Washington Safety'd Fable

The Rundown says the order required Anthropic to remove access for all non-U.S. citizens, including foreign nationals inside the U.S.; Anthropic responded by taking Mythos and Fable offline globally. Tech Brew adds that the order arrived Friday at 5:21 p.m. ET and Anthropic claimed it had about 90 minutes to comply.^{[1]The Rundown[2]Tech Brew}

Simon Willison linked the Axios behind-the-scenes account and highlighted the core unresolved issue: Anthropic says no universal jailbreak has been found, while administration sources appear to want the company and government to feel aligned and secure before access returns.^{[3]Simon Willison}

~00:00 Fireship framed the irony bluntly: a company known for AI safety was restricted by its own government. ~04:00 Low Level argued the precedent matters because cybersecurity research can be done with many models and harnesses, not only Mythos or Fable.^{[4]Fireship[5]Low Level}

~15:07 Theo's longer critique focused on trust: if a frontier model can disappear or its documentation can change under pressure, builders may start treating the provider itself as a supply-chain risk.^{[6]Theo - t3.gg}

~00:00 Theo's short explainer argued Fable 5 and Mythos 5 are effectively two access paths into the same model, with Fable adding heavier checks at the door. His longer video then contrasted strong benchmark behavior with refusal-heavy implementation details and an alleged silent change to Anthropic's system-card language.^{[7]Theo - t3.gg[6]Theo - t3.gg}

~00:00 Nerd Snipe's pre-ban discussion treated the model as genuinely special while still questioning price, access, and platform reliability. The useful signal: even model fans were already separating raw capability from deployability before the ban hardened that distinction.^{[8]Nerd Snipe}

~00:00 Jones noted that Google, Microsoft, Amazon, and Meta are on pace for roughly $700B in AI infrastructure spending and argued the key operating question is whether expensive tokens are doing work valuable enough to justify their cost.^{[9]Nate B Jones}

Sean Goedecke attacked a common bearish assumption: that inference GPUs last only one to three years. He points to GPU cluster survival data where many GPUs remain alive far past year three, making the depreciation story less fatal than critics suggest.^{[10]Sean Goedecke}

~00:00 Jensen Huang's Sequoia clip gives the clean bull case: AI became valuable when it started doing work people pay for. ~00:00 Last Week in AI zoomed out to the valuation problem: companies this large eventually need to grow the economy, not merely capture existing software budgets.^{[11]Sequoia Capital[12]Last Week in AI}

OpenRouter's governance post makes the control point concrete: an agent can retry, escalate models, and spend money unless the API key itself has workflow-specific limits, model allowlists, and request traces. The piece recommends separate keys for separate workflows because shared keys erase attribution and expand blast radius.^{[13]OpenRouter}

Its free-API comparison lists 13 usable free API options, including OpenRouter, Groq, Cerebras, Google AI Studio, Mistral, NVIDIA, Cloudflare Workers AI, and Hugging Face, while warning that free tiers shift cost into privacy, reliability, speed, or quota constraints.^{[14]OpenRouter}

~00:02 AICodeKing made that practical in OpenCode: NVIDIA's NIM catalog exposes development endpoints for models such as MiniMax M3, Step 3.7 Flash, and Nemotron 3 Ultra, giving builders a cheap way to test coding-agent loops before paying production rates.^{[15]AICodeKing}

Simon Willison's datasette-agent 0.3a0 adds execute_write_sql, an approval-gated write tool that respects user permissions before modifying a database. The CLI also learned approval flows plus --yes and --unsafe modes for direct database-modifying chats.^{[16]Simon Willison}

OpenAI's short June 15 clips showed three practical surfaces: an iOS build/test workflow without leaving the coding environment, Wayfair enriching a 40M-product catalog, and a solutions engineer using a coding agent to analyze Trustpilot reviews, mock up website changes, and save repeatable skills.^{[17]OpenAI[18]OpenAI[19]OpenAI}

~00:00 Real Python gave the compact definition: an agent is a model plus tools in a loop, and software tasks work well when framed with tests and room for self-correction. ~00:00 Logan Kilpatrick said Google's Gemini through-line is becoming an agentic product layer, powered by an agent harness across Google products.^{[20]Real Python[21]Sequoia Capital}

~00:07 Barthelet framed ChatGPT/MCP apps as a new acquisition surface for businesses: apps can be discovered in stores or summoned in chat when relevant. The implementation problem is that the embedded UI must be interactive without giving arbitrary third-party app code access to the parent DOM, cookies, or local storage.^{[22]AI Engineer}

~10:14 He walked through why a simple iframe srcdoc or direct app endpoint breaks down: one can inherit the parent's origin, while the other would require a CSP allowlist for every app. The double-iframe mechanism isolates app content while keeping the platform scalable for many app developers.^{[22]AI Engineer}

Import AI's lead item says Sequent will fund under-resourced alignment bets because current empirical lab programs may not deliver a priori confidence before ASI training. The same issue covers FrontierCode, a benchmark built from maintainer-selected multi-PR chains and freeform requests, graded for correctness, test quality, scope discipline, style, and mergeability.^{[23]Import AI}

The Hugging Face June 15 page featured agent-heavy papers: APPO for agentic procedural policy optimization, graph memory for LLM agents, a persistent autonomous AI paradigm, Orchestra-o1 for omnimodal orchestration, and HarnessX for composable adaptive agent harnesses.^{[24]Hugging Face}

~10:04 Nate Herk's key middle skill was context engineering: AI output quality depends on private context, subject-matter expertise, and iteration speed, not just prompt templates. He also argued the durable career move is becoming the person inside a team who can map AI tools to real workflows.^{[25]Nate Herk}

~00:00 GitHub Awesome's roundup emphasized self-hosted alternatives for tasks, wikis, language learning, notes, dashboards, backups, and certificates. ~00:00 Kage narrowed that instinct to webpages: strip JavaScript and save permanent offline copies of pages worth keeping.^{[26]GitHub Awesome[27]GitHub Awesome}

~00:00 The marimo notebook demo used pi digit search to turn a curiosity into an interactive probability lesson, ending with the idea that pi can be treated like a random seed. Simon's collected Julia Evans quote made the writing rule concrete: picture one specific reader, often your past self.^{[28]marimo[16]Simon Willison}

~00:00 Groww's founder described reading between the lines of customer requests: users did not ask for transparent product comparison and frictionless payments in those words, but those were the jobs to solve. ~15:13 The company later moved into zero-commission direct mutual funds even though it meant being a zero-revenue company for years, betting that organic acquisition and customer love would eventually unlock monetization.^{[29]Y Combinator}

~00:00 Paine's clip used the Qing destruction of the Zunghar and Tibetan empires to explain how Han control over arable territory was consolidated, then connected that history to modern concerns about Uyghurs.^{[30]Dwarkesh Patel}

~00:00 Acquired's clip returned to 2001, when Musk explored buying a post-Soviet intercontinental ballistic missile because U.S. rocket options were prohibitively expensive, a telling origin story for SpaceX's cost obsession.^[31]Acquired

Morning Brew reported that the U.S. and Iran reached a deal to reopen the Strait of Hormuz and end the U.S. blockade of Iranian ports, with a formal signing planned for Friday in Geneva and oil prices falling as traders priced in lower geopolitical risk.^{[32]Morning Brew}

Sherwood's June 15 Snacks issue said Elon Musk had added about $274B to his net worth through the prior Thursday and $351B year to date, pushing him into a paper trillionaire category far ahead of Larry Page, Jeff Bezos, and Bill Gates.^{[33]Sherwood Snacks}

Morning Brew's main daily issue was lighter, led by sports and KPIs, but it rounded out the same business-news mix as the Hormuz story in the feed.^{[34]Morning Brew}