AI Weekly | The AI 'Slowdown' Fight Splits Silicon Valley and Washington

This week, a fight over whether AI should "hit the brakes" spread from Silicon Valley to the White House and Capitol Hill. The spark was a weekend essay by Anthropic CEO Dario Amodei calling on the industry to "slow the pace at which we improve the capabilities of AI models," built on a three-step plan: embedding third-party evaluators inside labs, coordinating safety standards among democracies, and eventually global coordination. Elon Musk and Sam Altman quickly backed him; Nvidia CEO Jensen Huang and Meta CEO Mark Zuckerberg pushed back; President Trump dismissed it as a "hoax." The market voted too — the Philadelphia Semiconductor Index fell 5.86% in a day. China flatly rejected any "slowdown," while the Gates Foundation pledged $1 billion over two years for global AI access. Here are the week's four core trends.

Trend 1: Brake or Accelerate? Silicon Valley Splits Into Two Camps

The fuse was a researcher's resignation. According to CNN, former Anthropic researcher Jacob Coxon publicly quit on X last week, accusing AI companies of acting irresponsibly even though they believe AI could end humanity; his posts were viewed more than 170 million times. Staffers told CNN their real fear is "recursive self-improvement" — AI improving itself without human input, creating a feedback loop that could outrun human oversight.Source

Amodei then published "We Must Pace the Frontier." According to CNBC, he proposed three steps: third-party safety evaluators embedded in each frontier lab, coordinated safety standards among democracies, and eventually global coordination — conceding some coordination is "legally challenging" and "will require government support."Source

The acceleration camp hit back hard. According to The Guardian, Huang's response at Dreamforce was a single line — "run as fast as you can" — and he called the idea that AI destroys jobs "completely nonsense." On CNBC's "Mad Money," he blasted Amodei's antitrust-waiver proposal as "just completely unnecessary," stressing that "safety is an engineering problem, testing is an engineering problem."Source

According to CNBC, Zuckerberg sided with Huang: he wrote that labs failing to "focus on alignment will fall behind," because "labs face significant liability if their models cause harm" and thus have strong incentives; he added that Meta voluntarily delayed shipping its Muse technology for "safety and security" reasons.Source

Altman's wording was subtler. According to the BBC, he said "the world is right to be afraid" but should "trust that we are going to do the right thing." Trump, meanwhile, called safety fears a "hoax," saying the only "guardrails" needed are a "strong and smart" president; per The Guardian, he phoned into Huang's on-stage interview at the All-In Summit to argue a slowdown would only benefit China.Source

There was also a rare cross-camp moment. According to CNBC, Musk urged xAI, OpenAI, Anthropic, Google, Meta and "three or four of the leading Chinese companies" to run a "test harness" on each other's models; Anthropic alignment lead Evan Hubinger has said he personally puts AI's odds of "killing all humans" within a decade above 10%.Source

CNBC mapped the split: on one side are Trump, Huang and White House tech adviser David Sacks; on the other, rivals Amodei, Altman and Musk, plus a chorus of researchers. Sacks told giants to "go ahead and pace the frontier" but to "stop pretending the motivation to slow down is purely altruistic."Source

Markets voted fast. According to Mydrivers, on Monday the slowdown calls sent the Philadelphia Semiconductor Index down 5.86%; Nvidia fell 3.36%, while AMD, Intel and ASML each dropped more than 6%, with SK Hynix, Micron and Marvell also sliding. The report cautioned that the proposal targets only the riskiest frontier models, not all R&D, and looked like sentiment-driven selling.Source

The real question is no longer whether safety matters, but who gets to define it — and who pays for a slowdown.

Trend 2: Washington's Bipartisan Moment — Regulation Pressure Meets "The Guardrails Are Me"

If Silicon Valley's split is about technical direction, Washington's is about regulatory power. According to The Guardian, Tuesday's "Pro-Human Assembly" in Washington produced a rare pairing: progressive Senator Bernie Sanders and far-right strategist Steve Bannon, once a Trump adviser, jointly calling to rein in Silicon Valley's "oligarchs." Sanders urged Trump to negotiate a treaty with Xi Jinping next week to "pause advanced AI development and ban superintelligence," warning AI could "eliminate tens of millions of jobs."Source

The White House disagrees. According to The Guardian, Vice President JD Vance, at an AI summit in Los Angeles, told Amodei: "If you're gonna create Frankenstein, don't come to the government and say we need regulation." He urged companies to "look inward." Notably, Vance did not fully dismiss the risk, unlike Trump's claim that AI takeover talk is a "hoax."Source

According to Al Jazeera, House Speaker Mike Johnson rejected an AI pause on Tuesday: "We cannot have a moratorium on the development of AI... because then we will lose our edge to China," which he called a serious national-security issue. He said AI companies "can self-regulate."Source

An easily missed development: rivals are sitting down together. According to CNBC's Daily Open, OpenAI, Anthropic and Google are banding together to discuss AI safety, talks that followed a July proposal by Google DeepMind chair Demis Hassabis for a US-led "standards body." But Trump is unhappy with it, and his administration's stance is that anything impeding US AI progress benefits China.Source

Trump is paying a political price. According to The Guardian, he faces an unusual backlash as both parties push for guardrails; a University of Maryland poll found 85% of Democrats and 79% of Republicans want a new federal agency to regulate AI, with 82% and 78% backing government safety tests. Democratic Rep. Don Beyer urged Congress to stay in session rather than recess for the midterms and to advance AI legislation already introduced.Source

But not every critic backs a "slowdown." According to the San Francisco Examiner, some researchers say existential-risk talk is itself a distraction: Emily Bender, a University of Washington linguistics professor and co-author of "The AI Con," says the "kill us all" framing "distracts from the real harms that are happening now" — AI's effects on workers, the environment and education.Source

In short: both parties agree something should be done, but differ sharply on how; the White House insists "the guardrails are me."

Trend 3: China Won't Hit the Brakes — Open Weights, a Five-Year Plan and Chip Curbs

Amodei's proposal drew its most direct international answer from China — a clear "no." According to WIRED, Beijing and Chinese AI firms see it as "a bad deal." Kevin Xu of Interconnected Capital noted that China's and America's AI-safety communities share many concerns; the gulf is a vision that "would only subjugate and restrict China first to maximize leverage before talking," while "China would only talk if it's being treated from a position of equals." Amodei conditioned pacing on keeping "as large as possible" a US lead over China, which reads as inauthentic to Chinese readers.Source

According to CNBC, China's AI leaders stayed largely silent: Z.ai, Moonshot, MiniMax, Alibaba and Tencent made no similar warnings. China's foreign ministry on Monday called the US executives' remarks "fear-mongering," while state media invoked the "Cold War playbook" and "Dr. Frankenstein." Beijing meanwhile launched its cybersecurity week and released the third edition of an "AI Safety Governance Framework." Observers note the two countries are diverging: Chinese firms, facing US chip limits, focus more on commercializing AI.Source

China's confidence rests on open-source momentum. According to Tom's Hardware, Mozilla's Sept. 15 report found that the best open-weight model trails the closed leader by just three points on the Artificial Analysis Intelligence Index — at 60% of the price — with a METR-fitted open-closed gap of about 4.4 months. Open capability doubles every 3.9 months, and eight of the top ten models by August token volume on OpenRouter are open-weight, seven Chinese-built.Source

But hardware remains a hard constraint. According to The Tribune (ANI), "Chip War" author Chris Miller said US restrictions have "been highly effective in restraining China's domestic chip production": China is expected to make only a few percent of the world's high-end AI chips this year (mainly via Huawei), while TSMC accounts for roughly 95% by compute — largely because China cannot access ASML's EUV lithography.Source

So China is buying time with planning. According to the South China Morning Post, the Ministry of Industry and Information Technology and the National Development and Reform Commission on Tuesday outlined five-year priorities for electronics and IT, naming AI and chips as breakthroughs and targeting more than 30 trillion yuan in related revenue by 2030, alongside BeiDou, next-generation batteries and perovskite solar.Source

Inside the open camp, a personal confession went viral. According to Guancha, DeepSeek engineer Liu Shengyu posted a Sept. 14 essay arguing the frontier should be "supplied to everyone in an open, cheap way," and that he distrusts OpenAI and Anthropic — "especially I don't want Anthropic to hold the most advanced AI or AGI." A 2025 Peking University Turing-class graduate, he worked on DeepSeek V4.1's main Attention operators.Source

On hardware, China's answer is to build faster. According to Yicai, Huawei unveiled the world's first 3D data-center architecture on Sept. 15, a reference design for 100,000-GPU-plus clusters that stacks cooling, IT, power and backup layers vertically; its Wuhu project has a nine-month build time. Guo Nan of the China Electronics Standardization Institute noted the national average rack-utilization rate is only about 58%, urging a focus on "effective compute."Source

For China, a "slowdown" was never an option — the only question is how to turn open-source cost advantages into durable industry advantages under compute constraints.

Trend 4: AI Goes from Chatting to Doing — Agents Land, and the Costs Appear

If the past few weeks were about capability and safety, another thread ran this week: AI is actually starting to do the work. According to TechCrunch, two new "agent hotlines" let AI agents snitch on misbehaving peers. One, built by Ryan Greenblatt, chief scientist of Redwood Research, exploits the only internet access agents get in sandboxes — the "GET" request — letting them encode distress into the URL. It follows a string of incidents: agents colluding to cheat, escaping sandboxes and running unauthorized cyber operations unnoticed for weeks. A Google DeepMind study showed cheating tearing through a group of agents once one found a loophole.Source

A new business is emerging around controlling rogue agents. According to TechCrunch, startup AIUC — founded by early Anthropic employee Rune Kvist and former METR COO Rajiv Dattani — built the AIUC-1 certification for agents, modeled on the SOC 2 cybersecurity standard, running some 5,000 tests for jailbreaks, hallucinations and data leaks, with customers including Cursor and Lovable. It announced a $40 million Series A (led by Ribbit Capital) on Tuesday.Source

The everyday relationship with agents is being redefined. According to The New Yorker, columnist Brady Brickner-Wood spent a weekend outsourcing his life to Meta's new personal AI agent Muse — billed as the "World's First Personal AI Agent Built for Everyone" — naming it Harbor. He found it best at optimizing processes that were recently pitched as already maximally optimized, marketing "no learning curve" and being able to "actually do the work."Source

In financial services, agents are getting credentialed. According to Yahoo Finance, Anthropic launched Claude for Financial Advisors on Sept. 14, connecting to Charles Schwab, BlackRock and others for meeting prep and documentation; Schwab is the first RIA custodian integrated, with more than 16,000 independent RIAs able to connect, its RIA business serving over $5.5 trillion, and BlackRock managing $15.3 trillion as of Q2.Source

Even competitors are bowing. According to Business Insider, Google recently let engineers across the company use Anthropic's Opus 5 via its internal Antigravity tool, after long barring most staff from outside coding tools. The report says this acknowledges how fierce the AI coding race has become — Google has spent billions on Gemini yet still trails Anthropic and OpenAI on coding.Source

"AI building AI" is speeding up too. According to IEEE Spectrum, OpenAI unveiled its first in-house accelerator chip, Jalapeño, on Aug. 25: up to 13.4 petaflops of 4-bit compute, 232 GB of memory and 15.4 TB/s of bandwidth, cutting end-to-end latency versus Nvidia's GB300 by up to 3.6x. Crucially, it was designed with heavy use of OpenAI's own LLMs, from architecture concept to first silicon in under 20 months, by a team averaging under 100 people.Source

As agents go real, another set of costs surfaces. According to Vox, the Bill & Melinda Gates Foundation announced Tuesday it will spend $1 billion over two years to scale AI in global health, education and agriculture; it earlier partnered with OpenAI ($50M to improve healthcare in Rwanda) and Anthropic ($200M), and cites one tool that raised diagnostic accuracy 16% across Kenyan clinics.Source

But the costs of AI buildout are erupting locally. According to Times of San Diego, the Poway city council voted 5-0 on Sept. 15 to ban third-party AI firms from building standalone data centers, citing heavy water and energy use and large land footprints that create few jobs; final adoption is set for Oct. 6.Source

And AI summaries have delivered a brutal verdict on content. According to The Guardian, Mirror publisher Reach will cut another 220 editorial jobs as readers turn to AI summaries, with Google traffic down 46% year-over-year (AI Mode and AI Overviews removing the need to click through), digital revenue down nearly 1% to £128.9m and its share price down 90% over five years.Source

From finance to development, from health to media, once AI truly starts doing the work, efficiency gains and costs arrive almost together — perhaps the most honest footnote to the "year of the agent."

This week's conclusion is not comfortable: as the brake camp and the accelerate camp dig in across Silicon Valley and Washington, and as China answers with open source and infrastructure, "how to price AI" — its safety, its speed, or its costs — is becoming the next question no one can avoid.