Theo on A.I. · June 15th

In this article, we will walk through three essential Pandas tricks to clean and prepare your data efficiently: declarative method chaining, memory and speed optimization via categoricals and vectorized string accessors, and group-aware imputation using .transform().

Anthropic export controls turned an abstract policy fear into a live one last week: as of June 13, 2026, one US government directive took the company’s two most powerful AI models offline for users everywhere, including, briefly, Anthropic’s own foreign-born employees, and set off alarm bells across Europe and Canada about who really controls the AI the world runs on. The mechanics were startling in their speed. The reaction abroad has been louder still. On June 9, 2026, Anthropic made Claude Fable 5 and Claude Mythos 5 generally available, the public face of a model class the company had developed under controlled access since April through a programme called Project Glasswing. Fable 5 was described as a Mythos-class model made safe for general use, state-of-the-art on nearly all tested benchmarks, with strong performance in software engineering, scientific research, and autonomous work. Mythos 5, the more capable sibling, stayed restricted to Glasswing partners and selected biology researchers. Four days later, it was gone. Anthropic said it received an export control directive to suspend access to Fable 5 and Mythos 5 at 5:21 pm ET on June 12, with the letter not explaining the specific security concern in detail. Unable to filter users by nationality in real time, the company said it had to “abruptly disable” access for all customers to comply. The order, issued by Commerce Secretary Howard Lutnick in a letter to CEO Dario Amodei, called for suspending all access by any foreign national, whether inside or outside the United States. Washington cited national security, specifically, a method for “jailbreaking” Fable 5, or getting around its safety guardrails. Anthropic disputed the severity, saying the technique amounted to a limited capability to review programme code and identify errors, something rival models, including OpenAI’s GPT-5.5, can also do. The government’s account is sharper. David Sacks, co-chair of the President’s Council of Advisers on Science and Technology, said on X that the administration asked Amodei to either fix the vulnerability or pull the model from deployment, and that Amodei refused. Sacks pressed the contradiction directly: “In their blog post, Anthropic defended its decision by saying the jailbreak isn’t serious. That is not what the trusted partner and the US government believe; nor is that kind of minimising language consistent with Anthropic’s brand as the AI safety company. The Wall Street Journal reported the move was also shaped by Amazon CEO Andy Jassy, who told Treasury Secretary Scott Bessent and other officials that Amazon researchers had used Fable 5 prompts to obtain information that could aid cyberattacks. Amazon is one of Anthropic’s largest investors. A spokesperson said it is “not uncommon for governments to seek our counsel on potential security risks,” but declined to share details. None of this began last week. The dispute erupted earlier this year after Anthropic insisted its technology should not be used for mass surveillance or fully autonomous weapons systems, infuriating Pentagon chief Pete Hegseth. President Trump ordered every federal agency to stop using Anthropic’s technology, and Hegseth designated the company a “Supply-Chain Risk to National Security“, a label, the company’s lawsuit notes, usually reserved for foreign adversary firms like Huawei. Anthropic sued to reverse the blacklisting, warning it could jeopardise “hundreds of millions of dollars” in revenue. The result is a company simultaneously deemed too dangerous for the US government’s own use and too dangerous for foreign use, a contradiction not lost on observers. Dean Ball, an AI policy expert who briefly served in the Trump administration, called the order “simply cartoonish,” noting that an administration willing to export advanced AI chips to China now wants to ban Britain and every other non-American from using Anthropic’s best models. Outside the US, the response went straight past the jailbreak debate and landed on a single, uncomfortable realisation: a tool embedded in companies, research institutions, and public services worldwide had been switched off by a foreign government, with an email, in an afternoon. The European Commission confirmed it is examining the fallout. Spokesperson Thomas Regnier said the new generation of highly capable AI models offers real benefits, including for cyber-defence, but raises serious cybersecurity concerns that need addressing, adding that “contingency measures taken in this light should not be discriminatory against partners.” European politicians were blunter.

Consumers are showing a willingness to let AI agents take on more shopping-related tasks, according to new research from Accenture. The company’s 2026 Consumer Pulse Research, based on a survey of 25,590 consumers across 16 countries, found that 74% of respondents would trust a personal AI agent more than their best friend to make a purchase on their behalf. The report described this as a move beyond the use of chatbots or search tools. In this context, an AI agent refers to software that can act on a consumer’s behalf within set permissions. It can shop, negotiate, resolve complaints, manage subscriptions, and, in some cases, complete purchases. The survey found that 74% of consumers would allow an AI agent to handle routine tasks. These include deal negotiation, complaint resolution, subscription renewals, and product reorders. Accenture said this level of delegation does not mean consumers are ready to hand over every decision. Instead, the findings suggest that consumers are more open to delegating parts of shopping that feel repetitive, time-consuming, or low-risk. The report also found that 32% of consumers would ask an AI agent to make a purchase decision on their behalf within defined limits. These limits could include budget and brand preferences, with other conditions set by the user. In that scenario, the AI agent would choose the best available option, but the consumer would still review and approve the purchase before payment. The report categorised this as delegated decision-making, separate from task execution and autonomous purchasing. A smaller group of consumers is open to AI agents completing purchases without final approval. The report found that 9% of respondents would allow an agent to initiate and complete purchases within defined boundaries. The payment stage recorded lower openness to autonomous agent decisions. Accenture said only 12% of consumers are open to agents making purchase decisions autonomously at the payment stage. The report identified several conditions that affect consumer willingness to delegate more control. These include data safeguards, configurable permissions, and instant override options. Clear recourse, platform reputation, and perceived neutrality also affect trust. Consumers are more comfortable with AI agent autonomy in parts of the journey where effort is high and emotional stakes are lower. The report pointed to negotiation and post-purchase support as areas where consumers showed greater openness. The report said recurring services ranked highest across stages of delegation, while lifestyle and travel purchases showed a sharper drop as autonomy increased. It also said consumers are more likely to keep control over choices linked to identity or personal enjoyment. A consumer may delegate routine grocery restocking but still want to choose a hotel room, clothing item, or experience directly. The report said AI-assisted shopping requires brands and retailers to make product information clear and machine-readable. If consumers use agents to compare options, pricing, availability, policies, and claims will also need to be easy for agents to assess. AI agents can compare brands using structured attributes and verified claims. They can also weigh price-to-value ratios and fulfilment records. The report said this affects how brands appear across digital channels, including search engines, marketplaces, and social platforms. The report found that 56% of all consumers would tell their AI agent which brands to consider. Among behaviorally loyal consumers, 37% said they would allow an agent to switch brands if it found a better fit. The report linked brand switching to factors such as fit, price, availability, and service performance. Accenture also found that consumers are interested in agents that can work across providers. The report said 61% want an agent that can shop across multiple grocery retailers on their behalf, while 71% want an agent that can plan and book a complete trip across airlines, hotels, and activities. Brands and retailers need product data, pricing, availability, policies, and claims to be readable by the systems agents use to evaluate options, according to the report. The main reasons cited were existing knowledge of shopping preferences, trust built through service and support, and access to a broad selection of products and services. The report listed several possible roles for brands and retailers in AI-assisted commerce.

Four days after Apple confirmed that Siri AI would not launch in China, Huawei took the stage in Dongguan and declared HarmonyOS 7 the beginning of the agent era. The gap Apple could not fill, Huawei has moved into with an architecture built specifically for it. The headline change is the HarmonyOS Intelligent Agent Framework 2.0, which restructures the OS around what Huawei calls an “intent-as-service” model, compressing what previously required multiple app navigation into a single natural-language command. At the centre of this is Xiaoyi, Huawei’s AI assistant, rebuilt from a conventional voice tool into what the company describes as a system-level intelligence agent. Xiaoyi now controls over 2,100 system-level capabilities and coordinates with more than 2,000 third-party AI agents developed across Huawei’s developer ecosystem. Richard Yu, chairman of Huawei’s Consumer Business Group, framed the release as a generational inflexion point: “In 2019, HarmonyOS was born. In 2023, native HarmonyOS apps began. In 2026, HarmonyOS enters the Agent era.” Underneath sits openPangu 2.0, Huawei’s updated foundation model, with 505 billion parameters in its Pro version and 92 billion in the Flash variant, both supporting 512K context windows. On-device models at 30 billion parameters are due on Kirin chips by autumn 2026. HarmonyOS 7 also delivers a 15%-plus performance improvement over HarmonyOS 6.1, according to Huawei’s own benchmarks. The task execution rate claimed is above 90%, though that figure is Huawei’s own and has not been independently verified. The numbers shared at HDC 2026 reflect a shift that has already happened. In Q1 2026, HarmonyOS held 19% of China’s smartphone OS market against Apple iOS at 16%, with Android at 65%. HarmonyOS first overtook iOS in China in Q2 2025, according to Counterpoint Research. That trajectory matters more than any single feature because China is simultaneously the market Apple cannot currently operate in at the AI level and the one Huawei has fully optimised for. The agent network Xiaoyi coordinates includes partnerships with Ctrip for travel planning and Ant Medical for health data analysis, services woven into the Chinese consumer stack that Apple’s architecture does not reach. The scope of the challenge to Apple needs calibrating. HarmonyOS 7 is currently in developer beta, with the stable consumer release expected this autumn. The 2,000-plus AI agents are anchored in the Chinese app ecosystem. The platform counts more than 400,000 applications and services, which is significant but still a fraction of what Apple’s App Store carries. Huawei’s ambitions to take HarmonyOS international remain aspirational for now. There is also a design note that softens any clean divergence narrative: HarmonyOS 7 adopts the same Liquid Glass aesthetic Apple introduced with iOS 26, and Samsung brought to One UI 9. Visual language converges even as underlying architectures and regulatory environments pull in opposite directions. HarmonyOS exists because of US sanctions. When Huawei lost access to Google’s Android in 2019, it built its own OS from necessity. By January 2026, over 90% of Huawei devices were running the fully homegrown version. That forced independence is now a structural advantage in the one market where Apple cannot currently deploy its headline AI feature. Sanctions built the platform. Regulatory friction cleared its path. See also: Siri AI arrives with Google inside, and much of the world is locked out Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information. AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here. The post HarmonyOS 7 steps into the AI gap Apple left open in China appeared first on AI News.

Satya Nadella’s latest talk is less about hype and more about the mechanics that could tip the balance of whole industries. He’s nudging companies to treat AI like a new kind of capital—something you build on top of your own data, not just a service you rent. By looping proprietary data through custom models, firms can create a “token capital” that works alongside their people, keeping the value inside the organization.

His concern is that if firms rely only on the big, off‑the‑shelf models, a handful of AI systems will end up scooping up the bulk of the economic upside. Those models would act like a monopoly, pulling profit from sectors that once depended on diverse, niche expertise.

Nadella’s angle lines up neatly with Azure’s strategy: the cloud platform becomes the place where companies can spin up their own learning loops, store the data that fuels them, and keep the returns under their own roof. It’s a subtle shift from “use AI” to “own AI,” and it reshapes how value is captured in the age of generative tech.

The European Commission is assessing the implications of the US order that forced Anthropic to shut down Fable 5 and Mythos 5 worldwide. European researchers are debating the right response: building their own foundation models or securing access through contracts. But building out homegrown infrastructure would require computing capacity, energy, and competitive providers that Europe currently lacks, experts warn. The article Anthropic shutdown sparks sovereignty debate across Europe appeared first on The Decoder.

Volunteer AR scans from Pokémon Go players fed into Niantic's spatial AI models. That technology is now being combined with a US defense contractor's software for GPS-free navigation. The article Pokémon Go data helped train AI now linked to military drones appeared first on The Decoder.

How local optimization in last‑mile delivery can quietly break the system The post The System Always Knows: Why Local Efficiency and System Performance Are Not the Same Problem appeared first on Towards Data Science.

A single model hands you a single answer and no sense of how much it hinges on the dozens of choices buried inside it. The post I Built 11 Models to Predict the 2026 World Cup. They Crown Four Different Champions. appeared first on Towards Data Science.

I’ve been playing with the new vision‑LLM that can actually look at a PDF the way we skim a page, but it also “sees” the graphics. Instead of just OCR‑ing text, it feeds the image of a chart or diagram into its visual encoder, extracts the layout, axes, and data points, then stitches that into the same token stream the language model uses. The trick is that the model learns a joint representation, so a question about a sales trend can pull numbers straight from a bar graph without a separate OCR step.

What’s neat is the way they’ve wired the retrieval‑augmented generation (RAG) pipeline. The visual front‑end creates embeddings for each visual element, stores them alongside the text embeddings, and when you ask something like “how did Q3 compare to Q2?” the retriever pulls the relevant chart slice and the generator weaves the answer together. It feels like the system finally respects the way we actually read documents—words and pictures together.

The result is a smoother workflow for enterprises that have tons of mixed‑format reports. No need to hand‑craft separate parsers for tables or diagrams; the vision model does the heavy lifting, and the language side just fills in the narrative. It’s a subtle shift, but it means you can ask about a figure and get a precise answer without manually extracting the data first.