░▒▓▓▒░ ░▒▓▓▒░ ░▒▓▓▒░ ░▒▓▓▒░ ░▒▓▓▒░
╔════════════════════════════════════════╗
║ THE UNCERTAIN FUTURES ║
╚════════════════════════════════════════╝
No. 2 · Thursday, June 11, 2026
all the news that fits the context windowWelcome back. Today is fine-print day: the industry spent Wednesday reading the restrictions Anthropic built into its new model, and Anthropic's own CEO asked governments for the power to block frontier models outright. Wall Street, meanwhile, wrapped the compute buildout in another $35 billion of financing.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Fable 5's hidden throttle: Anthropic deliberately degrades its model on frontier-AI R&D tasks — without telling the user (update)
A day after launch, the sharpest controversy sits in the system card: Fable 5 carries interventions that degrade its output on frontier-model development work — pretraining pipelines, distributed training infrastructure, accelerator design — with no notice to the user and no fallback, applied through prompt modification, steering, and fine-tuning. Anthropic says the measure touches about 0.03% of traffic and exists to keep its research out of competing models; researchers pushed back hard — Prime Intellect's Will Brown called Fable "the first publicly available model I am explicitly not allowed to use for my work," former Anthropic researcher Behnam Neyshabur called capability concentration "net negative for humanity," and biologists reported fallbacks to Opus 4.8 on words as benign as "mitochondria." A separate cliff arrives June 23, when Fable leaves subscription plans for usage credits. Covered by AI Daily Brief, The Neuron, TLDR AI, Axios AI+, TechCrunch, and WIRED. Primary: Anthropic · Fortune
Anthropic's CEO asks Washington for the power to block models like his
The morning after Fable 5 shipped, Dario Amodei published "Policy on the AI Exponential" — and retired Anthropic's three-year position that disclosure should come before binding regulation. The essay calls for mandatory third-party testing of frontier models above a compute threshold in four risk areas — cybersecurity, biological weapons, loss of control, and automated AI R&D — with government power to "block or deter deployment" of a model that fails, and Anthropic says it will put "substantial financial backing" behind a matching legislative proposal. On work, Amodei floats wage insurance, retention tax incentives, and — if displacement endures — universal basic income or universal capital accounts; he also wants fully autonomous weapons banned from domestic law enforcement and the data-broker surveillance loophole closed. Published after yesterday's newsletters went out — added by your editors. Primary: darioamodei.com · Bloomberg
A Munich court rules AI Overviews are Google's own words — and Google is liable for them
The Regional Court of Munich issued a preliminary injunction, reported this week by The Decoder, after AI Overviews falsely tied two publishers to scams with claims "not even made in the search results." The court held that an AI answer is Google's own content rather than a list of links, so the liability protections search engines enjoy do not apply — and it rejected Google's check-the-sources defense: a casual reader has no reason to verify an AI answer against the underlying links, and the option to research further does not erase the liability. It is the first injunction of its kind, and it reads as a template for claims against any AI answer engine, in Germany and possibly beyond. Covered by The Information. Primary: The Decoder
Apollo and Blackstone put $35 billion behind Broadcom's AI buildout — Anthropic is first in line
Broadcom launched an "AI XPV Platform" with an initial $35 billion tranche led by Apollo alongside Blackstone, to finance more than 20 gigawatts of data centers running Broadcom-designed chips through 2028. The first project funds Anthropic's expansion of more than 1 gigawatt at partner Fluidstack — where Bloomberg reports Google is backstopping lease payments at five sites — and The Information reported in May that Broadcom and OpenAI were negotiating up to $18 billion in vendor financing tied to OpenAI's first custom chip, code-named Jalapeño. The buildout's bills are increasingly Wall Street's. Covered by The Information, TLDR AI, and Axios AI+. Primary: Apollo · The Information
India freezes Starlink clearances over its Iran-war role — two days before SpaceX prices its IPO (update)
Security agencies under India's home ministry have withheld the final approvals Starlink needs to launch commercially, citing the service's use in the Iran war — terminals operate there without a license, the WSJ reported in February that the US smuggled thousands in, and the US military runs operations on SpaceX systems. New Delhi's stated worry is a US company controlling domestic communications in a crisis; the freeze also stalls India's satellite-spectrum pricing framework. The timing pinches: SpaceX targets a Friday Nasdaq debut at a valuation near $1.75 trillion, and Starlink says talks with India remain "active and productive." Covered by TechCrunch and The Information AM, citing Bloomberg. Primary: Bloomberg
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░░░▒▒▓▓ TECH ▓▓▒▒░░░░▒▒▓▓ POLICY ▓▓▒▒░░░░▒▒▓▓ INFRA ▓▓▒▒░░░░▒▒▓▓ BUSINESS ▓▓▒▒░░░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░░░▒▒▓▓ RESEARCH ▓▓▒▒░░░░▒▒▓▓ REPORTS ▓▓▒▒░░Perplexity × Harvard Business School — How AI Agents Reshape Knowledge Work (June 8). Ten thousand identical queries run through both an agent and classic search: the agent performs 26 minutes of machine work per session versus 33 seconds, cuts full task time from 269 to 36 minutes, and shifts what people ask for — half of agent queries create something new (versus 26% for search) and 59% sit outside the user's own field (versus 50%). The time savings get the attention; the shift in what people attempt is the bigger finding. Report page
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Microsoft president Brad Smith answered graduates booing AI at commencements with a 3,000-word essay and a GeekWire interview: the boos are "telling us what we need to hear," AI will reshape work rather than end it, and "if the world's people don't have jobs, then neither do we" (GeekWire). Critics note the essay changes no policy. Takeaway: validating the heckler buys a hearing, but only until the audience checks what changed — pair acknowledgment with one concrete commitment or skip the essay.
Walmart used its Associates Week to tell 2.1 million employees AI will "power" their jobs, not replace them — "Technology will power our future. But our associates will lead it," said chief people officer Donna Morris — while executives separately guide to flat headcount as roles change (Storyboard18, via POLITICO). Takeaway: "no job losses" and "flat headcount while every job changes" are different promises; draft the one you can still defend in two years.
SparkToro finds under a third of Google searches now end in a click on the open web, while Posthog reports LLM-referred traffic grew 41x in two years and converts better than nearly every other channel (SparkToro · Posthog, via TLDR Founders). Takeaway: site traffic now undercounts your actual reach — add AI-engine presence to the measurement stack before the next reporting cycle.
░░▒▒▓▓ CORRECTIONS ▓▓▒▒░░░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Hiroki Tomiyasu, a self-taught farmer in Hokkaido, runs 100 hectares of broccoli, soybeans, and pumpkins with software he built himself using ChatGPT and Codex: satellite crop monitoring, a pesticide-log database, and a greenhouse whose vents he raises by text message. He describes the setup as an always-available engineer — which makes his farm better staffed, per capita, than most startups (profile, via The Rundown).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Movement 10 - Her Gift — Floating Points & Miriam Adefris
Anpfiff — your editors at The Uncertain Futures