Washington accuses Moonshot of stealing Fable, and the timeline doesn't add up

White House Office of Science and Technology Policy director Michael Kratsios said Wednesday the U.S. has “information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model,” and that the company built “a sophisticated internal platform to conduct large-scale distillation against U.S. models,” switching between access methods to dodge detection. He also claimed Moonshot obtained Nvidia GB300 chips through servers in Thailand, a possible end-run around export controls. Treasury Secretary Scott Bessent floated sanctions on any foreign company found to have stolen American AI technology.

The timeline doesn’t cooperate with the accusation. Fable has been publicly available since July 1st; K3 shipped less than two weeks later. Snorkel AI co-founder Braden Hancock put it plainly: “You can’t distill that much data, train a model, and release it in two weeks.” Gergely Orosz took the framing apart from a different angle, comparing model inspection via prompting to a carmaker tearing down a competitor’s vehicle: legal when Ford did it to a Tesla, apparently not legal when an AI lab does it to another AI lab’s outputs. Schmidhuber pointed out distillation has a long research lineage predating any single company’s claim on it, and Suhail argued the practical response to the whole mess isn’t prohibition but faster investment in open-weight domestic models. The accusation lands during the same week Hugging Face released The Stack v3, the largest open code dataset yet published: 114 TB raw, 224 million repositories, 44 billion files, 770 languages, and roughly 5 trillion deduplicated tokens after filtering, up from around 550 billion in v2. Whatever Washington decides about Moonshot, the open ecosystem it’s trying to fence off just got a lot more raw material to train on.

Black Forest Labs used the same week to ship FLUX 3, a single architecture trained jointly across image, video, audio, and action prediction rather than a family of specialized generators bolted together. The company’s early claims put it ahead of Seedance 2.0, Gemini Omni, and Grok Imagine on style range and typography generation, with an open-weights Dev version coming. The more interesting signal is FLUX-mimic: an early version of FLUX 3 already running on robots through a partnership with Mimic Robotics, combining the FLUX 3 backbone with Mimic’s dexterous-manipulation training data. If a video generation model’s world model transfers directly into robot control quality, as Mimic is claiming, image and video generation stop being a separate product category from robotics and start being the same research bet.

Stripe is in talks to buy OpenRouter for around $10 billion, according to the Wall Street Journal. OpenRouter’s pitch since launch has been the model-agnostic routing layer sitting between developers and every frontier API; a payments company acquiring it would fold inference routing into the same rail that already handles checkout for a huge share of the internet’s software. Read against Cursor’s own router launch last week and Replit’s move to cut autoscale deployment costs 80% by passing through volume discounts from cloud providers, model routing and inference-cost arbitrage look like the infrastructure layer everyone wants to own before the market decides it’s commoditized.

Back on the ground, Gergely Orosz’s Pragmatic Engineer flagged a trend that’s been building since January, when Opus 4.5 and GPT 5.4 started writing enough good code that engineering leaders watched the bottleneck shift from writing software to reviewing it. Since February there’s been a real boom in dedicated AI review tools (CodeRabbit, Greptile, Qodo, SonarQube’s new Gitar product), alongside review features bolted onto the harnesses themselves: Claude Code review, Cursor review, GitHub Copilot review. Uber built its own in-house tool, Code Inbox, with risk profiles that estimate the blast radius of a change and smart assignment to route reviews to the right person; Cloudflare, Faire, and HubSpot built similar internal tools after deciding vendor integrations didn’t fit. The uncomfortable part of Orosz’s reporting: engineers are starting to rubber-stamp PRs when the AI reviewer has no comments, while the reviewers still doing careful work feel buried under a rising volume of AI-generated changes. Nobody in his reporting has a proven fix; verification-by-testing is the theoretical answer, but nobody agrees on how much testing counts as thorough enough to replace a human reading the diff.

Grab’s engineering team published the origin story of LLM-Kit, the internal framework now running over 500 production agent services and 50-plus registered MCP servers behind a single LLM gateway handling billions of tokens a month. It started as one support bot for their infrastructure team in 2023, running on GPT-4-32k, and the post is candid about what actually broke as it scaled: no real evaluation strategy beyond “ship it and hope,” provider switching that required rewriting error handling, and debugging that meant grepping logs across three disconnected systems because there was no shared trace. Every fix became a framework component: a template that generates a working FastAPI service with evals wired in on day one, a gateway that turns “switch providers” into a config change, and OpenTelemetry instrumentation that ties agent, tool, and model call into one trace by default. The reasoning loop for their first bot took an afternoon to write; the production wrapper around it took two weeks. That ratio is the whole argument for building the framework in the first place.

GitHub gave Dependabot a three-day default cooldown before it opens pull requests for non-security version bumps, a direct response to the pattern behind last September’s chalk and debug npm compromise, where booby-trapped packages downloaded more than two billion times a week sat live for about two hours before the community caught them. GitHub’s own advisory data shows roughly 18 newly cataloged malicious npm packages a day over the past year, and a review of 21 major supply chain incidents since 2018 found malicious versions of axios, Solana web3.js, and ua-parser-js were all pulled within hours of publication, meaning a cooldown would have filtered out most of them before they ever reached a build. Security updates still land immediately; this only slows down the automated chase for the newest release, which is exactly the behavior attackers have been exploiting.

Watch what Treasury actually does with the Moonshot accusation next: a formal sanctions action would be the first real test of whether “distillation” becomes an enforceable legal category or stays a rhetorical cudgel, and either answer reshapes how every lab handles model access from here.