Pope Leo Denounces Big Tech's “Culture of Power”

Pope Leo has denounced the “culture of power” that has arisen around AI, warning that technology must be subject to the “most rigorous” ethical constraints.

• In a document entitled Magnifica Humanitas, he wrote that power over digital systems and infrastructure currently “does not rest with states but with major economic and technological actors”.
• When this power is concentrated “in the hands of the few”, it can become “opaque and evade public oversight, increasing the risk of distorted forms of development that give rise to new dependencies, exclusions, manipulations and inequalities”.
• The Pope also called for the “disarming” of AI, explaining that “To disarm does not mean rejecting technology, but preventing it from dominating humanity”.
• Last year, Pope Leo said he considered AI to be the biggest threat to humanity today.

This MIT Technology Review looks at Google’s AI strategy after the recent Google I/O conference.

• The company announced the Gemini for Science package which collects several of the companies LLM models under a single brand, including the hypothesis-generating AI Co-Scientist and algorithm-optimizing AlphaEvolve, even though these tools are still not publicly available.
• John Jumper, who won the Nobel Prize two years ago for AlphaFold, is now working on AI coding. This reflects Google’s worry that the company’s coding tools are not as good as models from Anthropic and OpenAI.
• The article reports that protein structure predictions from AlphaFold have been used by over three million researchers worldwide. A Google subsidiary called Isomorphic Labs that uses AlphaFold for drugs development has raised 2 billion USD in a Series B funding round.
• The author remarks that two different approaches to AI are being followed at Google. The first is the development of AI for specific domains (e.g., AlphaEarth, AlphaFold, WeatherNext). The second is the development of AI co-scientists to help the ideation and validation of research.

This VentureBeat article discusses new forms of technical debt that come with AI.

• Technical debt is the implied future cost of applying an easy or quick short-term modification to code. Changes can make code more messy or cause the architecture to become harder to maintain in the future.
• The importance of debt has increased in relation to AI. A 2025 MIT study found that 95% of AI projects fail to reach production and an S&P Global Market study showed that 42% of businesses scrapped AI initiatives in 2025 – a 17% increase on 2024.
• Traditional technical is confined to the code of the IT system. It can be identified through testing and repaired through refactoring. AI debt is more distributed since it can appear in prompts, model choice and data pipelines. It can also be intermittent due to the probabilistic nature of AI. All of these factors make AI debt harder to remove.
• Concretely, the first type of debt relates to prompts. These can be tweaked or stuffed (where extra context gets added to the prompt). This can have the effect of turning prompt into spaghetti code. One solution is to apply code versioning techniques to prompts.
• Model dependency debt arises when a mixture of external models are used or when access to models is built on API calls. Updates to models by the provider can hurt the application’s performance as prompts tuned for one model version could become less effective with another version.
• Retrieval debt can arise in retrieval-augmented generation (RAG) systems when the repositories used have messy data, duplicated documents, or outdated information. This can lead AI applications to return technically correct data, but which is outdated or irrelevant.
• Evaluation debt for models comes from a lack of standardization in testing and monitoring for AI models and applications. Most organizations do not have consistent testing standards or ground truth datasets.
• The article argues that explainability should be included by default to help compensate for lack of testing standards and limited reproducibility. Documentation should explain data lineage and models used.

This article outlines some of the motivations for Big Tech sponsoring or collaborating with sports teams. For instance, IBM, Anthropic, AWS and Oracle are collaborating with Formula 1 driving teams.

• Formula 1 is an extremely technical sport. Not only does it take 24 people working simultaneously only two seconds to change a tyre, but millions of data points relating to the car and driver are created in every second of a race. AI and data analytics can provide a competitive edge.
• The data can also be used to create a more immersive experience for fans. The Scuderia Ferrari HP team has a Head of Fan Development looking into this. The popularity of F1 has increased in the last few years, partly thanks to a Netflix series. Statistics show that 75% of new fans are women, many of whom are Gen Z.
• Partnerships by Big Tech with sports teams is also part of a marketing strategy by Big Tech to convince the general public of the pertinence and attractiveness of their technologies.

This InfoWorld article discusses the challenges of cloud environments to executing AI workloads.

• The core issue is that AI workloads are being deployed to cloud environments that predate agentic AI. All organizations surveyed by the journal report the need to migrate at least 25% of their data for reproducible model operations, standardized pipelines, and consistent policy enforcement.
• In essence, existing cloud platforms are optimized for application deployment, not governed agentic AI. This architectural mismatch is creating friction.
• Data migration is needed generally for governance due to regulation like the GDPR. When personal data is used in AI workloads, engineering teams need to retrofit access controls onto platforms that were not purposed for this.
• Many organizations today still see the AI infrastructure questions as a build-versus-buy decision. The real decision should be around architectural fit to governance requirements.

Anthropic has raised 65 billion USD in its latest funding round at a post-money evaluation of 965 billion USD.

• The main investors include Samsung, Sequoia Capital, Altimeter Capital, and Amazon (5 billion USD).
• One investor reportedly pledged 5 billion USD just to get a meeting with Anthropic’s CFO Krishna Rao.
• The funding round is rumored to be the last round before the company goes public.
• Anthropic says the funds raised will be used to “advance our safety and interpretability research, expand compute to meet growing demand for Claude, and scale the products and partnerships our customers rely on.”.
• The company also announced that its run rate revenue exceeded 47 billion USD this month. The company expects a revenue increase of 130% which could allow the company to have its first operating profit.

This MIT Technology Review article reports that despite fears of AI taking over many jobs, there is no evidence of this in the US – for the moment.

• The US Bureau of Labor Statistics reports that there is no evidence of people switching from jobs exposed to AI to jobs less exposed. At university level, there is increased interest in AI-related fields like data science and cybersecurity. Many artificial intelligence majors have appeared.
• The Federal Reserve Board has found that annual employment growth for coders has slowed by about 3% since ChatGPT appeared. However, overall employment for coders has continued to grow, though at a slower rate than before 2022. Wages in the coding sector have grown significantly.
• For a former commissioner of the Bureau of Labor Statistics, the real question is the speed of any job market disruption. “If it happens at the normal pace of technological change, labor markets will have time to adapt. If there is a sudden and severe disruption, then that will be a big challenge for policymakers”.
• In 2016, Geoffrey Hinton said that “people should stop training radiologists” because it was “completely obvious” the occupation was soon to be replaced by AI. Their are more radiologists than ever today despite the widespread use of AI. Human radiologists perform a variety of tasks that AI is poor at such as interpreting results and talking with patients.
• One fear is that young employees are the most vulnerable to AI job replacement. Stanford researchers found jobs where tasks could be automated with the help of AI accounted for the largest decrease in employment, e.g., software developers. Headcounts have grown in jobs where AI was mainly used but to augment human work, even for entry-level workers. Older workers, with more so-called tacit knowledge, are less susceptible to AI replacement.

A study has found that data centers in Ireland have used 22% of the country’s electricity. This figure compares to 6% in the US.

• The impact has been an increase on consumer electricity bills. The average household annual electricity bill could rise by 295 EUR to 644 EUR by 2034.
• One ecology campaigner said that “Even Trump, under intense pressure from voters, has acknowledged that big tech should pay its own energy bills”. He added: “Unless data centers are required to be powered by additional renewable energy, they could lock Europe into volatile and expensive fossil gas.”.
• The chair of Digital Infrastructure Ireland said data centers had injected 18 billion EUR into the Irish economy in recent years. A data center spokesman said that the data centers paid grid network charges and commercial electricity costs proportional to their usage.
• The main problem for Ireland is that natural gas accounts for half of the energy source of the electricity grid. This exposes prices to global supplies of gas.

The existing cloud infrastructure was designed for human users of the Internet who click, scroll and stream in a steady and predictable manner.

• AI agents however have different activity profiles. The exhibit bursts in which they query hundreds of databases and documents, and make API calls, within seconds and then quickly disappear.
• AWS has launched its next generation of OpenSearch Serverless that is specifically designed for agent workloads. It includes a fully managed search and vector database that can scale quickly to agent activity bursts. An important technical change is that compute and memory offers are becoming disentangled so compute bursts can be handled even more quickly.
• AI agents currently represent a very small portion of Internet activity. However, machine-generated activity is increasing. Cloudflare says that 31% of all Web traffic is due to automated bots. They expect machine-generated traffic to exceed human-driven traffic in 2027.
• The article highlights other activity in the cloud industry. Microsoft has released updates to Azure to handle agent activity bursts and for agent memory sharing. Databricks and Snowflake are repositioning their offers towards AI memory and enterprise data retrieval.

OpenAI announced a breakthrough on a mathematical geometry challenge set down by mathematician Paul Erdős 80 years ago.

• Described as “one of Erdős’ favorite problems”, the planar unit distance problem asks if n points are placed in a data plane, then how many pairs of points can be exactly 1 unit apart.
• Until now, a method called the “square grid” approach was believed to be the optimal solution to this problem. However, an OpenAI model has come up with an infinite set of examples that offer polynomial improvement on this existing method.
• The Fields medalist Tim Gowers called the result “a milestone in AI mathematics”.
• The breakthrough is notable mainly because it was made using a general-purpose AI model, and not a model trained for mathematical or geometry problems. It is also the first time that a prominent open problem in mathematics has been solved autonomously using AI.
• The OpenAI results have been validated by independent prominent mathematicians.