Anthropic, EU, DeepL: The AI Day at a Glance

Today makes it pretty clear what the AI world is currently wrestling with: compute, regulation, and the question of how models can actually become reliable. While Anthropic and OpenAI keep tightening the infrastructure engine, the EU is tweaking the legal fine print—and research is showing that “understand first, obey later” works surprisingly well for language models.

And yes: incidentally, it’s also becoming clear once again that the AI boom is not just scaling upward, but also bringing hard corporate reality with it. Welcome to today’s mix of supercomputers, rules, and restructurings.

🚀 Anthropic secures 300 megawatts from SpaceX

According to reports, Anthropic has secured the entire compute capacity of SpaceX’s Colossus-1 data center—including more than 220,000 NVIDIA GPUs and over 300 megawatts of power. That’s not just “more compute,” that’s a statement on a different scale: we’re talking about infrastructure that can not only train models, but also run them at a completely different magnitude. At the same time, Anthropic is raising rate limits for Claude Code and allowing higher API limits for the Opus models. For developers, that means very concretely: fewer bottlenecks, more throughput, more productivity promises.
Also interesting is the aside about “orbital AI compute” — the idea of moving data centers into space in the future. It sounds like science fiction, but given the current appetite for compute, it’s at least no longer a completely absurd thought experiment.
Source: The Decoder

🧠 Values first: New alignment method drastically reduces misbehavior

A study from the Anthropic Fellows Program shows that language models can be aligned much better with values if they first understand why they are supposed to behave in a certain way — and only then learn the concrete behavior. That sounds almost trivial, but it’s quite important for alignment and fine-tuning: instead of merely hammering in rules, a model is trained with documents that explain its values. For Qwen3-32B, the misalignment rate dropped from 54 to 7 percent, using 10 to 60 times less data than previous approaches.

The practical takeaway: if these results prove robust, alignment could become much more efficient and accessible—especially for teams without gigantic data budgets. The idea also fits a broader insight from AI research: models often behave better when they learn not only the “what,” but also the “why.” Surprisingly human, unfortunately even for the difficult parts.
Source: The Decoder

🌐 OpenAI and chip giants build MRC for AI supercomputers

OpenAI has developed the open-source network protocol MRC together with AMD, Broadcom, Intel, Microsoft, and NVIDIA. The goal is to distribute data simultaneously across hundreds of paths between GPUs and thus make supercomputers more efficient. The technical effect is enormous: instead of three to four switch layers, MRC is supposed to get by with two layers and thereby connect more than 100,000 GPUs more effectively. The protocol is already running in OpenAI’s Stargate supercomputer.

Why does this matter? Because AI infrastructure is long past the stage of simply “buy more GPUs.” Network architecture increasingly determines how much real performance a cluster can actually deliver. Open source is also important here: if such a protocol becomes broadly usable, it won’t just benefit the biggest players. For the rest of the industry, this means networking is turning from an invisible detail into a strategic competitive advantage. And yes, the same applies here: the bottleneck is often not the model, but the cable behind it.
Source: The Decoder

⚖️ EU delays AI rules and bans nudification apps

The EU has agreed on a simplified package for AI rules: the “Digital Omnibus on AI” pushes deadlines for high-risk AI to the end of 2027 or 2028 and eases the burden on small and medium-sized businesses. At the same time, so-called “nudification” apps are being banned—applications that undress or sexualize people via AI without consent. The labeling requirement for deepfakes and AI-generated text, however, remains at the original deadline of August 2026.

For companies, this is a classic case of “more time, but not infinite time.” Anyone working in regulated areas gets breathing room for implementation and compliance. For the EU, the line is clear: less bureaucracy in some places, but clear red lines on abuse and manipulation. Particularly relevant is the fact that labeling requirements were not postponed—exactly the issues that are central to trust in generative AI. In short: relief, yes; a free pass, no.
Source: The Decoder

🧩 Claude’s “Dreaming” is meant to let agents learn from mistakes

Anthropic is expanding its “Claude Managed Agents” with a feature called “Dreaming.” Behind it is an asynchronous process that analyzes past agent sessions, cleans up duplicate or outdated memories, and distills new insights. Along with features like “Outcomes” and “Multiagent Orchestration” in public beta, Claude is supposed to become more capable of learning across sessions.

This is interesting because memory is often the point where the hype collides with reality in AI agents. An agent that remembers things is nice. An agent that intelligently selects what to keep and what to forget is much more useful. “Dreaming” may sound like marketing with a pillowcase, but it addresses a real problem: how do agents stay consistent over the long term without drowning in old junk? Anyone building with agents should keep an eye on this feature—especially for workflows that run over several days or tasks.
Source: The Decoder

🏭 DeepL cuts around a quarter of its workforce

DeepL, one of Germany’s best-known AI translators, is laying off around 250 employees—roughly a quarter of its workforce. CEO Kutylowski says the move is part of a broad restructuring. That’s obviously a tough message for the company and those affected; at the same time, it shows how strongly even successful AI companies are under cost pressure, scaling pressure, and strategic realignment.

DeepL is interesting to the market because the company long stood as a showcase example of specialized AI products: strong product, clear benefit, internationally visible. The fact that it is now being restructured so visibly is a reminder that even good products are no guarantee of stable organizations. In the current phase of the AI economy, it’s not just technology that’s being optimized, but also the business structure behind it. Unfortunately, that’s often the part where Excel suddenly has more power than the model.
Source: heise

🌍 US and China plan official AI talks

The US and China are considering official talks about the risks of artificial intelligence, according to the Wall Street Journal. Politically, this is more than just symbolism: if the two biggest AI powers are actually talking to each other about safety, then the issues include escalation risks, misuse, model control, and possibly transparency around especially capable systems.

Of course, such talks won’t dissolve the geopolitical tensions around chips, export controls, and technological leadership. But they can open an important channel of communication—and that’s not to be underestimated with a technology moving this fast. For the industry, this matters because regulation and safety standards increasingly have to be thought about internationally. If there is movement here, it’s not because everyone suddenly became best friends, but because even rivals realize: with AI risks, not talking is often the more expensive option.
Source: The Decoder

🛠️ Tool tip of the day: take a closer look at OpenAI MRC

If you’re interested in AI infrastructure, GPU clusters, or network architectures, this is today’s topic to read up on: MRC could become an important building block for the next generation of AI supercomputers. It’s especially relevant for teams working with distributed training, large inference setups, or high-performance networking.
Affiliate note: #

Don’t want to miss any news? Subscribe to the newsletter