Upside-Down World: Trump Channels Bernie Sanders while Google, Meta & Co. Seek Cash on the Stock Market

Trump Channels Bernie Sanders and Wants to Nationalize Parts of OpenAI & Co.

According to a report from NOTUS, high-ranking US officials have held preliminary talks with major AI firms about the federal government acquiring stakes in the very companies whose technology it also aims to regulate. The common thread leads to OpenAI: Sam Altman pitched the idea directly to Donald Trump in early 2025 and has brought it up again with government officials in recent weeks. The talks are about companies voluntarily ceding stakes, rather than the state buying in. One option being discussed: The proceeds would flow as a dividend to all American households. The context is sensitive, as a NOTUS poll cited in the report shows that 55% of Americans believe AI does more harm than good in their daily lives. Critics immediately point to the structural conflict: A state that holds shares in a company it regulates is both a shareholder and a referee. Similar proposals are also circulating from the left, and for now, these are just talks, not a political decision or a deal. → Techpresso

Synthszr Take: Sam Altman has been repeatedly pushing this model since early 2025, and it fits with his appearance in April when he called for higher taxes and a universal basic income. Now he's providing the mechanics to go with it: The state holds stakes, and the returns land in every household as a dividend. The dividend is primarily about reputation management, especially when the majority of Americans already believe that AI harms their daily lives. The real problem lies in the structure. Anyone who holds shares in a company they are supposed to control will ultimately protect the value of their investment instead of the public, and no amount of feel-good dividend narrative can talk that away. The fact that the left arrives at the very same idea from the same job displacement argument only shows how open the question of who actually owns the profits from this technology is.

Et Tu, Meta? Everyone Wants to Tap the Stock Market

According to the Financial Times, Meta is considering a stock issuance in the tens of billions to finance its own AI ambitions. The report comes just days after Alphabet's increased issuance of $84.75 billion, which Google is also using to pay for the construction of data centers. Meta executives are reportedly looking for “creative” ways to raise money, and the talks have intensified following Alphabet’s success. Back in October, Meta filed for its largest bond offering ever (up to $30 billion) and agreed to a $27 billion financing deal with Blue Owl Capital. In April, the company raised its annual investment forecast to between $125 billion and $145 billion. The stock fell 6.6 percent after the report. Banks have not yet been hired, and according to the FT, a final decision is still pending. → www.reuters.com

Synthszr Take: When the world's richest advertising company starts borrowing money from the stock market, something has shifted. For years, Meta has paid for its investments out of its current cash flow; now, it's adding tens of billions from a stock issuance, plus a $30 billion bond and $27 billion from Blue Owl. The 6.6 percent drop in the stock price shows how nervous the market is becoming: an annual capex of $125 to $145 billion is a bet that all those data centers will pay off. As we wrote about Google's record issuance in early June, the capital markets have become the pacemaker in this race. Whoever gets oxygen the cheapest, builds the fastest. My impression: Capital is flowing freely right now; the open question is whether Meta can turn that computing power into useful products and not just stack expensive hardware in warehouses. The company has yet to provide this proof since its AI reorganization in May. Speed in raising capital is of little use if the products lag behind.

The AI Boys Are Damaging AI's Image More and More

Jeff Jarvis describes a paradox on Medium: 57 percent of Americans believe the risks of AI outweigh its benefits, making the technology more unpopular than the border agency ICE. Half are worried about the consequences for creativity, relationships, and elections, and Republicans and Democrats alike are blocking new data centers from New Jersey across the country. At re:publica, there's a workshop on poisoning AI training data, the FT is proclaiming a new Luddite movement, and commencement speakers are booed as soon as they utter the letters A and I. At the same time, actual usage is increasing: over half now use the tools for research, a 38 percent increase in one year. Jarvis blames the technology less than the “AI boys”—Musk versus Altman in the recently concluded OpenAI lawsuit, plus Thiel, Andreessen, Karp, Zuck, Ellison, and Bezos, who sometimes dismiss people as “meat computers.” His hope: In Jensen Huang and Yann LeCun, the industry at least has two gifted communicators. → Medium Weekly Digest

Synthszr Take: There's a gap here that any product person would immediately recognize. 57 percent say the risks outweigh the benefits, and in the same year, usage increases by 38 percent. People speak ill of AI and yet type their questions into ChatGPT every day. The product is convincing; its salesmen are ruining its reputation. Anyone who dismisses people as “meat computers” while investing billions in data centers that residents from New Jersey to Bavaria are blocking shouldn't be surprised by the bad mood. Trust is built in everyday life, when the tool makes work easier tomorrow morning, not in the next manifesto about superintelligence. Jarvis is betting on Huang and LeCun as better storytellers; I'd be more inclined to bet that tangible benefits will drown out the loudest podcasts.

Claude Writes 80% of Anthropic's Code and Makes It 52x Faster

Anthropic has published a report with numbers that are hard to ignore. In a consistent test to train a small AI model faster, Claude Opus 4 achieved a 3x speedup, while the new Mythos Preview model achieved a 52x speedup. Over 80 percent of the code that goes into Anthropic's production codebase is now written by Claude instead of humans. In research sessions where humans had taken a wrong turn, Mythos Preview suggested the better next step in 64 percent of cases, up from 51 percent just six months earlier. The length of tasks that the AI can handle autonomously is doubling roughly every four months: from four-minute tasks with Opus 3 to twelve-hour tasks with Opus 4.6. Anthropic emphasizes that true recursive self-improvement is not yet inevitable, but could arrive sooner than many expect. In parallel, OpenAI's Dreaming-V3 system is doubling ChatGPT's memory for Plus and Pro users and increasing factual accuracy from 41.5 to 82.8 percent. → AlphaSignal

Synthszr Take: The number that sticks is 80. Four out of five lines of code in Anthropic's production are now written by Claude itself, and the jump from Opus 4 (3x) to Mythos Preview (52x) shows how steep the curve has become. Anthropic cautiously states that recursive self-improvement is not yet inevitable. But when task horizons double every four months, from four-minute tasks with Opus 3 to twelve-hour tasks with Opus 4.6, the bottleneck shifts away from the model to the question of who can even direct this pace meaningfully anymore. In April, we wrote that the Mythos benchmarks would shock the competition; now, it's the speed at which the loop is closing that's shocking. The fact that OpenAI is doubling ChatGPT's memory in the same week (factual accuracy up from 41.5 to 82.8 percent) fits the picture: everyone is currently building the scaffolding before the ceiling disappears. Stop celebrating individual benchmarks and start asking which parts of your own code pipeline you will hand over in the next twelve months to models that already suggest the better next step in 64 percent of cases.

Gary Marcus Thinks Anthropic's Rhetoric is Nonsense

On his Substack, Marcus on AI, Gary Marcus deconstructs a wave of headlines claiming that Anthropic has called for a pause in AI development. His finding after a close reading: nothing of the sort. Anthropic speaks of an option that the company has no plans to use now or in the foreseeable future. Instead, it wants to continue accelerating, pointing to the “most reckless actors,” i.e., China, as justification. Marcus calls this a free piece of rhetoric, perfectly timed for the upcoming IPO. His conclusion to his readership: Caveat emptor, caveat lector—let the buyer and the reader beware. → Gary Marcus from Marcus on AI

Synthszr Take: Gary Marcus reads the press release more closely than most newsrooms, and it pays off. Anthropic talks about a pause it never intends to take itself. This rhetoric costs nothing and gains a lot: you appear responsible while continuing at full speed, pointing to the “most reckless actors” or China when needed. In March, Anthropic's valuation was nearly $20 billion, now the IPO is approaching, and at this exact moment, the word pause appears. Anyone who writes a 20,000-word manifesto on Claude's virtue ethics while simultaneously training models for the US military on a different constitution has learned to have it both ways. Read the headline, then read the paragraph below it. The difference between the two is the real business model.

Huawei's Chip Queen Declares Moore's Law Dead

He Tingbo has been leading Huawei's internal chip unit since 2003, with an annual budget of around $400 million. This month, she introduced the so-called Tao Scaling Law, an evaluation model for semiconductors that abandons the race for ever-smaller transistors and instead measures how quickly data flows through a chip. Jensen Huang, CEO of Nvidia, has publicly conceded the Chinese market for AI chips to Huawei, which will hold around 50 percent of it in 2026. According to Morgan Stanley, this market will grow from $21 billion this year to $67 billion by 2030. Huawei's plan: Without access to Nvidia's Blackwell chips, He Tingbo's team stacks and clusters domestic chips and couples the computing power with cheap solar energy from data centers near the Gobi Desert. The gap remains real; the best US chips are still estimated to be about five times more powerful by the Council on Foreign Relations, and in an optimistic scenario, Huawei will only supply about 4 percent of the computing power that Nvidia produces. → Alice Han & James Kynge

Synthszr Take: The 5x gap and the 4 percent are the numbers everyone is focusing on. They answer the wrong question. It's enough to gather sufficient computing power for domestic needs, and that's exactly what He Tingbo is organizing through stacking and clustering, coupled with cheap solar power from the Gobi. When the physics of miniaturization hits a wall (transistors are now smaller than a virus), the leverage shifts from transistor size to architecture and energy. This is textbook leapfrogging: those with little established infrastructure have little to migrate, and electricity from the desert scales cheaper than a 3D TSMC process. As early as late May, we noted that Trump's export controls are fueling China's AI rather than slowing it down, and the Tao Scaling Law is the next piece of evidence. Whether the model will ever become the industry standard for benchmarking remains to be seen; the underlying self-confidence is already real, and the +40 percent rise in Chinese semiconductor stocks since the beginning of the year is the picks-and-shovels bet on exactly this stance.

Meta's Glasses are Equipped with Hidden Face Recognition

Wired found a previously unreleased facial recognition feature in the source code of the Meta AI app, internally called “NameTag,” which is intended to run on Meta's Smart Glasses. According to the report, the feature can capture faces and later notify the wearer when it recognizes a previously saved face. None of this is currently active, and a security researcher confirms that no biometric data is sent to Meta's servers; however, earlier app versions already contained a “Connections” menu with the note “remember the people you met.” The New York Times had already described the tool in February under the same name, “Name Tag.” An internal memo reveals that Meta intended to launch the feature specifically in a “dynamic political environment” because critical organizations would then be otherwise occupied. Meta itself had shut down its Facebook facial recognition in 2021 for privacy reasons and reintroduced it in 2024 as an anti-fraud tool. Spokesperson Ryan Daniels emphasizes that nothing has been shipped to customers and that the company is not building a central facial database. → Techpresso

Synthszr Take: According to an internal memo, Meta wanted to launch the facial recognition feature in a “dynamic political environment” because critical civil rights groups would be busy with other concerns. A remarkably honest strategy note: you wait until the gatekeepers are looking away. In 2021, Meta buried facial recognition on Facebook for privacy reasons; in 2024, it returned as an anti-fraud tool; and now the same code is in the Smart Glasses app under the name NameTag. In mid-March, we wrote that Meta's glasses see everything; NameTag would turn “see everything” into “recognize everyone,” complete with the friendly reminder “remember the people you met.” Ryan Daniels claims they aren't building a central facial database, and maybe that's even true. But when a feature shows up in the logs three times, it's coming; the only open question is under what guardrails. We should lock those down while the code is still in the backlog and not on the noses of millions of people.

47 Seconds, and Then the Attention is Gone

Gloria Mark, a psychologist at the University of California, Irvine, has been researching how people interact with digital technology for 30 years and drew a sobering conclusion at SXSW London. In her “Living Laboratories,” she measured an average attention span of about two and a half minutes in 2003. By 2012, it was only 75 seconds; between 2014 and 2020, the value dropped to 47 seconds. Heart rate monitors showed a direct correlation between rapid attention switching and increased stress with decreased performance. In parallel, lawsuits are piling up: Meta and Google's YouTube were ordered to pay millions to a 20-year-old who claimed an addiction developed in childhood, a rural school district in Kentucky demanded over $60 million, and about 1,200 other districts are following suit. However, Mark also emphasizes that social media can create spaces of belonging, especially for marginalized groups; a 2024 LGBTQ+ survey describes platforms as places of rejection for some, but places of friendship for others. → The Download from MIT Technology Review

Synthszr Take: 47 seconds. That's how long an average person can maintain attention on one thing before switching, and that was from 2014 to 2020, long before AI chatbots. What Mark is describing is the backstory; ChatGPT and Gemini are now adding another layer on top by conveniently outsourcing thinking itself. I've noticed it in myself: since AI started pre-formulating my answers, I find myself asking less often if I've actually understood the issue. For anyone building products, here's the lesson. Engagement through constant interruption is a dead end that is bringing 1,200 school districts to court and costing Meta millions. Anyone developing software can decide tomorrow morning whether to fragment attention or protect it, and tools that keep us focused will become the most valuable part of the market. Attention is the last truly scarce resource, and those who respect it will win.