POLICY

 Pope's AI manifesto reframes the conversation

Pope Leo XIV's widely expected manifesto on AI delivered a very clear set of recommendations that some call naive, while others characterize as courageous. 

On Monday, the Pope's 44,000-word encyclical titled Magnifica Humanitas (Magnificent Humanity) was released to the public in eight languages, and the Pope made the rare move of summarizing it in a formal presentation that was livestreamed from the Vatican. 

What the Pope delivered was perhaps broader in scope than expected and showed greater technical literacy. He has been working with a study group on AI and reportedly consulted with scientists, technologists, theologians, moral philosophers, researchers, and business executives. Leo XIV is also the first pope known to personally use a smartphone and an Apple Watch, and the first pope from America, where he earned a bachelor's degree in mathematics from Villanova. 

When he was elected pope a year ago, he intentionally chose the name Leo XIV to associate himself with Leo XIII, who famously confronted the Industrial Revolution with his 1891 encyclical, Rerum Novarum (“Of New Things”), calling for the protection of workers and the need for collaboration between capital and labor. It became a catalyst for the labor movement. Leo XIV has publicly compared the current AI boom to the Industrial Revolution and has been building toward confronting the AI industry since he took office a year ago.

Here's a summary of the most prominent recommendations in his manifesto:

• AI for lethal force should be outlawed: The encyclical warns against autonomous weapons and AI-driven military decisions, arguing that life-or-death choices should never be handed to algorithms. 

• AI should not concentrate power: Unlike past technologies that governments controlled, today private tech giants control data, algorithms, and digital infrastructure in ways that can shape politics, culture, and even people’s understanding of truth. He calls for practical regulation, oversight, transparency, and treating data as a shared resource rather than something owned by a handful of corporations.

• Protect truth, kids, and education: Disinformation, deepfakes, algorithm-driven outrage, and addictive content are eroding democracy and harming young people's mental health. He urges parents, schools, and governments to form an educational alliance, push for age limits, and hold platforms accountable.

• Human dignity should be valued over efficiency: The ultimate solution Leo advocates is a society built on justice, solidarity, truth, care for the vulnerable, and the deepening of human relationships. He says technology should strengthen humanity, not weaken it.

• Humanity should recognize its choice: The Pope framed this moment as a decision between building technology for profit, pride, and dominance or building it together for the common good. He emphasized that technology isn't neutral, but reflects the values of the people who design, fund, and use it.

The most prominent member of the tech community to attend the event was Anthropic co-founder Christopher Olah, a self-described atheist. Olah also addressed the audience after the Pope, saying, "We need more of the world—religious communities, civil society, scholars, governments, and indeed all people of good will—to … take this seriously, to look closely, and to push events in a better direction. We need informed critics who will tell the labs when we are failing. We need moral voices that the incentives cannot bend."

Pope Leo XIV's recommendations on how we can best shape AI to serve the interests of all humanity are clearer and more comprehensive than anything that has come from AI labs, governments, or nonprofit organizations. Certainly, no one I know expected that out of this document. It follows OpenAI's call for an "AI New Deal," in which it advocated for broader public discourse on the future of AI. And just last week, Anthropic gathered leaders from 15 religious and cross-cultural groups to discuss the moral foundations of AI. I'll admit my skepticism toward the actions of OpenAI and Anthropic, since they are also trying to win over a skeptical public. As Anthropic's Olah mentioned, we need more voices in the important conversations ahead about governing AI. We can't afford to have the debate dominated by those whose primary incentives are financial.

IN PARTNERSHIP WITH LAMBDA

How to push Model FLOPS Utilization past 60%

Most large-scale training runs operate at 35-45% Model FLOPS Utilization (MFU). You're paying for more than twice the compute you're using.

Lambda's engineers benchmarked Llama 3.1 models from 8B to 405B on NVIDIA Blackwell GPUs and traced efficiency loss to its root causes:

• Memory overhead

• Parallelism strategy

• Serialized communication

The result: a reproducible framework that pushed MFU past 60% (25%+ improvement vs. industry benchmark) with no changes to model architecture.

Get the guide.

PRODUCTS

Inside the design of ChatGPT's personality

With over 900 million users, OpenAI’s flagship chatbot keeps getting more integrated into our daily lives. But most people don't realize that ChatGPT doesn’t need to be the same for everyone. 

ChatGPT's Personalization feature has been evolving since it first launched as Custom Instructions in 2023, giving users the ability to change how ChatGPT responds by being more professional, steering the model toward a warmer tone, or simply telling it to stop using emojis. 

Today, those controls live in ChatGPT's Personalization settings, where users can choose from preset traits like Friendly, Candid, or Cynical. They now also live in the Memory settings, where you can set a nickname, share your occupation, write custom instructions for how you want the model to respond, and toggle memory to let ChatGPT carry context across conversations.

The Deep View sat down with OpenAI’s Laurentia Romaniuk, product manager for model behavior, to learn how ChatGPT’s personality is formed and how users can customize it. 

User context is vital to making AI more useful. Without an understanding of a user’s data, these models can’t provide the most useful outputs. Personality preferences now represent part of that context, said Romaniuk. 

For instance, if a user wants a thought partner for brainstorming, they may lean towards warmer interactions. If a user simply wants a work assistant, they may lean towards a more concise and professional tone. Users can also get creative with exactly how precise they want their instructions to be, she said, such as configuring the model to use “highly specific communication styles” for writing, coding, or learning, or getting it to replicate exactly how they themselves act. 

“What stands out most to me is less the novelty itself and more how personal these preferences can be,” Romaniuk told The Deep View. “People often want the model to adapt to the way they think, learn, or communicate.” 

Romaniuk noted that adjusting ChatGPT’s personality doesn’t create a different system for each user. At its foundation, the model is the same for everyone. And by default, Romaniuk said that the goal is to offer an experience that’s “broadly useful and adaptable,” providing a model that’s palatable to as many people as possible. 

However, there is no “single perfect personality,” she said. Because everyone reacts differently to the way that the chatbot responds, “What feels concise and efficient to one person can feel cold or robotic to someone else,” she said. Small wording changes, for example, can elicit strong user reactions. 

The challenge is that the OpenAI team has to walk a fine line between warmth and clarity, without making the model sound overly deferential, verbose, or unnatural. 

In the future, ChatGPT’s personalization features will become more intuitive, she said. The goal is for people to eventually need to do little to configure the chatbot into their ideal conversation partner, making it more “naturally adaptive” while keeping transparency and user control at the forefront. 

“The goal is to build AI that feels genuinely helpful without compromising reliability, trustworthiness, or user agency,” Romaniuk added.

Personality can play a significant role in the ways we interact with chatbots. It's why people flock to platforms like Character.ai to get the personality they choose to interact with. OpenAI itself received backlash for retiring GPT 4o, with many users expressing that they preferred the warm tone of that model. But chatbots’ personalities can also have an impact on users themselves, sometimes in dangerous ways. OpenAI has said in its Model Spec that its goal is to allow users the freedom to create with AI while keeping guardrails up to “reduce the risk of real harm.” And while OpenAI may do everything within its power to prevent the worst from happening, the safest safeguards might be beyond the scope of any one company. Rather, the solution may be a combination of regulation and education.

IN PARTNERSHIP WITH GUILD

Every engineer is shipping agents. No one is governing them.

Shipping agents is easy. Knowing what each agent can access, what they did, how much they're spending, or who's responsible? Not so much.

Guild is the control plane for AI agents — a unified platform to govern permissions, track actions, control spending, and manage all your agents safely in production.

Bring your own models. Plug in your own tools. And fork and deploy any agent in minutes to share in the spoils. 

Guild is the control plane. The story is yours. Available now.

Get a demo

CONSUMER

Inside Google’s AI glasses strategy

AI smart glasses have quietly built momentum over the past several years, promising to put artificial intelligence directly in your line of sight. Now, Google has shown us the final product of its first AI glasses, which appear to have a clear competitive edge.

Juston Payne, Google's director of product management for XR, joins The Deep View Conversations straight from Google I/O, where the company pulled back the curtain and gave the world a first look at two of the pairs of glasses that will lead the collection when they launch in the fall: a pair from Gentle Monster and one from Warby Parker. 

Juston discusses how the AI smartglasses came to be, including the collaboration between Samsung, Google, Warby Parker, and Gentle Monster. In addition to discussing details of the new launch, including design, product choices, functionality, the roadmap, and more, Juston also sheds light on the broader AI glasses market and why people should give them a shot. 

Topics covered:

• The thought put into the aesthetics and comfort of smart glasses

• What products will be available for users to purchase at launch

• How the glasses act as an equivalent of a touchscreen on a phone for interacting with Gemini

• The computation offloading strategy that leverages the user’s smartphone 

• The choice to first launch with an audio-only product rather than in-lens displays

• How Google is approaching privacy concerns with the cameras on the glasses

• Real-world use cases for AI smart glasses

If you want to understand how AI glasses are reshaping the way people connect, this conversation will leave you much more knowledgeable about Google's strategy.

📺 Watch on YouTube
🎧 Listen in your favorite podcast player

Subscribe to Deep View Conversations for interviews with the leaders shaping the future of AI, business, and technology.

LINKS

• Reflection AI partners with Department of Energy on Genesis Mission

• Yusuf Mehdi, Microsoft’s longest serving executive, to leave after 35 years

• DeepSeek founder pledges to focus on AGI over commercialization

• AMD CEO Lisa Su expects CPU demand to grow 35% by 2031

• US government approves $9 billion for AI chips for CIA and NSA

• Anthropic shares initial findings on the security impact of Mythos

• Students cheer Apple cofounder Steve Wozniak for new definition of AI

• Runway Aleph 2.0: The video AI company’s editing tool that allows for frame-by-frame editing. 

• Perplexity Bumblebee: The AI answer engine open sourced its read-only scanner for macOS and Linux. 

• Meta Forum: Meta released an AI-assisted clone to Reddit, calling it a dedicated space for “deeper discussions, real answers and communities you care about.” 

• Nvidia AI-Q: Nvidia released an open source tool to give specialized deep research skills to agents

• Google DeepMind: Research Scientist/Engineer, Autonomous Security, DeepMind

• TSMC: Senior AI Research Scientist

• Fort: Founding Machine Learning Engineer, Health Algorithms

• SpaceX: AI Engineer, Special Programs - Top Secret Clearance

GAMES

POLL RESULTS

Do you think regulation can help with potential AI job displacement?

Yes (18%)
Somewhat (38%)
No (35%)
Other (9%)