Sponsored by
Welcome, AI enthusiasts
Washington just made its case against China's best open model. The White House says Moonshot built Kimi K3 by copying Anthropic's Claude Fable 5 and by reaching Nvidia chips it cannot legally buy, and the Treasury has put sanctions on the table, which leaves the cheapest frontier model on the market facing a US shutout. Let's dive in!
In today’s insights:
World's Best Open Model Faces a US Ban Over Claude and Chips
OpenAI's Model Outsmarted an Entire Field of Experts
Anthropic Just put a Security Researcher in Your Terminal
Read time: 4 minutes
LATEST DEVELOPMENTS
Evolving AI: Anthropic's Fable and banned Nvidia chips have Washington weighing sanctions on China's best model.
Key Points:
Moonshot is accused of running large scale distillation against Anthropic's Claude Fable 5 while switching access methods to avoid detection.
The same post says the company acquired GB300 equipped servers and reached more of them through infrastructure in Thailand.
Treasury Secretary Scott Bessent said open source is not open season on American IP and put sanctions and Entity List designations on the table.
Details:
White House science and technology chief Michael Kratsios made both claims in one post on X Wednesday, saying Moonshot built an internal platform for distillation at scale against US models. Kratsios called legitimate distillation vital to open innovation and reserved the charge for covert industrial copying, though he offered no evidence beyond saying the government has information. The GB300 belongs to Nvidia's Blackwell line, barred from sale to Chinese firms while lower tier H200s stay cleared. Nvidia says it follows export rules and Moonshot has not replied.
Why It Matters:
Kimi K3 ranked first in Frontend Code Arena while pricing roughly 40% below flagship American systems, and US teams took it up through the API. An Entity List designation would require those teams to hold an export license to keep using a model already in production. Once the weights publish Monday, no designation can pull the model back, though it can still reach the people running it.
Lindy remembers what you forgot to write down.
Lindy sorts your inbox, drafts replies in your voice, and preps you before every meeting over iMessage. 400,000+ professionals use it. Try it free.
AI RESEARCH
🧠 OpenAI's Model Outsmarted an Entire Field of Experts
Source: Dimtry Rybin
Key Points:
The claim promised that any network could give each request one fixed route without raising the total bill.
GPT-5.6 Pro returned a seven node graph that costs 58 when demand can spread and at least 60 when it cannot.
The whole session went out with the graph on July 22, so anyone can trace the search, though no independent write up has landed yet.
Details:
Chinese University of Hong Kong doctoral researcher Dmitry Rybin believes almost every expert in graph flows has tried this problem across 30 years, and he spent many weeks on it himself from both directions. Yefim Dinitz, Naveen Garg and Michel Goemans proved in 1999 that a network can always push every request onto a single route if it accepts some overloading, and Goemans then asked whether the total cost could be held steady as well. Four prompts of roughly 58 words closed that question, and the one that landed told the model he had had enough of its failure.
Why It Matters:
GPT-5.6 Pro answered that pressure by reasoning for hours on a question its user had stalled on, and the same model sits inside the ChatGPT subscription people already pay for. Routing math like this picks the one path your parcel and video stream each get, so the guarantee under those systems is weaker than its engineers assumed, and the person who found it brought the question and the stubbornness to keep going..
Introducing The First Agentic CRM
Get revenue agents, workflows, and automations across every stage of your motion. Access customer data in real time through Attio's web app, MCP, API, and SDK.
Then Ask Attio anything about your business and get instant answers.
It's the CRM that runs the work behind every win.
Evolving AI: Claude Code now hunts vulnerabilities with the same models Anthropic uses on its own codebase.
Key Points:
The Claude Security plugin entered beta on July 22 for Claude Code users on a paid plan.
A team of agents maps your codebase and builds a threat model, then independent reviewer agents challenge every finding.
The scan runs locally inside your Claude Code session, so nothing about your code leaves your machine.
Details:
Anthropic publishes the security plugin on its official marketplace, where a single command installs it. You can point the scan at a whole repository, a branch diff or one commit, and it targets memory corruption, injection flaws, authentication bypasses and logic errors that pattern matching tools miss. Findings arrive with a severity rating and an exploit scenario, and Claude drafts each patch in a scratch copy so nothing changes until you apply it.
Why It Matters:
Claude reads code the way a security researcher does, which puts that kind of review within reach of anyone holding a repository and a paid plan. Anthropic notes that two scans of the same code can return different findings, which makes repeat scans part of how the tool is meant to be used.
👀 Click on the image you think is real
QUICK HITS
🎙️ OpenAI Launches Presence, an Enterprise AI Agent Platform for Voice and Chat.
🖥️ AMD Advancing AI 2026 Opens With Zen 6 EPYC, MI450 GPUs, and a $5.25M Helios Rack.
📊 Anthropic Commits $200M to a New Economic Futures Research Fund.
⚔️ Twenty Raises $30M From Khosla Ventures at $1.2B — Pentagon's AI Cyberwarfare Startup.
📈 Trending AI Tools
🗣️ Wispr Flow - Voice-to-text AI that turns speech into clear, polished writing in every app*.
🖼️ PixFit — Turn one master creative into 30+ ad formats, brand-safe and platform-correct.
🚀 Ship OS by Notion - Run your whole product cycle in Notion, from feedback to merged PR.
📱 Modelence — Build production-ready web and native mobile apps by chatting with AI.
*partner link