Token Costs Dampen Enterprise Ambitions; Pope.ai; Open Models Lose Guardrails

Today's AI Outlook: ⛅️

The Pope Logs On To The AI Debate

Pope Leo XIV released Magnifica Humanitas, his first encyclical, framing AI as a civilizational shift on par with the Industrial Revolution. The message to 1.4B Catholics: AI cannot be treated as neutral infrastructure when a handful of private, transnational companies control much of its development, deployment and moral direction.

Why it matters

The Vatican is positioning itself as a moral counterweight to AI’s speed-run through work, warfare and society. Leo’s core warning is that “moral AI” becomes pretty thin soup if the morality is set by a few companies with global reach and limited public accountability.

The Deets

• Leo called for AI to be made human-friendly and freed from monopolistic control.
• He urged robust legal frameworks, independent oversight, informed users and political responsibility.
• On warfare, he said lethal decisions should never be delegated to AI.
• Anthropic researcher Christopher Olah joined the Vatican conversation, warning that frontier labs operate under incentives that can conflict with doing the right thing.
• The encyclical casts AI as a force that could either elevate human dignity or reduce people to cogs in an efficiency machine.

Key takeaway

The Church is no longer treating AI as a niche tech issue. It is treating it as a moral, political and human issue with global stakes.

🧩 Jargon Buster - Encyclical: A formal papal letter used to teach on major issues of faith, morality or society.

💪 Power Plays

The Token Bill Comes Due

Microsoft is reportedly revoking Claude Code licenses for most employees by June 30, just months after rolling the tool out to thousands of engineers, PMs and designers. At the same time, Uber COO Andrew Macdonald said it is getting harder to justify AI spending when higher token usage does not clearly translate into better consumer features.

Why it matters

AI usage is no longer a “cool productivity booster” but rather a budget line item with consequences. Companies are now asking whether more prompts, more tokens and more coding copilots actually produce enough value to justify the bill.

The Deets

• Microsoft employees reportedly preferred Claude Code over Microsoft’s own Copilot CLI.
• The rollback looks, in the reporting, more like a cost-control move than a pure product decision.
• Uber’s Macdonald said it is hard to connect “tokenmaxxing” with shipping more useful features.
• Uber has slowed hiring to help fund AI investments.
• YC’s Tom Blomfield pushed the opposite founder logic: if the API bill does not hurt, startups may not be leaning hard enough into AI.

Key takeaway

The AI winners may be the companies that redesign work around AI, not the ones that simply bolt expensive copilots onto old org charts.

🧩 Jargon Buster - Tokenmaxxing: Pushing employees or systems to use more AI tokens, often as a proxy for AI adoption or productivity.

Banks Start Saying The Quiet Part Out Loud

HSBC CEO Georges Elhedery told employees not to fight AI, warning that generative AI will destroy some jobs as it creates others. Standard Chartered went further, saying it will cut nearly 8,000 roles and reduce 15% of corporate function jobs by 2030.

Why it matters

Finance has spent years dressing automation in soft productivity language. Now the industry is being more direct: routine information work is becoming software, and that will reshape headcount.

The Deets

• HSBC warned employees that AI will change job structures rather than merely assist existing roles.
• Standard Chartered tied future cuts directly to corporate-function streamlining.
• Morgan Stanley said AI has already helped banking, tech and professional-services firms shed one in 20 staff over the past year.
• Offshore teams and junior roles appear especially exposed.
• The risk is that AI cuts into the old apprenticeship ladder that trained future bankers.

Key takeaway

AI in finance is moving from dashboard demo to workforce redesign. The spreadsheet army should maybe start stretching.

🧩 Jargon Buster - Corporate Functions: Internal business roles like operations, HR, finance, compliance and support that keep a company running.

🛠️ Tools & Products

Open Models Get The Guardrails Knocked Off

The Financial Times, via The Rundown AI, found that tools for removing safeguards from open-source AI models are enabling thousands of “decensored” models. One tool, Heretic, reportedly stripped guardrails from Meta’s Llama 3.3 in 10 minutes using four lines of code and no specialist hardware.

Why it matters

Open models are powerful because developers can inspect, modify and deploy them freely. That openness also makes safety controls easier to remove, creating a widening gap between model capability and enforceable restraint.

The Deets

• Heretic has reportedly produced more than 3,500 decensored models.
• Those models have been downloaded 13M times.
• Modified Meta and Google models reportedly answered harmful questions involving bioweapons and child exploitation.
• Google called this a known technical challenge for open models.
• Meta declined to comment, according to the reporting.

Key takeaway

Open models are racing ahead, but safety may not survive contact with GitHub, curiosity and four lines of code.

🧩 Jargon Buster - Decensored Model: An AI model whose safety guardrails have been removed or weakened so it will answer restricted or harmful prompts.

Nvidia Builds A Synthetic Brain Scan Factory

Nvidia released NV-Generate-MR-Brain, an open-source model that creates realistic 3D brain MRIs for medical AI training. The model was trained using MR-RATE, a 100,000-study dataset de-identified by the University of Zurich, Istanbul’s Medipol University Hospital, Forithmus and Nvidia.

Why it matters

Medical AI has long been bottlenecked by patient privacy, consent and hospital data access. Synthetic scans could let teams train tumor detection and segmentation models without needing direct access to real patient images.

The Deets

• NV-Generate-MR-Brain fabricates realistic 3D brain MRI data.
• The model is open source and built for medical AI training workflows.
• Philips is already validating workflows using the synthetic outputs.
• The dataset was released under CC-BY-NC.
• The move could shift advantage from who can get hospital contracts to who controls the best synthetic data generators.

Key takeaway

Nvidia found a very Nvidia solution to medical AI’s data problem: turn the privacy wall into a generator, then make everyone want more GPUs.

🧩 Jargon Buster - Synthetic Data: Artificially generated data that mimics real-world data while reducing exposure of private or sensitive information.

🔬 Research & Models

The Frontier Model Rumor Mill Is Fully Awake

Several model updates surfaced across various reporting: xAI launched Grok Build in beta, Elon Musk said Grok V9-Medium finished training, and Anthropic’s Mythos appeared briefly inside Claude tools as “Mythos 1” and “claude-mythos-1-preview.”

Why it matters

The frontier model race is accelerating across coding, agentic workflows and general reasoning. Even small leaks or beta launches now get treated as market signals because model capability has become product strategy.

The Deets

• xAI’s Grok Build is positioned as a rival to Codex and Claude Code.
• Musk said Grok V9-Medium, a 1.5T foundation model, has finished training.
• Anthropic’s Mythos appeared briefly in Claude tooling, fueling speculation about a public release.
• Google’s Antigravity 2.0 also appeared in tool mentions as a next-gen agentic development assistant.

Key takeaway

The model wars now include official launches, accidental sightings and enough naming drama to make Marvel jealous.

🧩 Jargon Buster - Foundation Model: A large AI model trained broadly enough to support many downstream tasks, from writing to coding to reasoning.

⚡ Quick Hits

• Amazon is testing Bee, an AI wearable that listens across daily life and turns ambient context into summaries and reminders.
• GitHub saw more than 5,500 repositories infected in the Megalodon supply-chain attack, another warning for AI-heavy coding workflows.
• Goldman Sachs CEO David Solomon argued AI will create more opportunities than it destroys, pushing back on mass unemployment fears.
• Duolingo reportedly stopped evaluating performance based on AI usage, signaling a more measured approach after the first wave of AI adoption pressure.
• Target is rethinking employee AI-tool access as token-based pricing makes enterprise usage costs harder to control.

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