FIFA World Cup Makes Extensive Use of AI

A union of nurses at New York’s Montefiore hospital has strongly criticized management for laying off 12 nurses and replacing them with AI.

• The union is not against the use of AI, but it is worried that the AI replacing the nurses is untested. It has created an AI bill of rights for patients and advocates for stronger guardrails for AI usage.
• One representative said: “We’re not against technology. There are several advances in healthcare utilizing technology. The issue is with new tech without evidence.”.
• Management has defended its position, saying “What is true is that we are always investing in new technology to ensure the best care and outcomes for our patients and will continue to do so for the betterment of the people we serve.”

This VentureBeat article looks at the impact of the redesign of the Google Search page.

• Users can now upload images, videos, files and the content of Chrome tabs into the search box, as well as hold an AI-driven search conversation.
• Another feature is “generative UI” which is the ability for Search to build custom interactive widgets that link to data on users’ queries.
• Also proposed are “information agents” which are AI agents that users instruct to monitor the Web 24/7 for specified information.
• Google’s usage statistics show that AI Mode has surpassed one billion monthly users and the summaries tool, AI Overviews, has 2.5 billion monthly users.
• Google's Gemini 3.5 Flash model is currently the engine of its AI search. The technical challenge has been to make model inference as instantaneous as traditional keyword search.
• The changes have implications for publishers because the engine is increasingly “zero-click”, meaning that a publisher’s content can be given to a user via AI Overviews without the user visiting the publisher’s web site. Lower visits to a publisher’s web site leads to lower revenue.
• There is also a change coming for advertisers because conversational search can indicate more precise “intent signals” that can be exploited for targeted ads.
• Finally, the article notes how Google's Search box had become a cultural icon over the last 25 years, training people to compress their curiosity into a few keywords. This era is over and the “think out loud” approach is the new search paradigm.

Over 200 researchers and economists have signed an open letter to governments and technology leaders calling on them to urgently create policies and institutions that address the economic impact of AI.

• One of signatories writes that “Steam, electricity, and computers each gave societies decades to adapt. AI may give us only a few years… We cannot improvise our strategy and institutions in the middle of the transformation; waiting for certainty means arriving too late.”.
• The signatories include leaders from OpenAI, Anthropic and Google.
• 15 Nobel laureates have also signed the letter, including Michael Spence, Daron Acemoglu and Simon Johnson.

Microsoft executives are openly encouraging their sales team to highlight weaknesses in both OpenAI and Anthropic AI models to boost Microsoft model sales – despite the partnership with OpenAI.

• The company’s CEO Satya Nadella told executives that a core emphasis over the coming year will be to monitor AI costs and deploy cheaper models that customers can fine-tune. Microsoft has already replaced OpenAI and Anthropic models in some of its products with Microsoft models.
• Microsoft is also betting on secure AI integration on a complete suite of products, arguing that an end-to-end business process integration is more valuable to companies than standalone models.
• In this regard, Microsoft is also addressing a fear among investors that AI might replace traditional software solutions. The company’s stock price has fallen 20% already this year because of this fear, as well as unease about the amount of money that the company intends to spend on data centers.

OpenAI has developed a large language model called GPT-Red that specializes in finding vulnerabilities in other language models by red-teaming.

• Red-teaming is traditionally done by humans and involves trying to discover vulnerabilities in a model. Vulnerabilities include getting a model to respond to a query which it shouldn’t, e.g., explain how to build a bomb at home, create child porn, leak confidential information, etc.
• OpenAI says that GPT-Red was used to strengthen the defenses of OpenAI’s latest flagship model – GPT-5.6.
• The need to automate red-teaming has increased over the past years as the size of the models increases, and also because the advent of agentic AI is expanding the attack surface of AI systems.
• Researchers at OpenAI explain that GPT-Red discovered a new type of prompt injection attack. This involves inserting a fake entry into a model’s chain-of-thought reasoning which then tricks the model into coming up with an incorrect answer or acting in an undesired manner. One researcher said “It’s like if I told you that 1+1=3 and that you have verified this already… The model’s like, ‘Oh, okay, of course,’ and it just spits out 3.”
• GPT-Red is now being used to test all of OpenAI’s models. The tool will not be released to the public.

This InfoWorld article discusses how one of the modern cloud infrastructure’s core technology, Kubernetes, is poorly adapted to agentic AI workloads.

• Kubernetes is a technology that allows applications to run in containers. Each container is independent of others. Kubernetes containers can be created very quickly, which helps application servers to scale under heavy load.
• In traditional cloud systems, each Web request is handled by launching a container to handle the request. Each Web request is stateless – which means that it does not depend on preceding requests, so a container can be removed when the request is removed. The whole Kubernetes infrastructure is optimized for the fast creation of short-lived containers.
• The concern now is that AI agent applications have different behaviors. Agents are long-running, they call external tools and can even write their own code. An agent workflow can last for minutes or even hours. Kubernetes is not adapted to this kind of application behavior.
• The key features that Kubernetes, and similar infrastructures, need to integrate are durable state management so that an agent state can be reloaded into a new execution environment, coordination tools for agent communication, and credential management for secure agent execution.
• An analysis of 23000 production clusters on AWS, Azure and GCP found an average CPU utilization of 8%, down 10% from last year, and a fall of memory utilization from last year of 23% to 20%. CPU over-provisioning jumped from 40% to 69%. These changes are due to the changing nature of workloads, and agents are contributing to these changes.

Google DeepMind and partners from industry and government have outlined a program to help address biological threats emerging from the use of large language models as well as advance use of AI to help in the event of outbreak.

• The balance for AI companies is to exploit AI to detect and defend against pathogens while at the same time, prevent threat actors from using models to create pathogens. This requires real-time analysis of queries to models to help determine the motivation of the user.
• DNA synthesis is one of the most concrete risks. DNA sequences researched are currently screened by the International Gene Synthesis Consortium which compares them to known harmful pathogens and toxins. However, AI can now help design DNA sequences that are as dangerous as existing pathogens but sufficiently different to known pathogens to avoid being flagged.
• DeepMind is considering the use of a mechanism similar to the SynthID image watermarking technology for tagging DNA sequences, and thereby detecting modified versions of these.
• AI is also part of the pathogen outbreak response pillar. DeepMind mentions that over 10000 publications exist that reference AlphaFold and treat tuberculosis and malaria transmission, as well threat maps for Mpox and Nipah.

Self-driving cars are in the headlines in San Francisco after two power-outage incidents brought Waymo self-driving taxis to a standstill, provoking major traffic jams in the city.

• Waymo is the largest self-driving taxi service in the Bay Area with an estimated 1000 robotaxis. The company operates in 11 US cities and handles half a million rides each week.
• The mayor of San Francisco has written to California’s Department of Transportation to complain about the current regulatory framework. He writes that recent incidents have “demonstrated that California’s current regulatory framework does not adequately address how autonomous vehicles operate during major incidents, planned or not. California’s challenge now is not just whether autonomous vehicles can operate safely under normal conditions, but also whether they can perform reliably during extraordinary ones.”.
• For the mayor, companies should be obliged to have contingency plans for emergency situations like moving cars to the side of the road or adjusting their routes. Robotaxis companies should also share information with authorities about immobile robotaxis and general service disruptions.
• The mayor stresses that his suggestions “will not undermine autonomous vehicles; they will strengthen them”.

This MIT Technology Review article raises concerns about manipulation of weather data due to the rise of prediction markets.

• Weather centers currently base their predictions on data from several weather stations, mainly at airports, utilities and transport services.
• The article tells the story of a suspicious temperature recording at Paris Charles de Gaulle Airport. The airport indicated 22 degrees Celsius whereas the actual temperature was 18 degrees Celsius. One person won 20’000 USD on an on-line gambling site for predicting that the temperature would exceed 20 degrees Celsius that day. Authorities suspect that a hand-held hairdryer was used to manipulate a thermometers reading.
• As weather forecasting becomes more AI-based, and as agentic AI is beginning to be deployed where agents react to specific weather events, the risk of general weather data manipulation becomes more important.
• The authors recommend implementing watch protocols at weather stations, developing adversarial robustness tools for AI weather models, and enforcing accountability for all steps of the weather prediction chain.

This FastCompany article looks at the impact of AI and technology at the FIFA World Cup, ending today in the US.

• The World Cup can be considered a planetary event. Bank of America Global Research data suggests that 75% of the world’s population will have followed matches online or on TV.
• AI has been highly present. Microsoft’s Copilot AI considered that France and Spain were equal favorites to win the tournament (though the final pits Spain against Argentina). The company Lenovo deployed an AI system for teams with 2000 football specific metrics and petabytes of historical data. The company also developed a system to help TV viewers understand what referees were saying. Finally, AI bots were deployed in all host cities to help fans.
• Ticketing has also been an issue due to very high prices. Selling practices have led to investigations in New York and New Jersey by the attorney generals. The problem partly stems from “speculative sellers” on on-line ticket platforms who sell tickets that they have not yet acquired, but hope to purchase at a lower price.
• The US is believed to have spent 1 billion USD on security at the event. Measures include AI powered cameras as well as drones. The Los Angeles Police Department is thought to have acquired software that allows them to take control of any unwanted drone in the vicinity of a stadium