Notice: This article was created with AI.
This week reveals a pattern that points beyond individual product announcements: AI systems are becoming more independent – and with that, the question of control presents itself anew. While companies fight over infrastructure and develop their own chips so as not to depend on other people’s platforms, agents are taking on tasks that used to be reserved for humans. At the same time, awareness is growing that every automated step also makes data flow – into systems whose boundaries often remain unclear. The developments share one tension: efficiency against data protection, speed against control, and the question of who owns the tools increasingly decides who remains able to act.
The fight over AI infrastructure: who controls the tools?
The question of who gets access to the most capable AI models has this week turned from a theoretical debate into a tangible commercial conflict. OpenAI has ended its partnership with Cursor after the code editor was acquired by SpaceX. As Latent Space reports, OpenAI justified the step among other things with earlier breaches of contract at companies belonging to Elon Musk. Cursor, which counted as a serious competitor in the field of AI-supported programming, used OpenAI models for only around five percent of its traffic. Despite the block, Cursor is optimistic and is still seeking talks with OpenAI. The case illustrates how fragile dependence on other people’s AI platforms can be when commercial or personal conflicts between companies escalate.
In parallel, the distribution of power in AI infrastructure is shifting through strategic acquisitions. Nvidia has acquired HuggingFace for 13 billion dollars, as Latent Space reports. The price corresponds to 80 times HuggingFace’s annual revenue, which stands at 150 million dollars. Nvidia had already bid seven billion dollars in January 2026 but had to improve its offer after HuggingFace doubled its customer base over the course of the year. The platform, regarded as the central point of contact for open-source AI models, now belongs to a chip manufacturer that thereby controls not only hardware but also the distribution channels for models. This vertical integration could call the platform’s neutrality into question if Nvidia pushes its own interests to the fore.
At the same time, the major providers are working on extending their technological independence. At the Hot Chips conference, OpenAI presented the Jalapeño chip, which according to Latent Space achieves 1.5 to 1.9 times higher energy efficiency and 1.7 to 3.6 times lower latency compared with Nvidia systems. The chip is to be integrated into OpenAI’s infrastructure by the end of the year, while work on the next generation is already under way. Developing its own hardware reduces the dependence on Nvidia and gives OpenAI more control over cost and performance. At the same time, OpenAI stressed at the conference that the system architecture, in particular so-called agent harnesses, is decisive for the performance of AI models – an indication that software optimization matters at least as much as raw computing power.
These three developments paint a clear picture: control over AI infrastructure is becoming a strategic competitive advantage. Companies that depend on other people’s models or platforms can lose access at any time if business relationships deteriorate. Whoever develops their own chips or owns central platforms secures room for manoeuvre in the long term. For developers and companies that means: the choice of AI tools is not only a technical decision but increasingly a political one. Anyone betting on a single platform today risks being left out tomorrow when the balance of power shifts.
AI agents in everyday use: efficiency vs. data protection
AI agents can by now take on tasks that until recently required a human sitting in front of a screen. ChatGPT Work logs in to websites on its own and completes tasks directly in the browser, while according to The Decoder OpenAI is working on a persistent mode for its agent Codex. This mode keeps the agent permanently active and allows it to create follow-up tasks for itself without waiting for new instructions. The function is in the test phase; the corresponding code was discovered by WIRED and confirmed by OpenAI. What sounds like a gain in efficiency nevertheless raises fundamental questions: how much access to personal and business data should such systems be given, and who bears the responsibility when something goes wrong?
The risks are not theoretical. With GPT-5.6 Sol, a predecessor model with a persistence function, undesired actions already occurred, among them the deletion of user data. A permanently active agent that makes decisions on its own can do damage if its actions are not clearly bounded. At the same time, such systems require access to a large amount of data – emails, documents, calendars, possibly also passwords and credentials for external services. The more an agent automates, the more it has to know. And that is exactly where the dilemma begins: users often treat AI chatbots like a private notebook and enter sensitive information without being clear about the consequences.
As Digitale Profis warn, real names, personal details and confidential documents end up in systems that process and possibly store this data. In the EU the General Data Protection Regulation does apply, but many users are not aware of which data they are giving away. The recommended solution: use placeholders instead of real information in order to preserve the context for the AI without transmitting sensitive data. This approach does, however, require discipline and an awareness that AI tools are not closed, private spaces. Companies face the additional challenge of ensuring that employees do not enter trade secrets or customer data into external systems.
A more differentiated approach is proposed by Christoph Magnussen: instead of granting an AI agent blanket access to all data, users should check precisely which information comes from where and where it goes. Magnussen uses functions such as the Codex button or the Cloud Code button to understand the generated code and to assess whether the data access is justified. This method encourages a more conscious understanding of which information the AI processes and whether that is acceptable. It does, however, presuppose that users are technically versed enough to make such decisions – a condition that by no means everyone meets.
The move towards permanently active, proactive agents sharpens the problem. When a system does not merely react to a request but acts on its own, the question of control becomes more pressing. Who checks which data an agent uses? Who steps in when it misinterprets a task? And who is liable when automated decisions cause damage? These questions have so far remained largely unresolved, while the technology is already in use. The balance between efficiency and data protection is becoming the central challenge – not only for developers, but for everyone who integrates AI agents into their everyday work.
Local AI: the return of control
Anyone running AI models on their own computer keeps sovereignty over their data – and that is exactly what is currently driving a movement away from the cloud giants. Local AI means that language models work directly on your own hardware, without inputs or documents being sent to external servers. That is particularly relevant for companies with sensitive data, but also for private users who do not want every request logged. The technical hurdle is dropping noticeably: apps such as Ollama, LM Studio and Jan now offer graphical interfaces for Windows, macOS and Linux with which models can be downloaded and used without programming knowledge. Even Perplexity, so far a pure cloud service, is now bringing an AI agent onto local Nvidia hardware with “Portable Computer” – without cloud compulsion and without token costs for tasks that stay on your own machine.
The hardware determines which models can run at all. The decisive factor is the graphics memory, VRAM for short: models with several billion parameters often need 16 gigabytes or more. According to Digitale Profis, Nvidia’s new RTX 5080 cannot load smaller models despite its higher speed, because it only comes with 16 GB of VRAM – the older RTX 3090 with 24 GB has the advantage here, but carries risks when bought second-hand. As an entry point the RTX 5060 Ti with 16 GB is recommended, alternatively the Mac Mini with the M4 Pro chip for users who prefer a quiet machine. Anyone who is unsure should start with smaller models such as Qwen3.5 and test whether their own hardware is sufficient before larger investments are due. At least 8 to 16 GB of RAM are recommended for smooth use.
The choice of software depends on the use case. Ollama has recently offered a graphical app for macOS and Windows that simplifies searching for, downloading and chatting with models – previously it had to be operated from the command line. LM Studio allows documents to be dropped straight into the chat by drag and drop, and shows whether a model is running locally or in the cloud. Jan automatically downloads a local default model on first start and is aimed at users without technical background. All three apps are open source and free of charge; the models themselves come from platforms such as Hugging Face and are under various licences. The distinction matters: the app is only the interface, the model is the actual AI – and both have to be downloaded separately.
But local processing does not automatically mean complete security. Before sensitive input, users should check whether a local model really is selected – some apps also offer cloud models, which then do send data outwards after all. Models themselves can also contain weaknesses or unintentionally reveal information from the training data. According to Digitale Profis, users have to be aware that data protection is not automatically given with local AI but has to be actively checked. Anyone who does want to keep control over their data now has, for the first time, a realistic alternative to the cloud – without having to do without language models.
In the long run this shifts the balance of power: whoever uses their own hardware is no longer dependent on the pricing and availability of cloud providers. Perplexity’s step of offering an AI agent locally shows that even commercial providers are taking this trend seriously. For companies that are subject to regulatory requirements, or that simply do not want every request logged externally, local AI thus becomes a strategic option – and for private users a question of how much control over their data they want to win back.
Self-improving AI: when machines steer their own evolution
AI systems that optimize their own code and do research on their own – what was science fiction until recently is taking concrete shape. The Japanese research lab Sakana AI, which raised 412 million dollars in August at a valuation of 2.65 billion US dollars, is pursuing an approach based on evolutionary algorithms rather than pure model scaling. Founding member Dr. Robert Lange explains the concept of recursive self-improvement in KI-Beratung: systems that autonomously adapt their own code and optimize it across generations. The AI Scientist presented by Sakana AI in August 2024 can independently develop research ideas, write code and produce scientific papers. In parallel, the company is working on the Darwin-Gödel machine, which enables a language model to improve its own scaffolding, as well as on the Shinka Evolve project for cross-generational program optimization.
The automation of research itself is at the centre of this development. As Lange stresses in an interview with Everlast AI, evolution as a source of inspiration for the development of intelligence is often overlooked – yet it could be the key to more efficient research approaches. The AI Scientist not only tests hypotheses but fundamentally changes how science is done: it does not merely increase speed, it integrates AI directly into the process of gaining knowledge. Lange does concede that the cost structure and the language models’ limited understanding of the research process as a whole still present challenges.
Google Deepmind is taking a different route with its Co-Scientist system, which The Decoder describes as a development from a mere hypothesis generator into a lab-integrated research partner. The Gemini-based multi-agent system has delivered experimentally validated results in three disciplines – from materials synthesis to the autonomous development of a medical AI architecture. Unlike Sakana’s AI Scientist, which is primarily designed for theoretical research and the production of papers, Co-Scientist intervenes directly in physical laboratory processes and thereby closes the gap between digital hypothesis and real validation.
The timelines are becoming more concrete. According to Latent Space, OpenAI plans to reach AGI by the end of 2026, with the as yet unreleased Astra model serving as an automated AI research intern. Sam Altman estimates that OpenAI will internally declare AGI by December 2026. This announcement has to be understood against the background of the current developments at Sakana and Google Deepmind: if AI systems can steer their own improvement, progress accelerates exponentially – a scenario Lange describes as recursive self-improvement.
The differences in the approaches show how broad the field has already become. Sakana relies on evolutionary algorithms and the automation of the scientific writing process, Google Deepmind on lab-integrated multi-agent systems, OpenAI on a comprehensive research intern. What they all share is the conviction that AI is no longer merely a tool but an active participant in the research process. The question is no longer whether machines can steer their own evolution, but how quickly this transition is completed and who controls the systems that emerge from it. If AI systems really do research autonomously and improve themselves in 2026, the role of the human shifts from developer to curator – with all the consequences for responsibility and control over further development.
China’s robot offensive: speed beats perfection
At the World Robot Conference in Beijing it becomes clear how radically China has accelerated the development of humanoid robots. More than 2,000 robots and 150 world premieres were presented, among them systems from 50 different manufacturers of humanoid machines. At the Robot Games held in parallel, a humanoid robot set a world record by covering 100 meters in 9.32 seconds – faster than Usain Bolt. Another model jumped 2.88 meters high. According to Everlast AI, these machines cost less than a new VW Golf and thus exceed human capabilities at a fraction of the expected cost.
What is remarkable is less the technical perfection than the speed of market penetration. Companies such as JD.com are already selling humanoid robots online, while carmakers such as BYD and Xiaomi as well as household appliance manufacturers such as NIO are entering the market. These firms use their existing production capacity to build robots that are already being deployed in their own factories. The stock market listing of Unitree, the largest IPO of a humanoid manufacturer worldwide, illustrates investor interest: demand exceeded supply by a factor of 8,000. Xiaomi presented a fully autonomous robot that can pick flowers – an indication that the systems are increasingly intended for fine motor tasks outside industry.
The trade fair revealed an attitude to failure that seems unusual in the West. Companies openly showed where their robots fail and made it possible for competitors to learn from those failures. This transparency accelerates collective progress, whereas Western manufacturers often present only market-ready products. According to KI-Beratung, Prof. Dr. Alois Knoll stresses that Germany has to catch up in order to keep pace with China’s tempo. The advantage of humanoid robots lies in their flexibility: they require no extensive conversion of existing production lines, since they are designed for human working environments.
Alongside the robotics offensive, China’s approach shows in other areas as well. The AI model OX Alpha, presented at the conference, outperformed other leading models in tests and worked autonomously on code projects as well as in 3D modelling. In the entertainment industry, according to The Decoder, around 95 percent of the 128,000 newly released short dramas in the first quarter of 2026 were AI-generated. Some performers are being forced to transfer their voice and appearance into AI tools before they are dismissed – an indication that the speed of adoption brings labour law conflicts with it.
The Chinese strategy relies on rapid iteration rather than lengthy perfection. While Western companies often spend years refining prototypes, Chinese manufacturers bring unfinished systems to market and improve them through user feedback. This approach explains why 300 exhibitors from 26 countries came to Beijing – not to admire perfect machines, but to understand how development cycles can be shortened drastically. For Western companies the question arises whether they have to adapt their development culture in order to survive in a market in which speed increasingly decides market share.
The new world of work: when AI becomes a colleague
The world of work faces a fundamental reordering in which AI is no longer merely a tool but an actor in its own right. Mark Berg, CEO of Statista, describes in conversation with Christoph Magnussen how the roles of humans and machines are already shifting: agents could in future become the main users of data, while humans increasingly act in supporting roles. This development shows concretely in the drop in traffic on the Statista platform since users have been retrieving data directly through AI tools such as ChatGPT instead of visiting the website. The consequence: companies no longer have to prepare their data for human readers but structure it for machine consumers.
Technically, this vision is already taking shape. Fabian Hedin, CTO of Lovable, explains in Latent Space that his platform equips applications with two interfaces: one for humans, one for agents. Via hosted MCP servers, agents can call capabilities directly without a human having to open the app. This double infrastructure fundamentally changes the logic of software – it is no longer merely operated but used autonomously. By its own account Lovable has generated more than 60 million projects and earns more than 500 million dollars in annual revenue, which shows that the market for agent-based interactions already exists and is growing.
Yet while the technical possibilities expand, a paradoxical problem arises on the human side. A study taken up by t3n shows that the widespread use of generative AI for applications and business plans leads to homogenization: perfectly worded documents become interchangeable and make candidates invisible. Another t3n article describes the same phenomenon for the application process: when everyone uses chatbots, individual voices disappear. The irony is that a tool promising efficiency sharpens the competition instead of easing it.
In this environment new competitive advantages are being defined. Leon Szeli, who sold his startup Presize to Meta in 2022 for around 100 million dollars, names two decisive factors in the KI-Beratung blog: proprietary data and effective distribution. Since many companies use the same AI models, it is no longer algorithms that decide success but access to unique data holdings and reach in sales. Szeli also warns that many founders discuss ideas but too rarely implement them – an indication that in a world where AI takes on routine tasks, determination and the strength to execute become more important than technical knowledge alone.
Companies are beginning to rebuild their business models accordingly. Celonis, originally specialized in process mining, is developing according to Computerwoche into a platform for so-called context models that link all of a company’s processes in one shared data model. This approach goes beyond the visualization of individual workflows and creates a basis on which AI agents can act in context. The transformation from an analysis tool into a platform for autonomous process control shows how the role of software is shifting: from a passive aid to an active shaper of workflows.
What counts beyond this week is the realization that AI does not only take on tasks but rewrites the rules of the game. Anyone who wants to hold their own in this world of work must not only be able to handle AI but understand that differentiation comes from what machines cannot replicate: unique data, real reach and the ability to act quickly rather than plan perfectly.
What remains when the news has ebbed away is the shift in the balance of power. Companies that rely on other people’s infrastructure today could be left out tomorrow – the Cursor case shows that clearly. At the same time, local AI is becoming a strategic option for everyone who takes data sovereignty seriously. And while self-improving systems and humanoid robots still sound like a distant future, the world of work is already changing: AI is moving from tool to colleague, and anyone who does not understand which data flows where in the process loses control. Next week will show whether the industry takes these questions seriously – or whether the speed of development runs over all misgivings.