Leaders Reject AI Slowdown; Corps Reject Anthropic Retention; Smarter Siri Arrives

Today's AI Outlook: 🌩️

The Pause Button Has No Political Backing

President Donald Trump and China’s Foreign Ministry have rejected Anthropic CEO Dario Amodei’s proposal to slow the development of the most advanced AI, according to The Rundown AI. The appeal, backed by Sam Altman and Elon Musk, ran into different objections in Washington and Beijing: Trump dismissed catastrophic AI risks, while China criticized the proposal’s push to restrict its access to leading chips.

Why It Matters

A coordinated slowdown needs governments willing to cooperate. With both rejecting the proposal, AI companies face pressure to explain what limits they will adopt themselves.

The Deets

• Trump characterized fears of AI destroying humanity as a hoax.
• Beijing objected to restrictions on advanced chips and criticized confrontation over AI.
• Amodei also told CBS he would consider handing AI governance to an appropriate group of governments to reduce concentrated power.

Key Takeaway

The proposal has prominent industry supporters, but government buy-in remains a major obstacle.

🧩 Jargon Buster - Frontier AI: The most advanced AI models under development, where new capabilities can bring unfamiliar risks.

♟️ Power Plays

Anthropic’s Logs Become a Dealbreaker

Nvidia, Palantir and Booz Allen Hamilton are reportedly restricting Anthropic models on sensitive work because of 30-day usage-log retention. AI Secret reports that Nvidia moved proprietary supply-chain work to its Nemotron models, Palantir sought an irrevocable zero-retention commitment, and Booz Allen barred the models from client cybersecurity work. The dispute puts control of confidential business data at the center of AI purchasing decisions.

Why It Matters

A capable model can still lose a contract when its data policies conflict with a customer’s requirements. Retention terms help determine which work companies are willing to entrust to an outside AI provider.

The Deets

• The trigger: AI Secret reports that Fable 5 introduced mandatory 30-day retention across Mythos-tier models in June.
• Customer response: The reported restrictions concern specific sensitive workloads, rather than a universal ban on Anthropic.
• An opening for competitors: Microsoft is reportedly pitching private servers to concerned clients.

Key Takeaway

For sensitive enterprise work, control of the records can carry as much weight as the model’s capabilities.

🧩 Jargon Buster - Zero data retention: A commitment not to keep specified customer data after processing, subject to the terms of the agreement.

🛠️ Tools & Products

Siri Gets More to Do Than Set Timers

Apple has released Siri AI in an English-only beta with iOS 27, giving its assistant the ability to interpret what is on a screen, draw context from messages, email and photos, and take actions inside apps. The rollout extends across Apple’s major device platforms, with the European Union and China excluded. The central promise is an assistant that can use information already on your devices to complete everyday tasks.

Why It Matters

An assistant that understands your context could reduce the hunting and app switching behind routine tasks. The beta’s usefulness will depend on how reliably it handles that access.

The Deets

• Personal context: Siri can use Messages, Mail and Photos, with actions inside apps including WhatsApp and Audible.
• Processing: The newsletter describes models built alongside Google’s Gemini, running on devices and Apple’s cloud servers.
• Usage: A dedicated Siri AI app syncs chats across devices; cloud features have daily limits, with extra usage purchases planned.

Key Takeaway

Siri’s expanded access makes it worth watching, while the English-only beta and regional exclusions limit who can try it.

🧩 Jargon Buster - On-device AI: AI processing performed on your phone or computer, rather than on a remote server.

🔬 Research & Models

DeepSeek Trims the Model, Raises the Score

DeepSeek’s V4.1 Flash is outperforming its larger V4-Pro model on the company’s own coding tests, according to AI Secret. The new model has 552 billion parameters, compared with V4-Pro’s 1.6 trillion, and activates a smaller subset for each token it processes. The newsletter reports that V4-Pro requests began redirecting to Flash on Sept. 14, making the efficiency push a change customers encounter in everyday use.

Why It Matters

Better results with fewer active parameters could make repeated coding and agent tasks more economical. The reported test result is promising, but it does not establish superiority across every workload.

The Deets

• Active size: Flash activates 8 billion to 16 billion parameters per token, versus V4-Pro’s 49 billion.
• Coding scores: AI Secret reports 74.2 for Flash and 62.7 for V4-Pro on DeepSeek’s own tests.
• Evidence limit: Those are company-reported results; the newsletter does not provide an independent comparison.

Key Takeaway

DeepSeek’s reported gains make efficiency worth watching alongside raw model size.

🧩 Jargon Buster - Parameters: Numerical values a model learns during training. Some models activate only part of their total parameter count for each piece of text they process.

⚡ Quick Hits

• Boston Dynamics’ IPO may have to wait. A senior Hyundai executive told Reuters a listing next year is unlikely, according to AI Secret, which reports a $393M loss in 2025 and says Atlas has yet to deploy at scale.
• Microsoft has a proposed house rulebook. Its 38-page draft code puts rules for in-house models above user requests, requires acceptance of shutdown and rejects simulated feelings; public comment is open for six weeks.
• OpenAI reportedly bought Glass Imaging for more than $300M. The startup uses AI to improve smartphone photos, and its founders previously worked on Apple’s Portrait Mode.
• AI hardware is reportedly getting pricier. AI Secret says Nvidia raised AI server prices by more than 15% as high-bandwidth memory supplies tightened.

🧰 Tools of the Day

• ELI5: A Claude Code skill for tailoring explanations of complex topics to a particular audience. The Rundown AI's guide also describes comparing its answers with a no-skill baseline.
• SWE-2: Cognition’s coding model claims advanced agent performance at lower inference cost, according to AI Secret. The brief mention gives no independent benchmark or pricing details.
• Fusion: Devin’s setup pairs a frontier planning model with a cheaper coding model, assigning the planning and implementation work to separate models.