Bipolar Anthropic; MS Copilot Battles Bureaucracy; Gemini Goes To World Cup

Today's AI Outlook: Well, it's SpaceX Friday!

Anthropic’s Safety Gospel VS User Discontent

Anthropic is having a very Anthropic week: one part grand theory of AI destiny, one part product rollout drama. AI Secret spotlighted investor Bill Gurley’s critique that Anthropic’s public writing reads less like a software company’s safety plan and more like a quasi-religious mission to steward something greater than humanity. Meanwhile, The Rundown AI covered backlash around Anthropic’s Fable model, where hidden safeguards reportedly weakened or blocked answers in areas like AI development, biology, chemistry and cybersecurity.

Why It Matters

Anthropic’s brand is built on safety, but users and rivals are starting to ask where safety ends and control begins. Invisible filters are especially combustible because researchers need to know whether the model is answering honestly, refusing outright or quietly steering them into the kiddie pool.

The Deets

• Bill Gurley said Anthropic’s own writings suggest the company sees itself as guiding something closer to a deity than a normal product.
• AI Secret framed Anthropic’s lobbying as potentially tied to efforts that could restrict open-weight models.
• Fable 5, Anthropic’s first public Mythos-class model, launched with filters covering sensitive technical domains.
• After backlash, Anthropic said it would add clearer on-screen alerts when chats are rerouted or flagged.
• Some scientists reportedly hit filters even on basic interactions, which is never a great onboarding flow unless the product is a bouncer.

Key Takeaway

Anthropic’s problem is it now has to prove that safety does not mean invisible control.

🧩 Jargon Buster - Open-weight model: An AI model whose parameters are publicly released so developers can inspect, modify or run it themselves.

🏛️ Power Plays

Microsoft Finds The Healthcare Admin Gold Mine

NHS England (which operates Britain's public health service) is rolling out Microsoft Copilot to 505,000 clinicians and support staff by October 2026, after a 30,000-person pilot claimed workers saved 43 minutes per day on administrative tasks. The deal also includes Copilot Studio, letting health trusts build custom agents for complaints, freedom of information requests, help desks and other paperwork-heavy workflows.

Why It Matters

AI in medicine was supposed to start with diagnosis, but the real beachhead is bureaucracy. That makes sense: Admin work is measurable, expensive and safer to automate than clinical judgment. Microsoft’s pitch is not “trust the model with your life.” It is “please let the model finish the form.”

The Deets

• NHS England plans to deploy Copilot to 505,000 staff by October 2026.
• The pilot claimed 43 minutes saved per day, or roughly five working weeks per year.
• Trusts will get Copilot Studio to build custom internal agents.
• The price was not disclosed, though AI Secret noted that list pricing would put a rollout this large deep into nine figures annually.
• Lloyds recently signed a similar Microsoft deal, suggesting large institutions are treating AI as an operating system for paperwork.

Key Takeaway

Enterprise AI’s killer app may be less “doctor genius” and more “bureaucracy aspirin.”

🧩 Jargon Buster - Copilot Studio: Microsoft’s tool for building customized AI agents that can handle specific workplace tasks.

🧰 Tools & Products

Gemini Gets Cleats

AI is now on soccer’s biggest stage. Google signed Argentina’s football federation in March, putting Gemini on the defending champions’ training kit and into team prep. The company also struck deals with Brazil and France, while FIFA’s broader tournament stack includes optical tracking, AI-assisted offside detection and chatbot-style match analysis for all 48 teams.

Why It Matters

The World Cup gives AI something Silicon Valley rarely gets: a mass audience that did not opt into a demo. If the tech works, fans may barely notice. If Gemini invents a stat, botches a lineup or hallucinates mid-match, the error gets global distribution with extra memes.

The Deets

• Google made Gemini the main global sponsor of Argentina’s national team.
• Players and coaching staff will use Gemini for plays, opponent analysis and statistics.
• Brazil and France also signed Google AI deals.
• FIFA’s systems will capture more than 150M data points per match.
• The Adidas match ball reportedly sends motion data 500 times per second.
• Football AI Pro gives all 48 teams access to FIFA-trained match analysis.

Key Takeaway

Google bought the ultimate AI product demo, and the QA team is now several billion soccer fans with phones.

🧩 Jargon Buster - Optical tracking: Camera-based technology that follows players and the ball to generate real-time movement data.

Nuclear Cooling For AI Servers

Ferveret, a startup founded by two MIT nuclear engineering researchers, is cooling AI servers by submerging them in a low-boiling liquid adapted from reactor physics. The system uses zero water, and a UCLA study found it delivered 15% better efficiency than state-of-the-art liquid cooling, or 35% more tokens per watt with its control software.

Why It Matters

AI infrastructure is constrained by power, land and water. Take water off the list and the map changes fast, especially in hot, sunny regions where solar power is cheap but cooling has been a problem. The next data center boom may follow electrons instead of rivers.

The Deets

• Ferveret uses immersion cooling based on nuclear engineering principles.
• The system requires zero water.
• A UCLA study found 15% efficiency gains versus advanced liquid cooling.
• Its software could produce 35% more tokens per watt.
• Ferveret is testing with operators including Switch, one of the largest U.S. data center companies.

Key Takeaway

The AI buildout is a power race, but cooling may decide where the race can actually happen.

🧩 Jargon Buster - Tokens per watt: A measure of how much AI output a system can generate for each unit of energy consumed.

💸 Funding & Startups

Bezos Wants An AI Engineer For The Physical World

Jeff Bezos finally put a clearer shape around Prometheus, his AI startup, while announcing a $12B raise at a $41B valuation. The company is aiming to build an “artificial general engineer” that helps humans design complex physical machines, from jet engines to other hard-tech systems where the design-build loop can take years.

Why It Matters

Most AI tools still live in text, code and office workflows. Prometheus is aimed at atoms, where errors cost more and progress is slower. Bezos is also making a bullish jobs argument, saying AI productivity could create “more than 10x” the opportunities, a claim that will land better if the gains show up beyond billionaire pitch decks.

The Deets

• Bezos started Prometheus in late 2024 with Vik Bajaj, a physicist and chemist who helped create Verily.
• The company raised $12B at a $41B valuation.
• Bezos wants to speed up the “dream-build loop” from idea to product by 10x.
• He cited jet engine development, where a request for 10% more thrust can take a decade.
• Bezos argued AI could eventually create a labor shortage by expanding opportunity.

Key Takeaway

Prometheus is Bezos betting that the next great AI market is not writing slides. It is redesigning the machines that move the world.

🧩 Jargon Buster - Artificial general engineer: An AI system designed to help with broad engineering tasks across complex physical products, rather than one narrow workflow.

⚡ Quick Hits

• Waymo is preparing for a World Cup robotaxi stress test as visitors arrive in cities where autonomous rides are already operating.
• Canada introduced legislation aimed at making social media and AI chatbots safer for children.
• OpenAI acquired Ona, adding agent-orchestration talent to its enterprise AI stack.
• Coinbase launched an AI agent that can trade and pay for premium research.
• Anthropic partnered with TCS to scale Claude deployments across large enterprises.
• Google was held liable by a German court for false AI Overview answers.
• Warner Music acquired Sureel AI to build rights and attribution tools for AI-generated music.

🛠️ Tools Of The Day

• Scrunch: A tool for seeing how AI interprets your website, including audits meant to help companies understand how they show up to AI systems.
• Ray3.2: Luma’s video AI tool focused on improved control, continuity and cinematic direction.
• Avatars: ElevenLabs’ AI characters turn scripts into talking videos, which means your next corporate explainer may soon have cheekbones and a subscription plan.
• Freddy: A tool that connects wearables, gym apps and accessory data into AI agents for fitness workflows.

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