AI, Compute Power, and Regulation: The Situation on 18/07
Today’s topics are AI models, compute, open source, and regulation: from Netflix to Kimi K3, from Torvalds to Anthropic.
Inhaltsverzeichnis
Today, the AI world is once again moving on several fronts at once: models are getting stronger, compute remains scarce, and regulation wants to pull Big Tech further apart. That sounds dry, but it’s actually highly relevant — because this is exactly where it gets decided who builds the next generation of AI and who only gets to use it.
On top of that, companies like Netflix are already showing very concretely how generative AI saves money and time. And while labs and corporations fight over GPUs, Linus Torvalds reminds the scene that open projects do not automatically become feel-good spaces. In short: today is about power, cost, and control. Not the worst combination for an exciting AI day.
🚗 Xpeng launches its Europe expansion with AI driver assistance
Xpeng is launching its electric SUV L03 in China and Europe at the same time and presented the first data and prices at a launch event in Munich. For AI Radar, this is not just an automotive story, but also a signal of how closely mobility, software, and AI are intertwined today. Modern electric SUVs are increasingly defined by driver-assistance systems, sensor fusion, and over-the-air updates — in other words, by software, not just sheet metal and batteries.
This is also exciting in the context of regulation and competition: European markets are attractive for Chinese manufacturers, but at the same time difficult because of standards, safety requirements, and political debates. Anyone who wants to succeed here needs more than a good product spec sheet. They need trust, infrastructure, and staying power. Source: heise
🎬 Netflix uses generative AI in 300 productions
Netflix is already using AI in around 300 productions, mainly in post-production. Co-CEO Ted Sarandos cites the documentary series “The American Experiment” as an example: 17 minutes of AI-generated material, produced in half the time and at half the cost. That sounds like efficiency, but above all it offers a glimpse into the future of media production. AI is no longer an experiment here, but a tool in day-to-day operations.
The key point is the broader context: Netflix does not want to simply pocket the savings, but invest them into more content. In plain terms, that means generative AI will not only cut costs, but also increase the volume of content. For creatives, the question remains how roles will change — and for the industry, whether AI ends up being more of an accelerator or a quiet shift of work into the machine. Source: The Decoder
🧠 Kimi K3 shakes up the LLM league from the top
Moonshot AI has released Kimi K3, a model that, according to early assessments, is said to be on the level of Anthropic’s Opus 4.8 — with a team of only about 300 people. This is exactly the kind of news that instinctively makes people in the West raise their eyebrows: How can a comparatively small team build a model that can keep up with the big names? The obvious answer is, of course: compute, data, training, and a great deal of engineering skill.
But the debate goes further. If open-weight models keep getting stronger, the gap between the “Big Three” and fast-learning challengers shrinks. That suddenly makes export controls, compute access, and infrastructure policy very concrete. Even OpenAI strategist Dean W. Ball calls the model “very good” and at the same time warns against an open model world — in his wording, “AI communism.” A bit polemical, but the direction is clear: control over models is also power over markets. Source: The Decoder
🔓 Thinking Machines introduces an adaptable AI model with “Inkling”
Thinking Machines is introducing its first AI model, “Inkling,” and thereby entering the competition for open or adaptable models. The focus appears to be less on maximum benchmark dominance and more on flexibility and customization. That is an important signal, because the LLM market is not determined only by “Who is the best?”, but increasingly by “Who is the easiest to integrate?”
For companies, that is often the really decisive question. A model that is easier to fine-tune, control, and embed into existing workflows can end up being more economically valuable than a theoretical front-runner. Especially in the enterprise and product context, we are seeing this more and more: not the loudest model wins, but the most useful one. Source: heise
🏗️ Anthropic is reportedly seeking more compute from Meta
According to reports, Anthropic is negotiating with Meta about renting compute from its data centers. At first glance, that sounds like a normal infrastructure deal, but in reality it is a very good illustration of the current AI market: even the big model providers depend on the availability of compute. If you have too little GPU capacity, you can write the best research papers in the world — without compute, you won’t turn them into a scalable product.
The fact that Meta is the one potentially providing the hardware also shows how complex market relationships have become. Competition and cooperation are happening in parallel, sometimes in the same sentence. For the industry, this means infrastructure is becoming an increasingly strategic currency. It’s not just models that are scarce, but also the hardware they run on. Source: The Decoder
🛠️ Tool tip of the day
If you are building AI projects, clean access to model APIs, logs, and experiments is worth its weight in gold. That is exactly why it’s worth taking a look at modern developer platforms and orchestration tools that let you test different models, prompts, and workflows — before everything creatively falls apart in day-to-day production. For teams that want to prototype faster and spend less time gluing JSON together, this is often the difference between “exciting demo” and “also works on Monday.”
🧑💻 Torvalds firmly backs AI tools in Linux
Linus Torvalds has spoken out clearly in favor of using AI tools in Linux development and has brushed off anti-AI criticism very directly. His message is about as charming as it is unambiguous: Linux is not an anti-AI project, and anyone who has a problem with that can fork the project or leave. Dryly put: this is not a soft launch of a debate, but its operating manual.
The trigger is the discussion around the Linux Foundation’s agentic AI code review tool Sashiko. Substantively, this matters because open-source projects are currently wrestling with how much AI assistance they want to allow. Torvalds takes a pragmatic stance here: tools are tools, not ideology. For the open-source world, this is a pretty clear marker of where the culture is shifting — away from ideological romanticism, toward productivity and acceptance. Source: The Decoder
🏅 MIT robotics researcher Daniela Rus receives Bavarian High-Tech Prize
US robotics researcher Daniela Rus is receiving the Bavarian High-Tech Prize from the minister-president. At first glance, this is a classic science note, but it fits perfectly into the broader AI and robotics context: the intersection of robotics, AI, and industrial application is becoming one of the most important fields of innovation there is.
Awards like this are not just honors, but also location policy. Bavaria wants to make it visible that high tech, research, and transfer belong together — and that the next industrial wave will not consist only of LLMs. Especially in combination with AI-supported robotics, we will see which regions attract talent, capital, and research over the long term. And yes: a bit of prestige never hurts in the global technology race. Source: heise
Don’t want to miss any news? Subscribe to the newsletter