AI power games, Google’s liability, and Claude Fable 5

Today’s about three things AI is especially good at right now: getting faster, building smarter tools, and causing legal trouble. Between new frontier models, agentic research workflows, and liability issues for AI answers, one thing becomes pretty clear: the next phase of AI development is not just technical, but also politically and legally charged.

🚀 Anthropic pushes Claude Fable 5 and Mythos 5 forward

Anthropic has introduced Claude Fable 5 and Mythos 5, two new models that are said to be significantly stronger than the current Opus generation, especially for coding and scientific work. What’s especially interesting is the real-world angle: Fable 5 reportedly completed a code migration at Stripe in one day, a task that would otherwise have taken a team two months. That’s the kind of number that gets CTOs paying attention and developers looking both nervous and curious. Source: The Decoder

For the market, this is more than just a model update. Anthropic is positioning itself even more clearly as a provider for demanding workloads: software development, research, complex analysis. Mythos 5 goes even further and is said to have autonomously designed drug candidates — but remains under wraps for now because of offensive cyber capabilities. This shows how strongly frontier models are now moving along the line between useful tool and risk. For you, that means: the best models are becoming not just smarter, but also more tightly separated by use case.

⚖️ Google is liable for false AI Overviews

The Munich Regional Court I has dealt Google a legal setback: for false answers in AI Overviews, Google is liable as a direct interferer. The previous, more limited liability for search engine operators cannot simply be transferred to AI-generated summaries. Specifically, the case involved false claims linking two publishers to fraud schemes — statements that did not appear in any of the linked sources. Source: The Decoder

The ruling matters because it could extend beyond this individual case. When AI systems generate content that looks like editorial statements, the operator’s responsibility increases too. That is uncomfortable for Google; for everyone else, it’s a pretty loud reminder that “we’re only showing summarized sources” is not automatically a legal shield. What will be especially interesting is how this ruling affects other AI products that deliver answers instead of mere results. The legal framework for generative search just got a good deal more serious.

🧠 NotebookLM becomes a real research agent

Google is giving NotebookLM a major upgrade: the tool now runs on Gemini 3.5, gets its own cloud computer for code execution, and can independently find sources via Google Search. That turns NotebookLM from a smart note-taking and research assistant into a much more agentic system that handles more steps on its own. Source: The Decoder

This is especially interesting for people who work with lots of documents, sources, and notes. The new version promises not just to organize research, but to actively structure it and connect it with tools. Internally, Google cites a win rate of up to 78.2 percent compared with the previous version — so not just a small tweak, but a real jump. But one thing is important: the more a tool researches and runs code on its own, the more important transparency, source checking, and control become. Otherwise you quickly end up with a very diligent but overconfident machine for mistakes.

⚙️ DiffusionGemma brings a new idea to text generation

With DiffusionGemma, Google is releasing an open language model with 26 billion parameters that generates text not in the classic token-by-token way, but via diffusion. You probably know the principle from image generation: results are formed step by step out of noise. According to Nvidia, the model reaches around 1,000 tokens per second on a single H100 GPU — about four times faster than comparable autoregressive models. Source: The Decoder

That’s exciting because it shows that text models do not necessarily have to work the way existing LLMs do. If diffusion approaches take hold, they could open up new paths for speed and latency — especially for applications that need to respond very quickly. The catch: the speed gain comes at the cost of quality, and that is exactly where the big research question lies. DiffusionGemma is therefore less an immediate product than a signal of where model architecture might be headed. In short: Google is not only competing on scale, but also on the mechanics underneath it.

🌏 China is building its own AI data center network

According to Bloomberg, China plans to invest around US$295 billion over the next five years in a nationwide AI data center network. The focus is clearly on domestic technology: at least 80 percent of the infrastructure is expected to come from vendors such as Huawei. At the same time, Taiwan is considering criminalizing the smuggling of AI chips to China for the first time. Source: The Decoder

This is geopolitically significant because data centers are no longer just infrastructure, but instruments of power. Whoever controls the chips, networks, and computing capacity also controls part of AI development. For Nvidia and AMD, that is far from good news; for China, it is another step toward technological independence. At the same time, the plan shows how much AI is now understood as a strategic state task. While the West debates productivity, China is simply building the foundation for the next decade.

💸 “AI-pilled” companies are spending serious money per employee

A new analysis from Ramp shows that especially AI-obsessed companies are spending an average of around $7,500 per employee per month on AI. Of course, that’s no more than an engineer’s salary — at least for now. Source: TechCrunch

The number is interesting above all because it makes the cost curve of AI applications in everyday business visible. Between API fees, tool licenses, and specialized workflows, the total quickly adds up to an amount that can be painful for smaller teams. At the same time, it shows that if you deeply integrate AI into processes, you’re no longer investing in just “a tool,” but in an entire operational layer. The big question, of course, remains whether these expenses will sustainably translate into productivity or whether companies are currently funding very expensive collective experiments. The truth probably lies — as so often — somewhere in between.

🛡️ Germany is planning an AI safety institute

The National Security Council has approved the creation of a German AI safety institute. The planned DE-AISI is set to test frontier models for safety risks, following the British model, including systems from Anthropic or OpenAI. Source: The Decoder

At first glance, that sounds like sensible oversight, but it is also a reminder of how dependent Europe still is on US and Chinese providers when it comes to frontier AI. As long as we lack our own top-tier models, we are mainly testing technologies built elsewhere. Still, such an institute is important: it can set standards, make risks visible, and translate political pressure into concrete testing procedures. For the AI governance debate, this is an important building block — not glamorous, but pretty necessary. Because if AI keeps getting more powerful, “we’ll take a look at it” is a rather mediocre safety strategy.

🛠️ Tool tip of the day

If you want to combine complex research with sources, notes, and AI assistance, take a look at NotebookLM. With the new Gemini 3.5 upgrade and its agentic features, it becomes a strong tool for knowledge work, research, and document analysis. Handy for anyone who doesn’t just want to search, but really wants to understand. #

---

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