The Anthropic Quake: The Day After

Amazon's CEO Gets Trump to Cut Off Anthropic's Models

Andy Jassy told U.S. government officials that Amazon researchers used a series of prompts to get Anthropic's Claude 5 model to release normally blocked information for cyberattacks. Shortly thereafter, the White House convened, security researchers verified the claim, and the government demanded that Anthropic either close the loopholes or shut down the model. Trump ultimately signed a ban prohibiting foreign governments, companies, and individuals from accessing the tools, despite stated concerns that it would stifle innovation. Anthropic then shut down Claude 5 for all users, and many of its own foreign-born researchers are now barred from working on the latest models. The company counters that the vulnerabilities are relatively trivial and can also be found in other freely available models, a view shared by some security researchers. At the same time, Amazon is a major investor in Anthropic and supplies the chips for its data centers. The move comes as Anthropic prepares for an IPO, possibly as early as this fall. → www.wsj.com

Synthszr Take: The investor who supplies your chips marches into the White House and delivers the diagnosis for your model ban along with it. Amazon holds shares in Anthropic, supplies its data centers, and at the same time uses Andy Jassy's line to Washington to take these very models off the market. This is vertical integration via a detour through security policy, packaged as civic duty. In March, we wrote that OpenAI would get the Pentagon deal while Anthropic landed on the blacklist, and in April, that the government fired its AI security chief after 96 hours. The pattern repeats itself: Whoever tells the better story in Washington decides whose model runs, and David Sacks (a known Anthropic critic and in the Trump camp) admits himself that the ban came “reluctantly.” For a company on the verge of an IPO, a ban on its top models is no small matter, but a loss of oxygen, while OpenAI is already cautiously rolling out its own cyber model to customers. Those who compete on AI through security briefings rather than products shouldn't be surprised when the one with the best lobbying access wins in the end, not the one with the best model.

How the 24-Hour Thriller of Shutting Down Claude 5 Unfolded

The Trump administration has imposed far-reaching export controls on Anthropic after a frantic 24-hour attempt to persuade the company to voluntarily withdraw a newly released AI model failed. According to two government officials and a senior White House representative, officials saw the model as a security risk. This was followed by several tense phone calls between Anthropic CEO Dario Amodei and top government figures, including Treasury Secretary Scott Bessent and White House Cyber Director Sean Cairncross. The details of these conversations have not yet been made public. The case shows how the U.S. government is struggling in real-time to regulate fast-moving and potentially dangerous AI models. For Anthropic, this is the next escalation after the Pentagon dispute in March, when the company was blacklisted while OpenAI got the deal. → www.politico.com

Synthszr Take: 24 hours, a few phone calls, then export controls against a U.S. company that is heading towards nearly $20 billion in revenue. This is what AI regulation looks like in 2026 when it's improvised: no law, no process, just Bessent on the phone and Amodei saying no. In April, we wrote that the U.S. government is fighting against China and its own people at the same time, and that's exactly what is continuing here. The really interesting part is what it reveals about sovereignty: even your home market won't protect you if the government deems your model too dangerous. For European companies, this is a data point, not a triumph. Anyone building agentic workflows on a provider that can be taken off the market with a phone call has an architecture problem, not a supplier problem. On-premise and data sovereignty aren't just compliance folklore; they are the only insurance against a call from Washington that you will never make yourself.

The Wake-up Call for Europe

Anthropic shut down Claude and Claude for all foreign users on Friday evening, following a letter from the Trump administration citing national security concerns. “The net effect of this order is that we must abruptly disable Claude and Claude for all of our customers,” the company writes. A wave of reactions followed across Europe: France's Minister for Europe, Benjamin Haddad, speaks of an accelerator in the geopolitical struggle for AI, while presidential candidate Bruno Retailleau calls it a wake-up call and points to Mistral, OVHcloud, Scaleway, and ChapsVision. British MP Al Carns calculates that British researchers, companies, and even hospitals were using the model just moments ago, and now they can't. Former Prime Minister Édouard Philippe compares AI to critical infrastructure like electricity or the internet, whose computing power Europe does not control. Jordan Bardella and Geert Wilders are unanimously calling for their own models, to be accelerated. → www.euronews.com

Synthszr Take: Retailleau has the diagnosis in one sentence: A nation that sources its technology from others can be shut down overnight. The only remarkable thing is how many tweets it took for someone to say what has been in the Draghi report for over a year, and from which hardly any recommendation has been implemented so far. Back in April, we wrote here about France's switch to Linux and 80 percent of Europeans being mistrustful; the dependency was never a secret, it was just convenient. Sovereignty now depends on compute capacity and energy at competitive costs, and that's exactly where, according to Tugendhat, the UK has applied the brakes instead of the accelerator. Outrage on X costs nothing, compute costs everything. If you're serious about Mistral, you provide capital, power, and data centers before the next letter from Washington arrives, not after the summit after next. That can be decided this week, and the bill for it has been on the table for months.

Meta's 'Applied AI' Unit Is a 6,500-Person Gulag

During a livestreamed employee presentation this week, someone interrupted the session with an outburst about being “the bitch of the company” and asked the speakers to tell a Meta AI manager that he was “a piece of shit,” as quoted by WIRED from a recording. The background is the Applied AI unit formed in March, with around 6,500 engineers and product managers tasked with improving the models from the Superintelligence Labs. Three current employees describe the work as soul-crushing: two tasks per week, such as generating riddles and coding problems to test models. “It is literally the gulag,” one says. In May, Meta had cut 8,000 jobs, 10 percent of its workforce; over 1,600 employees signed a petition against the tracking of clicks and keystrokes to gather training data. Chief Product Officer Chris Cox spoke internally of the “insanity of this company,” while Zuckerberg admitted mistakes in a memo and promised stability and no further mass layoffs this year. → www.wired.com

Synthszr Take: In May, we wrote that 15,000 employees were affected by Meta's AI reorg, with half having to leave. Now we see what's happening to the other half: highly paid engineers completing two tasks a week and calling themselves “draftees” because their only choice was the unit or termination. This is Meta's answer to the question of what people do when models eat up the interesting work, and the answer is a manager-to-employee ratio of 50 to one, meaning virtually no leadership. Cox is right with his “It is neither god, nor is it the devil,” but it's precisely this honesty that is missing in the organization's design. Ego gets in the way, and here it's the belief that you can stick 6,500 talented people in a data factory and hope for loyalty. Velocity isn't created by mass, but by trust in the judgment of the people building the product; “people over processes” isn't a poster, it's the prerequisite for someone even getting out of bed in the morning. If you degrade people into training data suppliers and record their keystrokes, you shouldn't be surprised by the next “spicy” opening in a livestream.

Kimi K2.7 Turns Coding Models into a Commodity

Moonshot AI has released the next open-source model for programming tasks with Kimi-K2.7-Code, and the leaps over its predecessor K2.6 are tangible: +21.8% on the Kimi Code Bench v2, +11.0% on the Program Bench, and +31.5% on MLS Bench Lite. In addition, there's a detail that matters more in practice than any benchmark bar: 30% fewer reasoning tokens for comparable results. vLLM documented the architecture in its support post: a 1-trillion-parameter MoE with 32B active parameters, MLA-Attention, and a 256K context. Weights and code are separately linked and open. With this, Moonshot continues the cadence that was already noticeable with K2.5 and K2.6. For those keeping count: This is the third serious iteration from Beijing in about a year. → AINews

Synthszr Take: The 30% fewer reasoning tokens are the real news, not the benchmark numbers. Tokens are the electricity bill of AI, and a coding model that thinks one-third less per task while performing better directly pushes down inference costs. The Jevons paradox applies here in full force: the cheaper the token, the more code you let the machine write. In the code crash framework, DeepSeek appears as a cheap China option with a compliance question, and Kimi belongs in the same column, with the same caveat. Wolf Köhler's objection remains valid: Open Weight is of little use to medium-sized businesses if no one hosts the 1T model themselves, so this also runs via Bedrock or Azure. For coding agents, Claude and Cursor will remain ahead in 2026, but the gap is closing faster than the price lists of foundation models can keep up. Anyone still narrowly defining their vendor map to U.S. providers should add a maturity indicator to the China column tomorrow morning.

Huawei Releases openPangu 2.0 (Open Source from June 30)

Huawei has announced openPangu 2.0 and plans to gradually open-source it starting June 30: architecture, weights, technical reports, inference code, as well as pre-training and post-training code, including training operators. The models use an MoE design with 512K context windows and high sparsity. The Pro version has a total of 505B parameters with 18B active, while the Flash version has 92B total with 6B active. Huawei claims that the inference throughput, optimized on its own Ascend chips, is up to 2x higher compared to common operators. This way, Huawei directly links the open-source model to its own hardware. Back in April, we described how Deepseek v4 is moving to Huawei chips, completely avoiding Nvidia. → AINews

Synthszr Take: The real leverage isn't in the 505B parameters, but in the two small letters: Ascend. Huawei gives away the weights and sells you the chips on which they run best. The 2x throughput claim is the marketing message; the moat is the integration of a free model and proprietary silicon. For a medium-sized German company, this is not an issue for now, as no one here can handle self-hosting on Ascend hardware, and the compliance question with a Chinese stack remains open anyway. But the strategic pattern is clear: whoever makes the open-source model free commoditizes the software and shifts the margin to the infrastructure, exactly where Nvidia has been the sole beneficiary so far. June 30 will be interesting, when the training code is actually released and we can see if the sparsity numbers hold up. Anyone building a model strategy in 2026 should no longer list China as a laggard, but as a second, hardware-driven axis alongside the U.S. camp.

Firecrawl is Becoming the Data Supplier for Agents

Firecrawl is positioning itself as a context API for AI agents: search, scrape, and interact with the live web, outputting clean Markdown or structured JSON. They claim to be trusted by over 150,000 companies, and their open-source repo has 132,400 stars on GitHub. The numbers they advertise are technically specific: 96% of the web covered (including JS-heavy pages), P95 latency of 3.4 seconds across millions of queries, and 93% fewer input tokens because navs, footers, and ads are stripped out. New is the Interact function, which allows an agent to scrape a page and then operate it via prompts (fill a search field, click the first result). There's also a /monitor feature that notifies the agent when a page changes, and an agent onboarding process where an agent can fetch an API key for itself via CLI. They list Lovable, Zapier, Replit, and Gamma as reference customers. → Techpresso

Synthszr Take: The most exciting number isn't the latency, but the 93% fewer tokens. Anyone building an agent fleet pays their bill in tokens, and most of that is currently burned on website junk that no model needs. Firecrawl essentially sells compute discipline: clean inputs so the expensive reasoning step doesn't have to digest 38,000 tokens of footer navigation. This is exactly the MCP tool layer I describe in the Agent Fleet Playbooks, the Search-MCP and the Vector-Store that the research agent connects to. The fact that it's open source but sold as a hosted API is the classic leverage: the repo builds trust and diffusion, while revenue comes from the infrastructure no one wants to run themselves. The interesting part is the moat, because you don't build 96% web coverage overnight; that's years of work on edge cases. Anyone building agents today shouldn't code the web access themselves, but buy it and invest the saved token budget in reasoning.

Protests Block $130 Billion in Data Center Projects

In the first quarter of 2026, at least 75 data center projects in the US, valued at around $130 billion, were blocked or delayed, the highest figure since Data Center Watch began tracking in 2023. Researchers at the AI intelligence firm 10a Labs explicitly state this is not a cyclical blip but a structural shift: the number of active resistance groups has more than doubled to 833 across 49 states. This quarter alone is approaching the $156 billion recorded for the entire year of 2025. Sociologist Tressie McMillan Cottom, who accompanied organizers in North Carolina, described in a guest essay for The New York Times how people are flocking to town hall meetings about water rights, land use, and thermodynamics, crossing party lines. Her conclusion: the resistance is giving many a first taste of political power. Both parties are increasingly sympathizing with the protests, and the issue is likely to shape the midterm elections. Where authorities were previously criticized for backroom deals, they are now facing backlash before anything is even signed. → Techpresso

Synthszr Take: The entire AI story of the last two years was based on the assumption that compute is a bottleneck of chips and capital. Turns out, the scarcest resource is the consent of the people who live next to the humming. 833 groups in 49 states are no longer a local zoning dispute; this is a movement with a playbook, and a playbook scales just as fast as a language model. In January, we wrote that Microsoft wants to proactively manage resistance, and in June, that Meta is setting up Mad Max-style tent data centers. It's exactly this brute-force mentality that's producing the headwind 10a Labs is now measuring. For Europe, there's an uncomfortable punchline in this. Anyone here talking about sovereignty should understand that the real leverage isn't in more concrete, but in the domain knowledge and data of the hidden champions that no hyperscaler can replicate. Compute discipline is becoming a question of location, and that's a task for industrial policy that can be tackled tomorrow morning, not after the next permit disaster.

The Laziest Senior Developer Writes the Best Code

A new open-source plugin called Ponytail (MIT-licensed, on GitHub) gives Claude Code a “lazy senior dev” mode. The idea behind it is a simple minimization checklist: before the agent writes new code, it checks if the standard library, native features, existing dependencies, or a one-liner would suffice. In a benchmark across five tasks, the author used about 16% fewer tokens and the whole process ran about 4x faster. The generated code shrank from 293 lines to 47. A single example, a 190-line countdown function, was reduced to a fraction of its original size. No more complex model, no new architecture, just an enforced way of thinking before the first token is generated. → AINews

Synthszr Take: Anyone who has ever worked with a real senior developer knows the reflex: the best pull request is the one that deletes code. Ponytail is now industrializing exactly that for agents, and going from 293 to 47 lines is not a rounding error; it's a factor of six less material that someone will later have to read, test, and maintain. The real gain isn't in the 16% saved tokens (nice, but compute is getting cheaper anyway), but in the maintenance load that is never created in the first place. We wrote in February that context is king; here, the other half of the truth emerges: constraint is queen. A 200-line prompt trick that forces a model to exercise restraint beats the next-larger model that produces unbridled feature-itis. Anyone setting up their coding stack in 2026 should treat such guardrails as the default, not as a gimmick for tinkerers. This can be written into the CI pipeline tomorrow morning, not after the next tooling offsite.

Degrowth Would Turn Europeans into 'Europoors'

Noah Smith attacks the degrowth movement in his latest post, contrasting it with the debate over European living standards. On one side are economists like Mario Draghi, Philippe Aghion, and Luis Garicano, who look enviously at the U.S. tech sector and call for liberalizing reforms because Europe has no real equivalent. On the other side, U.S. liberals like Paul Krugman and Brad DeLong defend Europe, attributing the lag to longer U.S. working hours and measurement problems. However, Smith identifies a second, more dangerous debate: Thomas Piketty and his World Inequality Lab argue in a manifesto that Europeans are too rich and should become poorer, with reduced working hours, growth caps, and less consumption to combat climate change. A Guardian essay by the same group of authors (Piketty, Stiglitz, Ghosh, Raworth, Hickel) was flagged by Pangram as 100% AI-generated and, according to Smith, reads like a mix of buzzwords. Smith dismantles degrowth “research” as vague conceptual acrobatics without a concrete plan. → Noahpinion

Synthszr Take: A manifesto against growth, written by an AI that exists only thanks to massive compute investments. The irony writes itself. Europe has a real problem, and it's not too much wealth, but a lack of oxygen for further development. In May, we wrote about SoftBank's €75 billion for French data centers and Mistral's billion-dollar revenue target; that's the direction that generates velocity, not growth caps. Piketty manages complexity on a societal level, just as our martech stacks with more code than a nuclear power plant manage complexity: lots of motion, no progress. If you want to lift 10% of the world's population out of extreme poverty, you need a productive economy, not a sermon about planetary boundaries in 100% AI prose. Becoming poorer has never solved any climate problem; it has only intensified the struggle over distribution.