The US government bans Claude
🐯 Before we get started
Right now a data center is being planned within shouting distance from the Nashville Zoo. While water and energy concerns around data centers are typically overblown, their sound pollution and growing CO2 emissions directly affect our world and animals, making construction near a zoo a nonstarter.
I urge you to sign the petition to prevent the Nashville Zoo build, and keep an eye out for unwise construction plans in your local community (data center or otherwise).
what to know for now
📖 Anthropic shipped Claude Fable 5, the first “Mythos-class” model. Fable 5 (claude-fable-5) is Anthropic’s most capable public model to date, claiming state-of-the-art results across software engineering, knowledge work, vision, and long-horizon agentic tasks, with a 1M-token context window and a reported 80.3% on SWE-Bench Pro against 69% for Opus 4.8. The catch is the safety architecture: in high-risk domains like cyber, bio, and chemistry it refuses and silently falls back to Opus 4.8, a downgrade Anthropic says fires in under 5% of sessions. Pricing lands at $10/$50 per million tokens, roughly double Opus.
⛔ Then the government made Anthropic pull it 72 hours later. Here’s the saga in one breath: Fable 5 launched June 9; by June 10 developers caught it quietly nerfing itself when it sensed frontier-AI work, and Microsoft yanked it from internal Copilot over data-retention terms; on June 12 a red-teamer handed the government a jailbreak, and at 5:21 PM ET Anthropic got an export-control directive ordering it to cut off all foreign nationals (including its own foreign-national staff), so it disabled Fable 5 and Mythos 5 worldwide. Anthropic publicly disagrees, calling the issue a narrow, non-universal jailbreak (essentially “ask the model to find bugs in code,” a thing GPT-5.5 does too) and warning that enforcing that standard would “essentially halt all new model deployments.” As of today, both models are still dark and Anthropic says it’s “working to restore” access. Read more
🍎 Apple gave up and handed Siri to Google. At WWDC on June 8 Apple unveiled “Siri AI,” the biggest overhaul since launch, and confirmed the long-rumored Gemini partnership: a reported ~$1B/year deal to run a custom Apple-tuned Gemini model (”AFM Cloud Pro”) on Google’s cloud for the heavy reasoning its on-device and Private Cloud Compute tiers can’t handle, shipping this fall in iOS 27. Developers got the real gift: a hybrid Foundation Models framework with a new LanguageModel Swift protocol that lets apps call Claude, Gemini, or Apple’s own models through one API with no code changes, plus free on-device model access for apps under 2M downloads. Read more
🛰️ SpaceX went public, and the market is being asked to price it as an AI company. The IPO priced after the close June 11 and started trading June 12 on Nasdaq under SPCX, raising roughly $75B at around a $1.75 trillion valuation (Bloomberg reported the target crept above $2T). It’s the whole company, not a Starlink spinout, pitched across three legs (space, connectivity, and AI) with Starlink throwing off two-thirds of revenue and the speculative upside resting on Starship and the orbital data centers SpaceX has filed to build (up to a million compute satellites, 1 GW of orbital compute by year-end). Recall that last month Anthropic rented all 220,000 GPUs in SpaceX’s Colossus 1. Read more
🌙 Moonshot dropped Kimi K2.7 Code and kept the open-weight pressure on. Released June 12 on Hugging Face under a modified-MIT license, K2.7 Code is a 1T-parameter MoE (32B active) with a 256K context, claiming double-digit gains over K2.6 on its own coding benchmarks and roughly 30% fewer reasoning tokens per task, at $0.95/$4.00 per million tokens. The asterisk: every number is from Moonshot’s proprietary benchmarks, with no SWE-bench Verified score at launch, so treat “it’s great” as a vendor claim until third parties run it. (I did a full Model Drop on this one if you want the deeper read.) The strategy is the story regardless: cheap, downloadable, commercially usable agentic coding aimed straight at the closed labs’ wallets.
🧪 AI Research of the Week
REDMOD: detecting pancreatic cancer on routine CT scans years early
From Mayo Clinic
Jake’s Take: Mayo built REDMOD, a model that reads the routine abdominal CT scans people already get for unrelated reasons and pulls out hundreds of “radiomics” features, subtle texture and structure changes in pancreatic tissue that a radiologist can’t catch by eye. In the validation study it flagged 73% of pancreatic cancers before clinical diagnosis, against 39% for radiologists reading the exact same scans, at a median of about 16 months early and up to three years out in some cases. They trained it on roughly 2,000 scans across different hospitals and scanner types so it isn’t just memorizing one machine’s quirks.
Pancreatic cancer’s five-year survival sits around 13% because it’s almost always caught late, and this is AI detection available on the imaging patients are already getting. No new scan, no new cost, no manual annotation. It’s not FDA-cleared and it’s not deployed yet (the follow-up AI-PACED trial is now testing it prospectively in higher-risk patients). It’s a landmark study, not a product. But a 16-month head start on the cancer that kills fastest is nothing but good news. Read more
what to know for later
⚖️ Dario Amodei reversed himself and called for binding AI regulation, two days before the government banned his model. In a ~5,000-word essay, “Policy on the AI Exponential” (June 10), Amodei argued it’s time to go “beyond transparency to more serious and binding regulation”: mandatory third-party testing for models above a compute threshold and federal power to block or reverse deployments that fail audits across cyber, bioweapons, loss-of-control, and autonomous-R&D risk, an FAA/FDA-style regime. Anthropic paired it with a $350M commitment to study AI’s labor-market hit. The timing is almost too perfect — Amodei asks for a government empowered to pull dangerous models, and 48 hours later that government pulls his; Sam Altman called the whole thing “fear-based marketing.” Read more