The open weights battle moved on three fronts in seven days: a near-final White House pre-release review framework, a Dario Amodei paper denying Anthropic ever wanted open models banned, and a 70-partner Nvidia open-model alliance.
The Trump administration is close to finalizing a voluntary framework under which AI companies would submit their most advanced models for government review before releasing them to the public, The Information reported. It is one of three open-weight developments in a single week. Open-weight models can be downloaded, run, and fine-tuned by anyone; you get the trained parameters that define the model, but not the training data or code that full open-source models provide.
The White House's Office of the National Cyber Director circulated a draft to OpenAI, Anthropic, and Google, which jointly submitted edits by an August 1 deadline set by an executive order. Applicable frontier models, the term for the most capable systems, would get up to 30 days of review by the National Security Agency and the Center for AI Standards and Innovation.
Amodei, meanwhile, published a position paper answering claims that Anthropic wants open models banned to protect its business. "Let me state it clearly so there's no doubt. Anthropic has never advocated for a ban on open weights models," he wrote, calling open models without dangerous capabilities a public good. His fixes: keep powerful chips from authoritarian governments, crack down on industrial-scale distillation (training a copycat on a powerful model's outputs), and mandate safety testing for all sufficiently capable models, open and closed.
Nvidia, with roughly 70 partners including Microsoft, IBM, Hugging Face, and Palantir, launched the Open Secure AI Alliance to build and share open models and tools for cyber defenders, citing an incident where defenders used open-weight Chinese models for forensic analysis. OpenAI and Anthropic are absent.
SmarterX founder and CEO Paul Roetzer broke down all three fronts on Episode 228 of The Artificial Intelligence Show.
7 - Number of days in which all three open weights developments evolved
August 1 - Deadline for OpenAI, Anthropic, and Google to submit edits on the White House draft
30 - Maxiumum days for a government review period for covered frontier models
~70 - Partners in Nvidia's new Open Secure AI Alliance
The people shaping AI policy mostly lack AI experience. Wired's breakdown of Donald Trump's AI brain trust is blunt: The president himself does not use a computer. National Cyber Director Sean Cairncross is a former political campaign lawyer with no tech or AI experience. The one adviser with real technical background, David Sacks, pushes a hands-off approach to his 1.6 million X followers. All this while the stakes of who decides are sky-high: "It's pretty realistic that by the end of '28 we will be talking about post-AGI worlds, post-AGI economies," says Roetzer, referring to artificial general intelligence, AI that can handle most work as well as a human.
Grok would not defend the government AI brain trust. Roetzer put the list to Grok, the AI built into X: "Are these really the best people to be deciding this? Be honest." Grok's answer, which Roetzer shared on X, was point blank: "No." It called the group "not the optimal technical or policy brain trust for deciding the future shape of the AI industry."
Amodei's warnings deserve a reading on the merits. Roetzer's ask: Take Dario's name off of the paper and judge the arguments independently. Amodei's two concerns: authoritarian governments using more powerful models for permanent military superiority and deep repression, and misuse for cyberattacks or biological attacks, plus alignment problems, which are models that do not always do what they are told. Open weights are harder to guardrail and, once released, cannot be withdrawn. He challenged the claim that broad access helps defenders more than attackers: "It seems at least as likely to me that the opposite could be true."
"You can not like Dario, that's fine. You can not share his concerns, that's fine, too, but the dude is one of the five people in the world that has a front row seat to what's coming in the next 12 to 24 months, and he seems very honestly concerned."
— Paul Roetzer, founder and CEO of SmarterX, Episode 228 of The Artificial Intelligence Show
The open-model camp is playing for unity, from the 1,300-signature letter to the new alliance. The more telling signal: AI's two fiercest rivals agree. "If OpenAI and Anthropic have found common ground on anything, then maybe we should all listen a little bit," says Roetzer, instead of writing it off as purely a greedy move.
His deeper worry: too many loud voices in this debate want to be right more than they want the right outcome for society.
Three definitions will decide how much the government review-process matters. How the administration defines a frontier model, whether it treats closed and open models differently, and how voluntary it stays in practice, will determine the effectiveness and fairness of the regulations.
Thinking Machines Lab has staked out the middle ground. Led by former OpenAI CTO Mira Murati, the lab published A Safe Path to Open Weights, which stages access from monitored APIs up to a full open-weight release. It widens only when the evidence supports it.
The open weights fight is a governance question, and most companies have not answered their own version. According to the 2026 State of AI for Business Report, only 13% of organizations have all four AI governance foundations in place: an AI council, an AI roadmap, generative AI policies, and an AI ethics policy. Nearly a third have none of them.
The report is built on more than 2,100 responses across roles, functions, and industries, and it is the clearest benchmark for whether your own AI governance house is in order. Read the full report →