AI Breakthrough, Security Risk, and EU Dispute in Today’s Snapshot
Today’s agenda is shaped by new research on LLMs, stronger AI security, controversy over cloud scanners, and the next boost for multimodal models.
Inhaltsverzeichnis
Today brings action on several fronts at once: new research is making low-precision training and reasoning models more efficient, while the alarm bells are getting louder in AI security. Add to that a clear reality check for AI subscriptions and agents, progress in volumetric videos, and a politically sensitive dispute over EU scanning practices. In short: lots of movement, very little boredom.
🧠 AdaHOP: Better Low-Precision Training for LLMs
Low-precision training is one of the most important levers for training large language models faster and more cheaply. The new paper AdaHOP tackles the problem differently from many existing methods: instead of applying a fixed rotation or transformation indiscriminately to all tensors, it detects different outlier patterns and adapts the rotation accordingly. This matters because outliers in weights and activations can quickly throw quantization off balance.
The practical effect: fewer errors, better stability, and potentially much more efficient training for large LLMs. In other words, if quantization has often sounded like “it kind of works,” this shows just how much the details matter. And yes, in AI research, progress sometimes means someone finally looks at the weird values in the data instead of sweeping them under the mathematical rug.
Source: arXiv: AdaHOP
⚙️ Qwen Team: Reasoning Models Think Deeper
The Alibaba Qwen team presents a new approach that weights reinforcement learning for reasoning models more finely. The key idea: not every token gets the same “credit.” Instead, the algorithm evaluates how strongly a step influences the subsequent chain of reasoning. That may sound minor, but it is an important lever, because classic reward systems are often too coarse for complex thinking.
According to the report, the approach even doubles the length of thought processes. That is interesting because longer chains do not automatically guarantee better answers — but they can indicate deeper internal processing. For developers and product teams, this means reasoning remains an active research area, where the quality of the thinking strategy matters at least as much as model size. Anyone relying on open-source LLMs should keep an eye on such training ideas, because that is exactly where competition is shifting.
Source: The Decoder: Qwen Team Gets AI Models to Think More Deeply
🧩 LiME: Lightweight MoE for Multimodal Training
Mixture-of-Experts remains one of the most exciting concepts for scalable models — but also one of the most annoying when fine-tuning becomes unnecessarily expensive. That is exactly where LiME comes in: the method combines expert specialization with parameter-efficient multitask fine-tuning, without the trainable parameters exploding linearly with each additional expert. This is especially important for multimodal models that need to process text, images, or other inputs together.
Why does this matter? Because many teams find MoE appealing, but fail because of the complexity. LiME could lower that barrier and make MoE approaches more practical in everyday use — especially for product teams that do not want to turn the whole data center into a furnace. For ambitious newcomers, the short version is: more specialization, less overhead. A rare moment when research and pragmatism seem to briefly shake hands.
Source: arXiv: LiME
🚨 AI Offensive Capabilities Are Growing Faster Than Many Would Like
A new study shows that the offensive cyber capabilities of AI models are rising rapidly — reportedly, progress has doubled every 5.7 months since 2024. Particularly concerning: models like Opus 4.6 and GPT-5.3 Codex are said to solve tasks that take experienced experts several hours. That is not a nice marketing feature, but a genuine security problem.
The takeaway is clear: AI is not only getting better at programming, but also at exploiting vulnerabilities. That shifts pressure onto security teams, pentesting, and incident response. Companies should therefore not only think “AI for security,” but urgently also “security against AI.” In practice, that means more red-teaming, more defensive testing, and less faith in the hope that attackers will eventually get tired. They rarely do.
Source: The Decoder: Danger from AI Hacks
📹 Volumetric Videos Are Finally Becoming More Mobile-Friendly
Volumetric videos have long been fascinating, but hard to distribute. Now there is an important breakthrough: dynamic Gaussian splats can be streamed in good quality to mobile devices and VR headsets. That is a real milestone for XR applications, because it slowly turns a lab feature into a product feature.
Why is this relevant? Because volumetric content is much more immersive than classic 2D video — especially for VR, training, entertainment, or virtual events. The main bottleneck until now was usually performance. If streaming and compression now work better, the mainstream gets closer. It is not yet the “3D Netflix from your phone” future, but it is already more than just a nice demo on a conference table.
Source: heise: Volumetric Videos on the Way to the Mainstream
🛠️ Tool Tip of the Day: Test Cloud and Agent Workflows Properly
When AI agents and tools become production-ready, “works locally” is no longer enough. A tool stack for workflows, API tests, and observability makes sense so you can systematically check agents, model calls, and tool calls. Especially with the current developments around third-party tools and agent usage, it is worth documenting dependencies cleanly from the start. For teams moving toward production agents, a reliable testing and monitoring stack is almost mandatory.
If you are looking for suitable solutions for this: #
🔒 Anthropic Blocks Third-Party Tools for Claude Subscriptions
Anthropic is drawing a clear line with Claude: third-party tools like OpenClaw are being blocked for subscription users. The reason is quite sobering and also very typical for the industry: flat-rate plans and agent-based continuous use often do not mix. Once a model is not just responding, but triggering tools for minutes or hours, the cost model quickly becomes questionable.
That is annoying for users, but understandable for providers. Agents generate many calls, high load, and unpredictable usage patterns — hard to reconcile with a flat monthly subscription. The case also shows how sensitive the market currently is to product boundaries. Anyone relying on AI workflows should therefore not only look at features, but also at terms of use, the tool ecosystem, and long-term platform strategy. Otherwise, you end up building smartly on sand — only with more expensive sand.
Source: The Decoder: Anthropic Blocks Third-Party Tools
🕵️ EU Chat Control: Scanning Remains Politically Sensitive
The topic of chat control remains complicated: the legal basis for indiscriminate searches for abuse material has expired, yet Google, Meta, Microsoft & Co. continue the practice. This is highly sensitive politically, because security interests, child protection, and privacy are colliding head-on here.
For the debate around AI and platforms, this is a lesson: technical possibilities do not disappear automatically just because the legal basis is shaky. Conversely, “can be done” does not mean “should be done.” Especially in Europe, this issue is likely to keep causing disputes — between regulation, platform interests, and fundamental rights. Anyone planning AI-assisted scans or content monitoring should understand this tension very precisely. It is rarely a good idea to treat data protection issues as casually as a cookie banner.
Source: heise: Tech Giants Want to Keep Scanning
Don’t want to miss any news? Subscribe to the newsletter