Sunday, October 18, 2026
TL;DR
Zvi Mowshowitz says a leaked federal AI bill Anthropic is shopping lawmakers preempts stronger state laws while weakening enforcement. OpenCode's breach exposed stored API keys, and Microsoft pulled out of a UK data-center deal over grid-power delays.
Act on this
- Revoke and reissue any OpenAI, Anthropic or Google key stored in OpenCode Console via each provider's own dashboard, not an emailed link, then check billing since Oct 17. Cyber Security News, OpenCode breach #
Signals
Anthropic is shopping Congress a weaker AI safety bill than its own public framework calls for #
Zvi Mowshowitz's Oct 17 read of a leaked draft, first reported by Politico, found preemption of state laws stronger than the bipartisan FRONTIER Act's but enforcement and auditor licensing weaker - weaker, Alex Bores said, than Anthropic's own June framework. Nathan Calvin called it less protective than existing law; Nate Soares said it would make courts less able to block a dangerous AI. "Anthropic needs to course correct, quickly," Zvi wrote. Zvi, Leaked Anthropic Draft AI Regulatory Bill · Reuters, via The Daily Guardian
The gap between the best American and Chinese open-weight models has closed by about half, one close tracker says #
Nathan Lambert wrote Oct 17 that Meta's Muse Spark Open, "the strongest American-made open model by quite some margin," plus Reflection's Beam keep more than 20 Chinese open models from holding every top spot; together, he says, "the gap between the best American and Chinese models has been reduced roughly by half." He calls Muse Spark Open's custom license - mandatory "Built with Muse" attribution plus a separate license required above $20 million a month in revenue - "an unnecessary strategic risk." Lambert, Meta shakes up the frontiers, open and closed
Compute scarcity is routine enough now that practitioners are building their own workarounds, not just complaining about rate limits #
Geoffrey Huntley's Oct 17 post describes three tools he built and runs himself: Underclass, which pools multiple ChatGPT subscriptions on one account to keep the prompt cache intact; Preflight, which redacts secrets from outbound model requests; and Geoclass, an experimental router that retries capacity-limited requests from other regions over a Tailscale mesh. "We're in a cost of compute crisis," he wrote - "we're software engineers, we can solve these problems." Huntley, pooling OpenAI subscriptions and bypassing regional inference capacity limits
News
WATCHMicrosoft pulls out as anchor tenant of Nscale's UK data center #
Microsoft ended its agreement to lease 23,000 Nvidia chips at Nscale's planned Loughton, Essex site, after grid operators said power wouldn't arrive until the early-to-mid 2030s. Nscale says it remains committed and is seeking alternative power; Microsoft has not commented. GuruFocus
The long view
A year ago, California had just signed SB 53, the country's first binding frontier-AI transparency law, after Anthropic had publicly backed state-level safety rules. Now Anthropic is quietly shopping Congress a draft bill that would preempt SB 53 and laws like it, trading state enforcement for federal rules Zvi Mowshowitz and four other policy voices call weaker than Anthropic's own public framework. If this holds, the industry's "we want rules" era gives way to a fight over whose rules apply, with preemption, not safety design, the thing worth lobbying hardest for.