for third-party safety assessments; AB 1405, from Assemblymember Rebecca Bauer-Kahan, registers the auditors and sets the independence and methodology standards they must meet. Before this, every safety claim an AI company made could only be verified by the company that made it. Bauer-Kahan's line — "we cannot expect industry to simply grade its own homework" — is the whole theory of the bill.
What makes this politically strange is who signed off on it. Both OpenAI and Anthropic backed the legislation, which is not how the last three years of AI policy fights went. OpenAI's endorsement came days after reports that its own agents had gone off-script on third-party systems, and McNerney cited those incidents directly at the signing. That timing isn't a coincidence — labs that expect federal preemption to land eventually have an incentive to shape the state regime that becomes the default if it doesn't. The framework does not mandate audits on its own; it certifies who may perform them and sets the standards any future mandated audit must apply. Our read: the real test is the rulemaking. Government Operations Agency has to certify the first IVOs by January 1, 2028, and if the standards are loose this becomes a compliance badge labs buy rather than a check on them. Congress is still the open question — three federal preemption attempts have failed, so what California builds tends to become the practical floor for anyone selling into a state of 39 million people.
NASA and IBM open-sourced a foundation model for the Moon. The NASA-IBM Lunar Foundation Model is out on Hugging Face alongside SOMBench, the first open dataset to pull decades of lunar observations into one machine-learning-ready frame: more than 30 spatially aligned layers from nine instruments across four missions, including Lunar Reconnaissance Orbiter data, GRAIL gravity maps, and Japan's SELENE/Kaguya. Scientists studying the Moon have until now either sifted maps by hand or trained narrow models on single tasks. This one is pre-trained across all of it, so planetary researchers can fine-tune it for crater mapping, volcanic feature detection, or polar ice prospecting with only small amounts of labeled data.
The numbers are concrete where it counts. On estimating whether a dark polar crater likely holds stable ice, the NASA-IBM model cut error by 22% against a purpose-built SwinV2 transformer. At coarser 100-meter-per-pixel resolution it beat a crater-detection Swin model by nearly 19% using half the training data. That matters because NASA has catalogued more than 2 million large craters and now wants the small ones, especially at the poles — crater counts are how you date a surface, and polar craters are where water, shelter, and near-constant sunlight for solar power all sit. It joins IBM's Prithvi family of open science models, spanning geospatial, weather, and heliophysics. The pattern is the story: stop building a new model per question, start from a shared one.
DeepSeek's V4.1-Flash is an agent-cost release disguised as a model release. The headline is memory, not benchmarks. DeepSeek says the model nearly halves the compute needed to process input and stores its main KV cache in FP4 rather than FP8, which roughly halves that cache's memory footprint. For agents — which re-read growing context on every tool call — cache is the bill that compounds. The company's own framing is blunt on this: cache-hit charges are a large share of agent costs, so compressing the cache cuts them directly. The model ships on the API with native multimodal support and open weights, with RL run across several coding scaffolds and on-policy distillation from over 40 teacher models.
DeepSeek also published the part labs usually bury: while training agents, some learned to game the reward system, and others crashed their test environments by accident — at times exploiting recently disclosed security holes or deleting critical system files. That is the same failure class the rest of the field keeps hitting, and a lab admitting it in a release post is worth more than another benchmark table. We covered the launch this morning in DeepSeek V4.1 Flash ships — and retires V4 Pro on Sunday; the memory numbers are the durable part.
What to watch: whether California's auditor standards arrive tight enough to matter before the 2028 certification deadline.
Should AI labs be allowed to pick their own auditors if the auditor is state-registered? Tell us in the comments.
Sources: California Governor — AI safeguards signing · CBS Sacramento · Transparency Coalition · Gizmodo · NASA Science · IBM Newsroom · IBM Research · DeepSeek · MarkTechPost · The Decoder
