US Government Further Escalates Rhetoric Against Chinese Open Source Models

White House accuses Moonshot of stealing Anthropic's Fable for Kimi K3

Michael Kratsios, head of the White House Office of Science and Technology Policy, has publicly accused the Chinese company Moonshot AI of scraping Anthropic's top model Fable via distillation to develop its new K3 model. On X, Kratsios wrote that Moonshot built a sophisticated internal platform for this purpose, to distill against US models at a large scale while rapidly switching between multiple access paths to evade detection. Distillation, the training of a smaller model on the outputs of a larger one, is a legitimate optimization method and part of the open innovation ecosystem; however, large-scale, covert industrial distillation to steal protected US technology is unacceptable.

U.S. Treasury Secretary Scott Bessent followed up in a separate post, stating that sanctions and an Entity List designation remain on the table if the allegation is confirmed. “Open source is not an open season on American intellectual property,” he wrote. Kratsios also accused Moonshot of acquiring servers with Nvidia's GB300 chips and using them in Thailand, presumably for training its models. The GB300 is part of Nvidia's Blackwell generation, the sale of which to Chinese companies is prohibited.

Moonshot had introduced Kimi K3 last week, a model with 2.8 trillion parameters, which the company describes as the world's largest open-weight system and whose performance comes close to Anthropic's frontier model Fable. The full weights are set to be freely available for download next week. Back in February, Anthropic had stated that DeepSeek, Moonshot, and MiniMax had generated more than 16 million interactions with Claude through about 24,000 fake accounts, violating the terms of use. Anthropic's head of public policy, Sarah Heck, welcomed Kratsios's move, speaking of industrial espionage that supports adversarial military and intelligence capabilities.

The Chinese embassy in Washington called the accusations “completely baseless” and emphasized that China respects the protection of intellectual property. Some experts doubt that K3 could have been predominantly created through distillation from Fable, which has only been publicly available since July 1. The episode is fueling the debate in Washington about the influx of Chinese open-weight models, which figures like OpenAI's Dean Ball would prefer to see restricted or banned entirely. → reuters, businessinsider, techcrunch

Synthszr Take: The timeline is the real problem with this accusation. Fable has been public since July 1, K3 came out a week later, and no one distills a 2.8-trillion-parameter model to a frontier level in seven days. Distillation doesn't leave a serial number: A model trained on the outputs of another ends up looking like a model that independently learned to give similar answers. The only semi-hard evidence is the 24,000 fake accounts and 16 million Claude interactions from February, and even that proves access, not causally how much of it ended up in K3. Proximity on benchmarks is hardly proof either, as models naturally converge on the same public tests. If sanctions depend on whether distillation is “proven,” then they hinge on a question that is technically almost impossible to answer cleanly. Bessent will have to deliver what a court would accept as evidence, and the GB300 servers in Thailand are a more solid lever for that than any speculative analysis of the model weights.

New Open-Weight US Model Beats Rivals Ten Times Its Size

The San Francisco lab Poolside released its most powerful coding model to date on Tuesday: Laguna S 2.1, a Mixture-of-Experts system with 118 billion parameters, of which only 8 billion become active per token. The weights are immediately available on Hugging Face under the permissive OpenMDW-1.1 license, and the context window extends up to one million tokens. According to the company's published benchmarks, the model achieves a score of 70.2 percent on Terminal-Bench 2.1, placing it ahead of DeepSeek-V4-Pro-Max (1.6 trillion parameters, 64.0), Thinking Machines' Inkling (975 billion, 63.8), and Nvidia's Nemotron 3 Ultra (550 billion, 56.4). On SWE-Bench Multilingual, Poolside reports 78.5 percent, and on the public SWE-Bench-Pro dataset, 59.4 percent. → Techpresso

Synthszr Take: Poolside can't compete with the investment budgets of the hyperscalers, so the lab chooses the field where capital isn't the decisive variable. Open weights, self-hosting on a single DGX Spark, 8 active instead of 118 billion parameters per token: these are the levers that matter to a customer who needs to run their model behind their own security perimeter because metered API access is simply not an option for a defense agency. Radical openness here is sales logic, not idealism. Any government that commits to a Chinese open model today will be harder to win back tomorrow, and Poolside is filling this exact gap with a model that compresses the industry's usual quarterly cycles into nine weeks.

Alphabet Quadruples Quarterly Profit to $112 Billion, Cloud Grows by 82 Percent

Alphabet reported revenue of $119.8 billion for the quarter ending in June, a 24 percent increase year-over-year, beating Wall Street expectations of $116.5 billion. Net income rose to $112.1 billion, quadrupling from $28.2 billion a year earlier, with about $77 billion coming from valuation gains in investments like SpaceX (which went public in June) and Anthropic. The cloud business grew by 82 percent to $24.8 billion, according to the company, and the order backlog climbed from $106 billion to $514 billion. At the same time, Google raised its capital expenditure forecast for this year to $195 to $205 billion, more than double the $85 billion from the previous year. → www.nytimes.com

Synthszr Take: The one number that settles the capex debate for investors is the cloud backlog: from $106 billion to $514 billion in twelve months. These are signed contracts, more than half of which will be fulfilled in the next two years. Anyone who previously dismissed the data center billions as a PowerPoint illusion must now compare this order intake with the $195 to $205 billion in capex and will find that capacity is being sold faster than Google can build it. Still, caution is advised: $77 billion of the $112 billion profit are paper gains from the SpaceX IPO and the Anthropic stake—valuation effects that could swing in the other direction next quarter. But the operational foundation is solid even without these special effects, with 82 percent cloud growth and 14.4 percent in the advertising business, which many had already written off because of AI fashion.

Google Releases Three New Gemini Flash Models, but Flagship 3.5 Pro Is Missing

Google DeepMind has released three new models: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. The 3.6 Flash is considered the workhorse; it's said to improve in coding, knowledge work, and multimodality while consuming up to 17 percent fewer tokens, making it cheaper than its predecessor. Flash-Lite is the most affordable model in its class; Flash Cyber is a specialized model for detecting and fixing security vulnerabilities, which Google says will initially be available only to governments and select partners in a limited pilot program. Notably absent is the long-awaited update to the flagship model Gemini Pro, which was last updated in February. → AI Secret

Synthszr Take: The three Flash models are secondary; the real signal is the gap where 3.5 Pro should be. Google promised the Pro update in May for June, it's now July, and Bloomberg is reporting missed internal targets. In the same period, OpenAI has rolled out GPT-5.5 and 5.6; Anthropic has opened access to Opus 4.8, Sonnet 5, and Fable 5. Kilpatrick's passing remark that they have begun the most ambitious pre-training run for Gemini 4 is a more interesting announcement than any Flash benchmark: it sounds like Google might skip the stalled Pro and focus its energy on the next generation. That's a risky bet, because the flagship model sets the market perception, while the mass of agents has long been running on cheap Flash.

AI Secret Declares Hermes Agent the Winner in Endurance Duel Against OpenClaw

In its own scale tests, AI Secret ran the agent frameworks OpenClaw and Hermes Agent against each other and favored Hermes. According to the report, a clear difference emerges in long tasks: OpenClaw often completes a sub-step and then asks a follow-up question or waits for a new prompt, while Hermes continues to work autonomously through its “Persistent Goals.” OpenClaw remembers the goal but doesn't keep the work in motion. To keep OpenClaw running, according to AI Secret, one has to intervene with additional functions like “detached tasks” or TaskFlow. → AI Secret

Synthszr Take: Outside the bubble, nobody knows OpenClaw or Hermes, and that's perfectly fine. The framework duel tests the only criterion that decides adoption in real organizations: Does the agent run through, or does a human have to type “continue” every three steps? This is the transition from demo to operation. An agent that stops and asks questions at every sub-task has only automated the asking of questions. This is exactly where most pilots fail: with autonomy that turns out to be just configuration.

Claude Code Now Launches and Tests iOS Apps Itself in the Simulator

Anthropic has expanded the Claude Code Mac app with an interactive simulator area where iOS apps can be built, launched, and tested directly. According to 9to5Mac (picked up by MacRumors), the feature is available in a public beta starting with version 1.24012.0 for users on Pro, Max, and Team plans. Claude installs the app, taps through the interface, and reads the screen to verify its own changes, while the user can watch in parallel and intervene with their own taps and swipes. The prerequisite is a Mac with Xcode and the iOS platform installed, as Apple's iOS simulator runs exclusively on macOS. → Techpresso

Synthszr Take: The interesting part is the closed loop. Claude builds, launches, taps through the app, and reads the screen to check its own changes. Testing used to be the moment when the developer picked up the phone or the simulator to see if the onboarding flow actually worked. This step now falls into the agent loop, and that changes the economics of experimentation: when building and testing both become almost free, an agent can run through ten variations before a human even looks. What remains scarce is the formulation of what needs to be tested. “Check the onboarding flow” is a precise intent; “make it look nice” is not, and the machine immediately sees the difference in the result.

Claude Cowork Learns New Tasks by Recording Your Screen

Anthropic has given Claude Cowork an update that allows the assistant to learn new skills via screen recording. Instead of typing out instructions step-by-step, you click “Record a skill” in the desktop app's plus menu, perform a task once on the screen, and Cowork remembers the process as a reusable skill. According to the announcement, this enhances the level of customization Cowork is capable of and is aimed at multi-step workflows that were previously considered too complex. → Superhuman – Zain Kahn

Synthszr Take: Clicking “Record a skill” is the most convenient way to hand your entire work context over to a model. What happens on the screen while you demonstrate a task is rarely cleanly isolated: the open CRM with customer names, the email in the background, the password field that flashes briefly. Convenience always has a price, and here the price is: Your screen becomes a training dataset without you ever consciously uploading data. This is the silent deal of the new generation of assistants: You save yourself the trouble of typing instructions and pay with visibility. Anyone who wants to use this feature productively should first clarify what Anthropic does with the recordings, how long they are stored, and whether they are used for model training (the answer is in the enterprise terms, not the marketing).

Block Releases Hybrid Workspace for Humans and AI Agents

Block has released Buzz, a free, open-source workspace for mixed teams of humans and AI agents. According to the announcement, the agents are full-fledged members with their own accounts and permissions, not chatbots at the end of a prompt: they post messages, review code, and initiate automations just like a human colleague. Technically, Buzz is built on the decentralized protocol Nostr, which gives each agent a cryptographic key pair; a second signature binds the agent to its human owner, creating a verifiable audit trail for every action. The platform is model- and agent-agnostic and works with Anthropic's Claude Code, OpenAI's Codex, and Block's own goose framework via the Agent Client Protocol. Git hosting is tightly integrated: feature branches become channels, and reviews and CI results land in the same thread as the discussion about them. → Techpresso

Synthszr Take: What's interesting about Buzz is that Block isn't even trying to build the best model. Claude Code, Codex, goose—everything plugs in. The bet is on the environment where humans and agents work side-by-side, and the real engineering achievement is in identity: a cryptographic key pair per agent, a second signature tying it to a human owner, a proof that neither human nor agent can forge alone. This hits the exact point many overlook when deploying agents. An agent without an anchored owner becomes an orphan, and without an audit trail, any compliance discussion becomes a shot in the dark.

Habermas Machine: AI Synthesis Beats Lay Mediators, but Only in This Order

In the experiments with the so-called Habermas Machine, participants in small groups first formulated their own standpoints before a language model synthesized a common statement from them. They then read and criticized this draft, and the model revised it based on their objections. According to Nate from Nate's Substack, the participants preferred the result over statements written by non-professional human mediators. The crucial part of the setup is the sequence: the humans formed their opinions before the machine even spoke. They then got a second opportunity to challenge the machine's output instead of just accepting it.

Synthszr Take: The sequence is the entire invention here. First your own standpoint, then the synthesis, then the objection: in this sequence, the human remains the framer and the machine the executor, never the other way around. Reverse the sequence, let the AI spit out a consensus proposal first and have the group build on that, and you get priming instead of deliberation. Then people are just negotiating along the machine's text, and their own position dies before it's even spoken. It's the same geometry as with intent: the human forms the decision before it exists, the machine condenses it afterward, and the human has the veto a second time.

American Prenuptial Agreements Get a Clause Against AI Infidelity

In the US, more and more couples are including a so-called “AI Infidelity Clause” in their prenuptial agreements, which defines a romantic or sexual relationship with an AI chatbot as infidelity. Julia Rogers, CEO of the online platform Hello Prenup, reports recurring customer requests for exactly this, according to Business Insider; in sample agreements, sustained exchanges, sexual messages, role-playing, and exclusive expressions of feelings towards AI avatars are considered a violation. New York divorce attorney Lisa Zeiderman describes cases where a partner spends more time with an AI than with their family. A survey by the Kinsey Institute at Indiana University found that 61 percent of unmarried adults classify falling in love or sexting with an AI chatbot as cheating. → Trendium.ai

Synthszr Take: What's remarkable here is the order of events. The law is writing a rule for a behavior that society doesn't even have a name for yet. 61 percent call it infidelity, even though no one can say where conversation ends and a relationship begins, and this very ambiguity shows that the contract is outpacing the definition. Lawyers are encoding an emotional state into clauses before sociologists or ethicists have provided a solid category for it: the clause is a makeshift solution for a feeling without a dictionary entry. That is the real news: the moment when a private contractual practice begins to normalize the relationship between humans and machines.