Ninety-nine to one. The Senate just blew up the one federal lever that could have made any AI rulebook coherent — and the fight didn't end. It moved into fifty rooms at once. If you're just joining: state AI regulation was already narrowing before this fight. Colorado swapped its 2024 comprehensive AI Act for the ADMT Act — cutting explicit governance duties and expanding regulated-sector coverage, while keeping discrimination liability on the books. Across the states, rules are getting quietly revised under pressure from overregulation worries, federal preemption pushes, and international regimes that keep slipping. This is the AI Daily Briefing. Today: a benchmark survey lands the same week the measurement regime gets its own reckoning. And we ask which state actually files something first. Cassidy, the vote. This one's from arXiv:
We systematically review the current status and development of large language model benchmarks for the first time, categorizing 283 representative benchmarks into three categories: general capabilities, domain-specific, and target-specific. General capability benchmarks cover aspects such as core linguistics, knowledge, and reasoning; domain-specific benchmarks focus on fields like natural sciences, humanities and social sciences, and engineering technology; target-specific benchmarks pay attention to risks, reliability, agents, etc.
Two hundred eighty-three benchmarks. That number matters here because the measurement regime starts to look like a junk drawer. And the survey's own authors flag data contamination as a known way scores get inflated. So the tool everyone uses to declare a winner is saying, in print, that the numbers can be cooked. Right, and they say benchmarks guide model development. So if the yardstick's bent, you don't just mismeasure — you start optimizing for the bend. Step seven of a ten-step chain doesn't show up on any of these 283. Here's the timing that gets me, Bill. This drops the same week Oracle commits to fifty thousand MI450s. If a benchmark score can't tell you what a model's worth, what justifies a purchase order that size? Somebody's pricing capability off proxies the literature just called unreliable. And we still don't have Oracle's ROCm numbers back at scale. So when they do report, this survey's the frame: 'Oracle says it worked' and 'we have a valid read on whether it worked' are two very different claims. So the Senate just killed the federal AI moratorium. Does that actually leave every state free to write its own rules? And which ones are going to hit companies first? Yes — and the vote was about as decisive as it gets. Senators voted 99-to-1 to strip the moratorium from the reconciliation bill — that's per IAPP reporter Caitlin Andrews — after a last-minute compromise between Senators Marsha Blackburn and Ted Cruz collapsed. Blackburn crossed the aisle and joined Democrats on the removal amendment. So the federal firewall is gone. In practice, though, states aren't starting from zero. Kevin Collier and Bruna Horvath at NBC News note the bill moves forward without the proposed ten-year block on state AI laws, and states have already been busy. Marc Levy's reporting has states zeroing in on chatbots interacting with children, employers using AI in hiring and workplace decisions, and developer duties to prevent catastrophic AI failures. The sweeping bills — the big liability frameworks — have mostly been vetoed or stalled because governors worried they'd kneecap development. What's moving instead, per Debevoise and Plimpton's analysis right after the moratorium died, are narrower fit-for-purpose rules: harder to kill politically, easier to enforce. If companies thought the big comprehensive bills were the main threat, and those keep dying, are they underestimating how much the targeted stuff could cost them? They are. Cooley's state AI law tracker flagged this directly: compliance deadlines were already arriving in 2026 on laws passed just a few years ago, and several have since been amended or delayed — which creates its own whiplash for legal and engineering teams. The first laws to watch are the ones with clear enforcement hooks: child safety rules for AI products, employer disclosure requirements, and developer obligations around high-risk systems. If you're building or investing in AI right now, plan for the patchwork. The next twelve months of state legislative sessions will decide how jagged it gets. If AI Daily Briefing helps you keep up, consider subscribing and leaving a quick review wherever you're listening. It really helps other people find the show, and it helps us keep making it better.
You'll find links to all of today's stories in the show notes, so if one caught your ear, you can go deeper there. Thanks for listening. We'll be back tomorrow. That's AI Daily Briefing for today. This is a Lantern Podcast.