AI Companies Call For Slowdown... Trump Yells 'No'; Ultrasound Gets AI Boost

Today's AI Outlook: 🌩️

AI’s Biggest Labs Reach For The Brake

Anthropic CEO Dario Amodei published a lengthy proposal calling for AI companies to deliberately pace capability improvements so safety systems can catch up. His concern centers on AI increasingly helping develop better AI, potentially compressing years of technical progress into much shorter cycles.

The idea quickly drew support from leaders including Sam Altman, Elon Musk and Demis Hassabis, according to The Rundown AI. Amodei proposed independent evaluators embedded inside labs, shared safety standards among major developers and international coordination around increasingly capable systems.

AI Secret offered a more skeptical interpretation, arguing that competitive pressure is also part of the backdrop. It pointed to increasingly capable lower-cost models such as DeepSeek’s Flash releases and suggested economics could be complicating the business case for expensive frontier models.

Meanwhile, Axios reports President Trump continues to push back, writing on Truth Social: "The only control or 'guardrails' that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT, and the U.S.A. has that, in spades!"

Why it matters

Public agreement among competing labs around tighter evaluations and capability pacing is significant because these companies normally have powerful incentives to ship faster. Any meaningful slowdown, however, would require competitors to cooperate while cheaper models, open development and global competition continue advancing.

The Deets

• Amodei warned that AI is beginning to accelerate parts of its own development process.
• He proposed greater use of third-party evaluations inside frontier AI labs.
• Altman, Musk and Hassabis publicly backed elements of the proposal.
• AI Secret connected the safety discussion with growing price competition from lower-cost models.
• OpenAI also said it would not pursue an IPO in 2026, with Altman citing the current environment around AI safety.

Key takeaway

AI’s leading companies are beginning to talk seriously about pacing capability growth, but technical competition and model economics will make coordinated restraint difficult.

🧩 Jargon Buster - Frontier model: An AI model operating near the leading edge of current capabilities, typically built using enormous amounts of computing power and training data.

⚡ Power Plays

The AI Boom Finds A Very Big Credit Card

The AI infrastructure race is increasingly being financed with debt that does not always appear directly on Big Tech balance sheets.

AI Secret reports that Amazon, Microsoft, Google, Meta and Oracle collectively have $831B in off-balance-sheet lease obligations, while total obligations across five of the world’s richest companies reach $2.13T under the newsletter’s calculation.

One example is Meta’s Hyperion data center financing. A project tied to the facility reportedly borrowed $27.3B, while only $2.37B appeared on Meta’s books, with much of the financing housed through separate entities.

Why it matters

AI infrastructure requires extraordinary amounts of capital for chips, power and data centers. Moving portions of that financing outside corporate balance sheets can preserve financial flexibility, while investors such as insurers and pension funds ultimately absorb more exposure to the AI buildout.

The Deets

• Off-balance-sheet leases across five major technology companies reportedly total $831B.
• Meta remains economically exposed to parts of its Hyperion data center financing despite the structure.
• Big Tech stock buybacks have reportedly fallen sharply as infrastructure spending rises.
• Institutional investors including insurers and pension funds are purchasing some of the debt associated with these projects.

Key takeaway

The AI infrastructure boom is becoming a financing story as much as a technology story, with more of the capital burden spreading into credit markets.

🧩 Jargon Buster - Off-balance-sheet financing: A financing structure where certain obligations are held outside a company’s primary balance sheet, often through leases, partnerships or separate entities.

🛠️ Tools & Products

The Sales Follow-Up Writes Itself... Almost

A workflow highlighted by The Rundown AI shows how sales teams can automatically turn meeting notes into a polished one-page proposal shortly after a call.

The setup combines an AI meeting notes app, Google Docs, Zapier and Gmail. AI extracts details such as the prospect’s company, needs and budget, inserts them into a proposal template and prepares an email draft for review.

Why it matters

This is the type of AI automation with an immediate business payoff. Salespeople spend less time copying notes between systems while prospects receive relevant follow-ups faster.

The Deets

• Meeting notes are stored in a dedicated folder.
• Zapier extracts the relevant information using AI.
• The information populates a templated Google Docs proposal.
• Gmail prepares the follow-up as a draft rather than automatically sending it.
• A Google Sheets step can create an audit log of each automation run.
• Human review remains important. One test reportedly left the AI’s internal commentary about a budget inside the finished proposal.

Key takeaway: Automate the paperwork, but keep a human approval step before anything reaches the customer.

🧩 Jargon Buster - Workflow automation: Software that automatically moves information and triggers actions across multiple applications based on predefined steps.

💰 Funding & Startups

AI Infrastructure Keeps Pulling In Bigger Checks

Money continues pouring into the physical layer supporting AI.

Beyond Big Tech’s growing debt load, Nvidia is reportedly considering a major investment in Anthropic, while Nvidia CEO Jensen Huang said the chipmaker could grow revenue 70% next year, citing visibility into AI infrastructure contracts.

Why it matters

AI’s software boom still rests on an expensive physical foundation. Data centers, accelerators, networking and power infrastructure remain among the biggest constraints on growth, keeping capital providers deeply involved in the AI race.

The Deets

• Nvidia is reportedly discussing a major investment in Anthropic.
• Nvidia says demand visibility supports potentially substantial revenue growth next year.
• Technology companies continue spending aggressively on data-center capacity.
• Financing structures are expanding alongside direct corporate investment.

Key takeaway

Compute remains one of AI’s most valuable commodities, and capital is continuing to chase companies that control it.

🧩 Jargon Buster - Compute: The processing power used to train and operate AI models, usually supplied by large clusters of specialized chips such as GPUs.

🧪 Research & Models

AI Gives Prenatal Ultrasound Second Set Of Eyes

A randomized trial across five hospitals in China found that an AI assistant improved sonographers’ ability to detect certain fetal brain malformations during prenatal ultrasounds.

The system, called PAICS, was trained to identify 10 specific abnormalities in real time. Detection sensitivity increased from 78.6% to 87.3% when sonographers used the system, while false-positive rates remained steady.

Why it matters

The strongest result came from combining AI with clinicians. PAICS performing alone was often less effective, while trained sonographers successfully overrode roughly 60% of its errors.

The Deets

• The trial focused on high-risk pregnancies.
• PAICS looked for 10 fetal brain malformations.
• Detection sensitivity increased by 8.7 percentage points.
• AI-assisted examinations took roughly 40 seconds longer.
• The trial took place entirely in China, limiting how broadly the results can immediately be generalized.

Key takeaway

Medical AI appears most useful here as clinical decision support, giving specialists another signal while leaving judgment with the human examiner.

🧩 Jargon Buster - Sensitivity: The percentage of actual cases a medical test successfully identifies.

Mathematicians Want AI To Show Its Work

Twenty-five Fields Medalists, including Terence Tao, signed a statement raising concerns about the way AI companies use difficult and unsolved mathematics problems to demonstrate model capabilities.

Their argument centers on incentives. Racing to claim that an AI system solved a famous problem can reward speed and benchmark performance while making verification, attribution and deeper mathematical understanding harder.

AI Secret connected the warning with increasingly large-scale AI mathematics experiments, including reported attempts using thousands of agents and enormous token budgets against open mathematical problems.

Why it matters

Mathematics depends heavily on verification. As AI systems generate increasingly sophisticated proofs, the challenge becomes establishing whether the reasoning is correct, original and understandable enough for other researchers to build upon.

The Deets

• 25 Fields Medalists signed the statement.
• The group criticized using unresolved mathematics problems primarily as AI capability benchmarks.
• Researchers emphasized verification, attribution and mathematical understanding.
• Large agent systems are increasingly being applied to difficult open problems.

Key takeaway

AI may accelerate mathematical discovery, but researchers want verification and understanding to keep pace with the raw volume of machine-generated results.

🧩 Jargon Buster - Benchmark: A standardized problem or test used to compare the capabilities of different AI systems.

⚡ Quick Hits

• OpenAI hits the capacity wall: The company temporarily paused new $200 ChatGPT Pro subscriptions after demand for Astra strained infrastructure.
• AI audio scales fast: Pocket FM says its revenue run rate has reached $500M, with AI now powering 93% of its catalog and 99% of newly produced audio.
• Shopify picks up Tailwind: Shopify acquired Tailwind Labs, while Tailwind CSS will remain open source.
• Microsoft gets buried in AI code: Microsoft says AI-generated Edge extensions are arriving faster than its existing review process can handle, prompting more automated quality checks.
• Safety researchers switch seats: Anthropic researcher Joe Benton and Google DeepMind researcher Josh Engels left their companies to join METR Evals, with both raising concerns about AI capabilities advancing faster than safety work.
• China pitches shared AI development: At the BRICS Summit in New Delhi, China proposed an open-source AI community focused on collaborative development, training and industrial applications.

🔧 Tools Of The Day

• Smaug Flash: Abacus AI’s open-weight DeepSeek Flash fine-tune is positioned as a lower-cost model option, with The Rundown AI reporting pricing roughly 3x cheaper than DeepSeek Flash.
• SWE-2: Cognition’s new coding model runs inside Devin and is aimed at more capable autonomous software development workflows.
• Music v2.5: ElevenLabs’ latest music-generation model targets richer melodies and longer-form song creation.
• Muse: Meta’s newly highlighted personal AI agent is designed around persistent, ongoing assistance rather than one-off prompts.
• Oracle Database 26ai: Oracle highlighted a technique for giving AI writing agents multiple memory layers so they can learn a user’s style over time instead of recreating a voice from scratch.

Today’s Sources: The Internet, The Rundown AI, AI Secret