AI #176 Part 2: Plan B
Zvi's AI Roundups · Zvi Mowshowitz · 2026-07-10
Zvi Mowshowitz surveys AI policy, open-weight model safety risks, and alignment research, covering the DoW-Anthropic contract collapse over surveillance redlines, Nvidia's chip export misrepresentations to the US government, and Anthropic's new GRAM capability-routing training technique.
Appears in
Extraction
Topics: ai-policyopen-weight-modelsai-safetyalignment-researchai-governance
Claims
- The Trump administration will not create a formal AI licensing regime, instead relying on ad hoc executive authority that Zvi characterizes as opaque and arbitrary, with winners chosen by the whims of those in power.
- Open-weight models at roughly 'Mythos-level' capability now pose credible national security risks, a threshold both US and Chinese governments are beginning to act on through export controls and release restrictions.
- Nvidia provided materially false information to the US government about Huawei's AI chip production capacity in order to influence export control policy in its favor.
- Anthropic's contract negotiations with the Department of War collapsed because the Pentagon demanded blanket 'anything lawful' usage rights that eliminated Anthropic's redlines on mass surveillance and autonomous weapons targeting.
- Anthropic and AE Studio developed GRAM, a gradient-routed auxiliary module technique that routes dual-use capabilities into removable model compartments, potentially replacing suppression-based safety training with a cleaner developmental architecture.
Key quotes
"This simply amounts to a blanket posture of 'anything lawful,' and the statement from the Pentagon's spokesman confirms that this is the intent. I unfortunately don't see a way forward given these categorical statements." — Dario Amodei, Feb 26 2026
"The answer to 'where do we start to get big risks of serious problems' seems now to be established at 'somewhere between Opus 4.8 and Mythos.'"
"Open weights advocates that are doubling down on wanting 'Mythos-level' open weights models ASAP are revealing their decision processes to be a rock with 'Open Is Good' written on it."