Everyday artificial intelligence is pushing its boundaries further, enters into daily habits and turns on new economic challenges, scientific and diplomatic. The United States They accuse six Chinese companies of systematically exploiting American models to accelerate the development of their products, fueling tensions with Beijing. Meta launches Muse, an assistant capable of performing tasks on behalf of users, while OpenAI Strengthens Alliance with Samsung on chips and claims a solution to one of the great puzzles of mathematics.
We compete for technological leadership, to conquer new markets, and to transform AI into a tool for scientific discovery. But with every advancement they are accompanied by questions increasingly concrete on data security, the reliability of the assistants and the solidity of the announced results.
The United States accuses Chinese companies of copying American AI
Washington raises the bar in its technological clash with Beijing. In a joint notice Published on September 8, FBI, NSA and CISA accuse DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI have conducted industrial-scale campaigns to extract proprietary capabilities from leading US models.
At the heart of the accusations is the “distillation”, or distillation, is a technique that allows a model to be trained using the responses produced by another, generally more advanced and expensive, system. This is a legitimate practice in artificial intelligence research. According to US authorities, however, the companies involved employed it in a targeted and systematic manner, circumventing access restrictions and conditions of use. Agencies claim that, at least since the end of 2024, billions of tokens have been mined through millions of requests for variants of Claude, GPT, Gemini, and Grok. The goal would have been to acquire expertise in reasoning, programming, and other specialized areas, reducing the investment and time required to develop competing models. To hide These operations would have used cloud providers, intermediaries, and proxy networks called "transfer stations." The activities likely occurred with the knowledge of the Chinese government, a circumstance Washington presents as an assessment by its own agencies.
The story obviously also has a military and cybersecurity dimensionThe United States maintains that the capabilities thus obtained could strengthen cyber defense and attack tools that could be used against the country and its allies. Treasury Secretary Scott Bessent has proposed sanctions and Entity List placements for companies involved in operations that constitute intellectual property theft.
Beijing rejects the accusationsForeign Ministry spokeswoman Mao Ning attributed China's progress to the country's scientific and technological autonomy, urging Washington to avoid baseless accusations and strengthen cooperation. The clash comes at a delicate diplomatic juncture, ahead of the planned meeting between Donald Trump and Xi Jinping at the end of September and the bilateral debate on the risks of AI.
Meta launches Muse, the assistant that acts for users. Security remains a key issue.
If international competition at the diplomatic level becomes more intense, Meta tries to bring the intelligent agents in everyday life. Mark Zuckerberg's company has introduced Muse, a personal assistant designed for perform activities on behalf of the user, from planning a trip to sending emails, to making purchases and managing long-term goals.
Il post-sales service It debuted in the United States through a dedicated app and WhatsApp, with web access as well. The basic version is free, while subscriptions for more intensive use cost $20 and $100 per month. It is expected to be available on Meta's smart glasses at a later date. Powered by the Muse Spark modelThe agent runs on a personal virtual machine in the cloud and can continue working even after the app is closed. Users choose which services to connect and which permissions to grant, allowing them to interact with email, calendar, shopping tools, and home devices.
The integration with the Meta ecosystem It should allow the assistant to use shared preferences and information to suggest relevant activities. A recipe saved on Instagram, for example, can become a shopping list. For payments, integration with Stripe Link is planned, which generates single-use cards while keeping the actual card details hidden. The assistant is at the heart of Mark Zuckerberg's strategy to generate revenue from AI beyond advertising, while massive infrastructure investments are fueling investor pressure on financial returns.
The bet must however dealing with safety and reliability. Second ReutersMeta had postponed its April launch to strengthen its protections. Internal tests, while positive results in travel planning were noted, also revealed disconnections, interrupted activity, and issues with personal information management. In one instance, Muse allegedly circumvented the protections by exposing private photos stored on iCloud.
To monitor the system, Meta introduced Sentinel, a second agent that monitors Muse's actions and requests user authorization when necessary. The company ensures that sensitive operations, such as sending emails or making purchases, require consent, and that activity is accessible in a log. Users can revoke access and prevent interactions from being used for training, while conversations and virtual machine data, according to Meta, are not shared with advertising systems. A version with the entire environment encrypted and a key held exclusively by the user is also planned for release by the end of the year.effectiveness These protections will have to be measured against everyday use. AI product manager Vishal Shah acknowledged that errors can't be completely ruled out, but he maintained that the additional work has made the product ready for launch.
OpenAI Strengthens Alliance with Samsung on New Chips
Meanwhile, to support the development of increasingly complex assistants and models, increasing computing power is needed. OpenAI has expanded its collaboration with Samsung Electronics for next-generation semiconductors. Harrison Kim, general manager of OpenAI Korea, listed joint research and production of new chips among the areas where the relationship with the South Korean group has seen the greatest progress, without providing technical details. Samsung did not confirm customer information. The initiative complements the path started with Jalapeno, OpenAI's first custom chip, unveiled in June and designed in conjunction with Broadcom. This processor is intended for inference, the processing by which models respond to user requests, and will be manufactured by TSMC.
Samsung and SK Hynix had also signed letters of intent to supply memory to the Stargate data center project. OpenAI expects demand for these components to continue growing as systems capable of faster and more complex processing become available. The collaboration also involves the use of AI in businessesAccording to OpenAI, Samsung hosts one of the largest enterprise adoptions of ChatGPT in the world. As of the end of August, ChatGPT Enterprise users in South Korean companies and institutions were approximately 28 times higher than a year earlier, though the company has not disclosed absolute numbers.
OpenAI defies mathematics and claims to have solved a Millennium Problem.
From computing power to its scientific applications with its great enigmas, OpenAI relaunches with an announcement that, If confirmed, it would mark a turning point in mathematicsThe company claims that its in-house model, not yet publicly available, has solved the existence and regularity problem of the Navier-Stokes equations, one of the seven Millennium Development Goals.
They are seven big questions selected in 2000 by the Clay Mathematics Institute, which offered a prize of one million dollars for each solution. So far only the Poincaré conjecture has been recognized as solved, thanks to the Russian mathematician Grigori Perelman, who refused the prize. Navier-Stokes equations They describe the movement of liquids and gases and are used in aircraft design, weather forecasting, and the study of blood circulation. The problem is to understand whether a three-dimensional fluid, starting from a regular motion, can develop a "singularity" over a finite time, a condition in which the velocity increases without limit and the mathematical description loses regularity.
According to OpenAI, the model would have demonstrated that this can happen Even in the presence of viscosity, which tends to dampen the motion. The solution described by the company features a vortex that shrinks and stretches, concentrating ever-increasing velocities in an ever-smaller region, while the overall energy remains finite. However, the result awaits confirmation from the scientific communityTo get there, OpenAI would have mobilized approximately 10.000 artificial intelligence agents simultaneously, capable of exploring different approaches and exchanging intermediate results. The solution would emerge 88 hours after the project's launch. Another 17 hours would be needed to formalize and verify the demonstration in Lean, a computer system that allows for the control of the logical steps of mathematical proofs. The company has published the work and formalization, specifying that he does not intend to claim the million dollars.
The announcement also opened a dispute over research priority with Tristan Buckmaster, a mathematician at New York University, and Levent Alpöge, a researcher at Anthropic, authors of a related paper on Euler's equations. Buckmaster disputes OpenAI's reconstruction of the approach. The group denies using their prompts or proofs and acknowledges the two researchers' precedence over the Euler problem with an external force, distinguishing it from the result they claim.
The floor now turns to the scientific examination of the demonstration. If validated, the discovery would offer one of the more significant evidence of AI's capability to contribute to the solution of problems that have remained open for generations.
