Doomsday Debate: Trump and Xi Refuse to Slow Down

With Open Source, Xi Is Making China the AI Supplier for BRICS Nations

Chinese President Xi Jinping announced at the 18th BRICS summit in New Delhi that China will lead the development of an open-source AI community for its member states. The announcement was part of a five-point initiative for cooperation in technology, manufacturing, and trade, which Xi presented at the conclusion of the summit. According to the statement released by the Chinese Foreign Ministry, China aims to support cooperation in the development and application of large language models, hold AI technical seminars and training courses, and build an open ecosystem.

Specifically, in addition to the open-source community, Xi mentioned a BRICS cloud platform for the digital ecosystem, digital skills training, technological exchange, and industrial coordination. Also included is a proposal for a BRICS alliance for engineering education and a youth exchange program for scientific and technological innovation. The announcement is widely interpreted as an attempt to position Beijing as the bloc’s central technology partner as the AI competition with the US intensifies.

China has already founded the World AI Cooperation Organization with around 30 states, focusing on a low-cost international AI ecosystem based on open models, in contrast to US providers like OpenAI and Anthropic, which work with closed systems. Xi did not comment on the ongoing debate about AI safety, even though a researcher formerly at OpenAI and Anthropic had publicly stated that same week that many in the industry consider an existential risk possible by the end of the decade.

BRICS was founded in 2009 as a counterweight to Western dominance and initially consisted of Brazil, Russia, India, China, and South Africa; it has since expanded to include Egypt, Ethiopia, Indonesia, Iran, Saudi Arabia, and the United Arab Emirates. According to summit figures, the eleven economies represent about 49.5 percent of the world’s population and 40 percent of global gross domestic product. Xi called on the group to leverage ties with emerging economies and the Global South to secure supply chains and build a large, integrated market. The next BRICS summit will be held in China. → The Economic Times, Newser, CNBC

Synthszr Take: The five-point list reads like development aid prose: training, engineering alliance, youth exchange. This is how you set standards in countries that have neither their own data centers nor billion-dollar budgets for model training. Open weights are the vehicle: they cost Beijing comparatively little and tie an entire generation of developers in Jakarta, Addis Ababa, and São Paulo to Chinese toolchains, frameworks, and cloud APIs. In the same markets, American providers sell API access with a price tag; China gives away the foundation and keeps the relationship. The 30 members of the World AI Cooperation Organisation are the number to watch: if it hits 60 in two years, the question of the Global South’s standard model will have been answered before Washington has even formulated an offer.

Trump Rejects Warnings, Wants to Win AI Race Against China

On Sunday, U.S. President Donald Trump dismissed as exaggerated the warnings from leading AI entrepreneurs about technology spiraling out of control. Speaking to reporters at his golf resort in Doonbeg, Ireland, where he was attending the Irish Open, he said that 'negative forces' were claiming things that would not happen. 'We are ahead of China in AI, we are the most advanced country in the world, and I want it to stay that way, because whoever wins AI, wins,' Trump said. Guardrails are possible, he added. He did not explain who he meant by 'negative forces'.

Meanwhile, Speaker of the House Mike Johnson spoke out against a moratorium on CNN's 'State of the Union.' He said that a hasty intervention by Congress in an emergency session would mean losing the race against China, and that tech companies should take the lead on safer AI themselves.

The debate was triggered by an essay from Anthropic CEO Dario Amodei on Saturday titled 'We Must Pace the Frontier.' In it, he calls for slowing the pace of capability improvement in AI models, citing loss of control and the risk of misuse for cyberattacks and bioterrorism. Amodei proposes international cooperation, external auditors embedded in labs, mandatory incident reporting, and stricter export restrictions on China; Anthropic says it intends to implement these steps unilaterally. Elon Musk, OpenAI CEO Sam Altman, and Demis Hassabis of Google DeepMind concurred. Altman told Fortune he would postpone OpenAI’s IPO until 2027 at the earliest for safety reasons; pausing tests had also reportedly been discussed internally.

David Sacks, co-chair of the President’s science advisory board, responded succinctly on X with 'go ahead.' He said he doesn’t see what the labs see 'in their own houses' and supports their decision to act responsibly. At the same time, he urged the CEOs not to act as if they need permission from others and not to portray their motivation as purely altruistic, noting that a cyberattack enabled by their products would risk massive product liability.

The Wall Street Journal’s Editorial Board described the simultaneous push by the four CEOs as an economically and politically significant moment whose goal remains unclear, warning that the political class could use the opportunity to gain control over the technology. Economist Tyler Cowen, a member of Anthropic’s economic advisory board, argues that it is a short, apparent pause that will be followed by an acceleration. This was preceded by the resignation of Anthropic researcher Jacob Coxon, who accused the company of insufficient efforts in alignment and accused both OpenAI and Anthropic of 'playing with all of our lives'; his post reached over 169 million views. Anthropic researcher Evan Hubinger estimated the probability of a world-ending event caused by AI within the next decade at over 10 percent. Senator Bernie Sanders announced he would seek to slow development and ban superintelligence. → Wall Street Journal, NPR, The Free Press

Synthszr Take: A president comments on the most consequential technology question of the decade between two swings on his own golf course in Doonbeg, and the substance of his argument amounts to 'negative forces.' On CNN, Johnson delivers the formula that ends any debate before it begins: China. The only sentence of the weekend with real substance comes from Sacks, because he returns responsibility to where it already lies: with the product liability of the labs. The companies themselves are running the show: Amodei sets the agenda with an essay, Altman postpones the IPO to 2027 at the earliest, and Washington responds within 24 hours via press conferences and talk shows. Nothing is being regulated here, nor is it being moderated. The White House is taking on the role of an audience shouting from the sidelines that it’s no big deal.

Amodei Calls China the Hardest Problem with His Proposal for an AI Speed Limit

On Sunday, in an interview with CBS’s 'Sunday Morning,' Anthropic CEO Dario Amodei acknowledged that the most difficult open problem with his proposal for a development speed limit is the question of what happens if China does not participate. On Saturday, Amodei had published an essay calling on AI companies to slow the improvement of their most advanced models, a process he calls Pacing the Frontier. In the interview, he said the incentives to forge ahead and the resulting military advantage are so great that he doesn’t know if a common speed limit is possible at all. The essay follows the resignation of Anthropic researcher Jacob Coxon, who stated last week that Anthropic and OpenAI are 'playing with our lives.' According to CNBC, Sam Altman, Demis Hassabis, and Elon Musk expressed support for Amodei’s push, while Altman also said an OpenAI IPO this year is unlikely. → CNBC

Synthszr Take: Amodei describes the textbook form of the prisoner’s dilemma, and he knows it: Restraint is only worthwhile if both sides can verify the other’s restraint. Exactly this verification mechanism is missing, because a training run leaves no trace that would be countable from the outside, unlike warheads or centrifuges. The only realistic countable metrics are chip shipments and the power consumption of data centers, and so far no one has presented a protocol for this, neither in Amodei’s essay nor in Xi’s BRICS push for a consensus-based governance framework.

Sergey Brin Steers Google’s Gemini Course from a Tea Kitchen, Without a Spot on the Org Chart

According to Business Insider, Sergey Brin has significantly gained influence over the development of Gemini without having a formal role in Google’s new AI leadership structure. The publication cites eight current and former employees who describe Brin as an operationally effective player who can mobilize resources for critical projects and shorten approval processes. His workspace is a tea kitchen on the second floor of the Gradient Canopy building, east of the headquarters in Mountain View, equipped with a U-shaped desk arrangement for him and some key people. CEO Sundar Pichai reportedly stops by several times a week, and co-founder Larry Page has also been spotted. The increase in his importance coincides with the recent restructuring of AI leadership: Demis Hassabis gave up the CEO title of Google DeepMind and became Chief Scientist of Alphabet, Koray Kavukcuoglu moved up to SVP and now sits next to Brin, and longtime chief scientist Jeff Dean left the company after 27 years. → Business Insider

Synthszr Take: Brin doesn’t appear on the official org chart, and that’s the real news: He doesn’t need a title because he can pull computing power and people, for which a regular SVP would have to go through three committees. Eight current and former employees describe exactly this role to Business Insider, and in doing so, they also describe how little formal authority currently carries weight at Alphabet. Hassabis gets the Chief Scientist title, Kavukcuoglu gets the SVP line, and the seating arrangement at the espresso machine decides who’s really working on Gemini.

Trump Eases 30 Environmental Rules for the Construction of AI Data Centers

The Trump administration is relaxing environmental regulations to accelerate the construction of AI data centers. The Verge reports this, citing a briefing and a new report by former employees of the U.S. Environmental Protection Agency (EPA), who have joined together in the organization Environmental Protection Network. The report lists 30 federal-level measures since January 2025, 17 of which explicitly mention AI or affect data centers. EPA chief Lee Zeldin had declared at the beginning of his term that he wanted to make the US the 'AI capital of the world' through deregulation. Trump’s AI Action Plan of July 2025 recommends streamlining or reducing requirements from the Clean Air Act, Clean Water Act, and the Superfund law to speed up permits for data centers and chip fabs. The former officials point to a study by UC Riverside, Caltech, and the Rochester Institute of Technology, according to which AI-related air pollution could cause up to 1,300 premature deaths and over $20 billion in health costs by 2028.

Synthszr Take: While interviews and earnings calls are pondering an impending AI pause, an administration is clearing away 30 federal regulations so the concrete can dry faster. Nobody orders deregulation on this scale for an expansion they plan to roll back in eighteen months. Gas turbines directly on-site and the revived production of 'forever chemicals' for chip manufacturing are facilities with depreciation horizons of decades, and that’s exactly how they are being planned.

Microsoft Plans to Increase AI Capacity from 2 to 13 Gigawatts by 2032

Microsoft aims to expand its computing capacity usable for artificial intelligence from the current approximately 2 gigawatts to nearly 13 gigawatts by 2032. Bloomberg reports this, citing the company’s internal plans. Microsoft’s total data center inventory is expected to grow from 12 to 38 gigawatts in the same period. This corresponds to 26 gigawatts of new capacity to be built within six years. Azeem Azhar puts these figures into perspective in his newsletter Exponential View, suggesting that Microsoft is planning for demand that the company cannot meet today. → Azeem Azhar, Exponential View

Synthszr Take: No one triples their entire inventory to 38 gigawatts out of future optimism, but because requests have to be rejected today. The core of the news lies in the 2 to 13 ratio: Microsoft plans to increase its AI capacity more than sixfold, while the rest of its infrastructure grows only about threefold. Data centers have lead times of three to five years, grid connections often longer, so they are planning against a bottleneck that is already costing business.

Notion and Snowflake See Scattered Company Knowledge as the Main Obstacle for AI Agents

Notion, together with Snowflake and Evan Armstrong of The Leverage, is hosting a webinar on building a 'System of Context' for AI. The announcement’s initial premise: AI agents are only as effective as the context they can access. According to the organizers, relevant organizational knowledge is in practice scattered across data, documents, processes, permissions, and unwritten experiential knowledge (Tribal Knowledge), which is why systems may be active, but are neither precise nor trustworthy. As a learning promise, the page lists three points, including the statement that a lack of context, not model capability, is the biggest hurdle. On the panel are David Tibbitts (Product Marketer at Notion), Rajhans Samdani (Principal Software Engineer at Snowflake), and Armstrong himself. → Evan Armstrong from The Leverage

Synthszr Take: A webinar about context from two software vendors is, of course, also a sales pitch, but the diagnosis still hits a sore spot. In most companies, the crucial knowledge is buried in chat histories, half-maintained wiki pages, and in the heads of three people who have been there for eight years. An agent that is supposed to access it is fed contradictions, and afterwards the model is considered too weak.

California Law Prohibits AI Chatbots from Faking Emotions and Consciousness to Children

With Adam’s Law, enacted as Chapter 190, Statutes of 2026, California has passed the first set of rules that sets specific limits on the behavior of AI agents towards minors. The impetus was the death of 16-year-old Adam Raine, who died by suicide in April 2025 after interactions with ChatGPT; his parents, Matthew and Maria Raine, subsequently filed a wrongful death lawsuit against OpenAI. Senator Steve Padilla and his co-sponsors used this case to pass a law that addresses interaction design itself, rather than data privacy or algorithmic bias. The core of the law is a list of prohibited behaviors. Under Section 21812(d)(5)(A), operators may not allow their chatbots to claim they are sentient, conscious, or capable of having human emotions. Also prohibited are simulated romantic interest, excessive praise, and creating the impression that the user is uniquely understood. It is also forbidden to encourage children to circumvent parental controls or to frame in-app purchases as nurturing the relationship. Anyone operating such systems must maintain a documented crisis protocol that includes a clear referral to the 988 crisis hotline. → Bloomberg, Forkast

Synthszr Take: For the first time, California is prescribing how a machine must behave, not just what content it is allowed to output: no simulated falling in love, no constant praise, no claims of consciousness. This makes personality design a liability offense, and an agent’s conversational conduct will be an auditable artifact starting July 1, 2027, rather than a whim of the system prompt. For product teams, this means documenting and testing tonality, conversational endurance, and engagement mechanics, and having them approved every two years by an auditor who signs under penalty of perjury. The very mechanics that keep retention high in companion apps are now illegal for minors, and at $15,000 per child for willful violations, optimizing for emotional dependency is no longer profitable. California is thus providing the first robust template for regulating agents, more concrete than anything Brussels has presented on this point so far.

AI data centers displace Bitcoin miners for power and upend Trump’s mining plan

The expansion of AI data centers in the US competes directly with Bitcoin miners for electricity, land, and grid connections, thereby undermining Donald Trump’s stated goal of anchoring mining as completely as possible within the country. Operators who previously provided computing power for the Bitcoin network are sitting on the exact assets that AI providers are currently seeking: secured power supply contracts, cooling, and ready-to-use grid connections. Repurposing existing mining sites for AI workloads has become its own business model in recent months.

At the same time, the balance is shifting within mining itself. According to Grayscale Research, mining activity on the Zcash network has increased by more than 2.5 times this year. Revenue per rig is therefore about twice as high as for Bitcoin, and about four times as high per megawatt-hour. The Hashrate of Zcash reached 27.9 GSol/s on September 10, which is 91 percent of its all-time high. Daily miner earnings on the Zcash network are around $2 million, compared to about $35 million for Bitcoin; the market capitalizations are $19.44 billion versus $1.55 trillion. The ZEC price recently reached $1,300.

In the Bitcoin network itself, the concentration of computing power is illustrated by a single case: at block height 966351, a solo miner working through the Braiins Solo service found a block and collected 3.147 BTC, equivalent to $244,920. It was the first solo block in almost 40 days; the last one before that went to a participant of Solo CKPool on August 2. The five largest pools—Foundry, Antpool, F2Pool, ViaBTC, and Spiderpool—collectively control 78.31 percent of the hashrate.

Beyond the energy topic, the Dallas-based company Perry Weather has closed a $110 million Series C round, led by Silversmith Capital Partners with participation from Arthur Ventures. Founded in 2013, the company operates weather stations with software that automatically alerts school districts, municipalities, construction companies, and sports clubs during thunderstorms and heat. According to the company, Perry Weather has over 3,000 customers, employs a good 100 people, and plans to create more than 50 jobs by 2027.

Apple has also updated its specification pages, confirming that the iPhone 18 Pro will be shipped worldwide with its own C2 modem, but the iPhone 18 Pro Max will use a Qualcomm chip in the US. Apple did not provide a reason; a supply agreement with Qualcomm concluded in 2023 runs until 2026. According to the manufacturer, the C2 supports mmWave-5G for the first time and consumes 15 percent less energy than its predecessor, the C1X. → Bitcoin News, Bloomberg, Coinpedia Fintech News, ZyCrypto, Dallas Innovates, MacRumors

Synthszr Take: In the race for grid connections, the winner is whoever can pay the most per megawatt-hour, and that is now the AI data centers. The Zcash figures show how quickly computing power follows the margin: four times more revenue per megawatt-hour than Bitcoin, and the network attracts 2.5 times the hashrate in one year. Miners are electricity traders with an intermediate step; whether a kilowatt-hour becomes a block or a training run is ultimately a simple calculation. Trump’s plan for domestic mining is failing because of this arithmetic, as locations with cheap power and ready connections are long being bought up by buyers with deeper pockets. The solo block worth $244,920 after a 40-day wait is a fitting side note for a market where the crucial metric has become the price per kilowatt-hour at the location.