Time for a catch-up!

Well, as the title says, this week is a mixed bag of news, with new software, major business deals and some positive developments in terms of AI benefitting us through improvements to medical technology.

Probably the biggest headline though was the news that in its attack on Russian warplanes last week, the Ukrainians utilised AI both in the planning and the actual execution of the attack. This opens up a Pandora’s Box around concerns with where this could lead, particularly following the revelations I highlighted last week about AI trying to prevent itself from being shut down by employing blackmailing techniques or simply ignoring instructions.

Just a thought 🤨 Since all of these LLMs are trained on information from the internet, giving them access to the storylines of both the Terminator movie series, but also Kubrick’s 2001: A Space Odyssey was probably not the best idea 😩

Why do I raise this now, and not under the legislation section? Well, governance and oversight are key, and this week Meta announced that it plans to replace humans with AI for assessing societal and privacy risks across its platforms like Instagram and WhatsApp.

Although Meta claims that the AI system would just be used to assess “low-risk” releases, the AI system is reportedly being used to make decisions on AI safety, youth risk, and integrity (including misinformation and violent content moderation), which are high-risk areas.

So, with that issue flagged, let’s dive in and catch up on everything else that’s been happening …

Big Sharks vs. small fish

Staying with Meta for now, they announced that they have entered a 20-year agreement with Constellation Energy to procure 1.1 gigawatts of power using nuclear power in Illinois, starting June 2027. This marks Meta’s first venture into nuclear energy, following similar initiatives by Microsoft, Google, and Amazon as tech companies seek additional reliable energy sources to support expanding AI infrastructure.

Meta also plans to enable brands to autonomously generate entire advertising campaigns including images, videos, text, and optimised targeting using AI across Facebook, Instagram, and WhatsApp by the end of 2026. This is likely to have a massive impact on what is a $26 Billion industry and moves us to the possibility of AI agents engaging each other in the sales process. This move aims to attract direct marketing spend from brands and will disrupt traditional advertising and media agencies.

Amazon announced a $10 billion investment to build a data centre and AI campus in North Carolina, and Amazon Web Services (AWS) is expanding its data centre footprint globally, with new facilities planned in Chile, New Zealand, Saudi Arabia, and Taiwan. This expansion aims to support the growing demand for cloud services and enhance access to advanced AI chips. Amazon is also developing an AI foundation model to enable robots like Proteus to understand and act upon natural language commands. Additionally, not forgetting where its empire started, the company introduced "Wellspring," an initiative leveraging generative AI to enhance last-mile delivery precision and driver experience.

Whilst on the topic of logistics, DHL deployed machine-learning algorithms to optimise delivery routes, built AI image recognition systems to sort parcels, and automated many back-office tasks. This resulted in 60% quicker processing of shipping documentation and a 50% reduction in errors. It also cut delivery times by 20%, fuel costs by 15%, and increased sorting speed by 30%.

Elon Musk’s xAI also announced an infrastructure investment of $5 billion for a data centre in Memphis. This currently has 200,000 GPUs, but the investment supports broader AI compute plans, including scaling to 1 million GPUs. The move signals growing capital needs in AI infrastructure.

Talking of Elon Musk and following on from the speculation last week it was finally confirmed that Telegram will host Elon Musk’s Grok AI chatbot starting this summer through a $300M deal with xAI. There was confusion when Musk went onto X to deny it, only for the two companies to finally confirm the deal in a joint statement. Telegram gets cash, equity, and half of the subscription revenue, and in return, Grok gains access to Telegram’s billion users, crucial for scaling paid AI services. It is worth keeping an eye on AI chatbots moving from apps to messaging, where engagement is constant and monetisation is growing.

Also, in infrastructure development Applied Digital signed a $7 billion, 15-year lease with CoreWeave to provide 250MW of AI data centre capacity in North Dakota. Two centres will go live by mid-2026, with a third planned. CoreWeave can scale to 400MW. This shows how AI demand is reshaping infrastructure - and why access to power, not compute, may soon be the real bottleneck, linking back to Meta’s decision to invest in nuclear power to support that demand.

In software news Kuaishou’s new AI tool, Kling 2.1, creates high-quality videos in seconds. It lets users upload an image and give a text prompt to guide how the image should move or transform. It delivers 5–10 second clips in under a minute for between $0.30 and $1.50 depending on the quality. This shows AI video tools are getting faster, better, and more affordable, which could change content creation for many users.

Countering that is Microsoft, who have also added a new AI video generator - Bing Video Creator - into its Bing app using OpenAI’s Sora model. Users can create short, vertical videos from text prompts at no cost, with a cap of 10 clips before redeeming Microsoft Rewards points. The tool is mobile-only for now, but the move makes generative video more accessible and nudges marketers to rethink content workflows. All of this highlights how competition in AI video tools is heating up, with Tencent, and Google also recently releasing new products.

Microsoft are not the only one to be giving away stuff for free this week, with Anthropic making Claude’s web search free for all users, along with beta voice access on mobile. This lets anyone get real-time answers and use voice to interact with Claude. The update follows Claude 4’s release and helps it catch up with AI tools like ChatGPT. Free access may drive user growth and improve Claude’s position in live, voice-driven research use cases. This is linked to the news of Anthropic’s "Claude Opus" which now has conversational voice capabilities, allowing users search Google Docs, Drive, and Calendar with vocal commands.

There were also rumours of OpenAI possibly buying Anthropic, off the back of news that the “start-up” had managed $3 billion of revenue in the last year. In other buyout new Salesforce announced that that had bought Informatica for $8 billion. This was primarily to fix AI trust issues, by improving the platform’s data governance, transparency and compliance for enterprise sized users.

In terms of business applications Microsoft announced that it will be supplying Barclays with 100,000 Copilot AI licenses to automate tasks and support decision-making across its workforce. The deal joins others from Accenture and Toyota who are already using Copilot at scale.

Toyota also announced that its repair technicians are using AI conversational tools for quicker troubleshooting, boosting productivity and reducing dealership bottlenecks.

Not to be outdone by video, there were two major developments this week in relation to AI-generated voice. The first was ElevenLabs’ new Conversational AI 2.0, which has redefined the voice assistant game, which can precisely pause, take turns in a conversation, and essentially sound human.

Next was Phonely who in collaboration with Maitai and Groq hit a groundbreaking 99% conversational accuracy. It’s cracked the dreaded "awkward AI pause" by reducing voice latency by more than 70%, allowing customers to barely distinguish between human and AI callers. This tech is already replacing 350 call centre agents in one company this month.

This matters because these parallel breakthroughs signal major disruption in industries relying on human-sounding communication, from call centres to entertainment.

Legislation, policy and other news

In legislative news, the UK House of Lords passed an amendment requiring AI companies to disclose the copyrighted materials used in training their models, challenging the government's data bill and emphasising the need for stronger copyright protections.

Perhaps mindful of this and the general shift towards copyright protection, Amazon announced that it had entered a multi-year AI licensing agreement with The New York Times, allowing them to use its editorial content for AI products like Alexa. This marks the publisher's first licensing deal involving generative AI.

It’s not all smooth sailing though. Following the release of its advanced AI reasoning model R1-0528 last week, researchers have been left asking whether DeepSeek trained the model using data from Google’s family of Gemini AI models, which has taken Google years to build/train. According to experts, R1-0528 uses language patterns, words, and expressions that Gemini 2.5 Pro often uses, and exhibits the same reasoning/thought process that Gemini models use to reach a conclusion.

DeepSeek faced similar accusations earlier this year when OpenAI found evidence that it was using a distillation process (which trains AI models using more advanced ones to save time, data, and money) to train its models. Plus, OpenAI investor, Microsoft, found that data was being taken from OpenAI developer repositories. Although distillation isn’t illegal, OpenAI’s rules state that its models can’t be used to train competing models.

Also, Reddit filed a lawsuit against AI startup Anthropic, alleging unauthorised use of Reddit's user data to train AI models without proper licensing or user consent. So, expect to see continued news in this area as the law tries to come to terms with the whole idea of intellectual property in the age of AI.

Ending as always on a positive note, there was some positive news of the benefits of AI in terms of medicine (for humans and animals!). The first was ONC, who showcased its deep learning radiomic model, which predicts overall survival in late-stage non–small cell lung cancer patients by analysing changes in routine CT scans over time. Then Northwestern Medicine announced the introduction of a generative AI system that significantly enhances radiology by improving speed and accuracy, aiming to boost productivity and diagnostic precision in medical imaging.

The next news was a new tool using AI that analyses lymph node and skin masses in animals, providing veterinarians with faster and more accurate diagnostic capabilities. Obviously if it works with animals, it should only be a matter of time before this is translated into use for humans.

Finally, Yoshua Bengio, a Turing Award winner and one of the "godfathers of AI," has launched LawZero, a nonprofit focused on building safe-by-design AI systems. The organisation also raised $30 million to develop "Scientist AI", a tool designed to promote truth, transparency, and responsible AI use. LawZero aims to build AI that gives probabilistic answers, not false certainties, accepting that AI can be unsure. The upcoming "Scientist AI" will help advance science, watch other AIs for lies, and reduce risks.

Stay informed. Stay critical. And wherever possible—stay ahead.

Regards

Tom Carter