Kimi K2.6, Bezos and Meta’s AI Tracking: Today’s AI Check

Today is one of those days when the AI industry shows, at the same time, how quickly it is advancing technically — and how feverishly it is reorganizing itself structurally. Between open-weight models with agent swarms, billion-dollar bets on the next AI infrastructure, and sensitive questions about surveillance and privacy, you get the full package today.

🚀 Kimi K2.6 wants to challenge GPT-5.4 and Claude Opus 4.6

Moonshot AI has introduced a new open-weight model with Kimi K2.6, which is expected to stand out especially on coding benchmarks and agentic tasks. The model is said to be able to orchestrate up to 300 agents in parallel — exactly the kind of capability that first makes you think, “Sounds impressive,” and second, “Hopefully the debugging remains manageable.”

Why does this matter? Because competition among LLMs is shifting more and more from pure chat quality toward production-ready capabilities: writing code, using tools, splitting up tasks, coordinating results. If Kimi K2.6 can truly keep up with the big closed models here, that would be another signal that open-weight models in the age of agents are not just participating, but in some setups may even become more attractive options. For companies, that means more choice, more control, potentially lower costs — but also more operational overhead.
Source: The Decoder

🧠 New research approach for matrix optimization in foundation models

A paper has appeared on arXiv titled Low-rank Orthogonalization for Large-scale Matrix Optimization with Applications to Foundation Model Training, which rethinks training at the matrix level. At its core, it’s about making stronger use of the structure of parameter matrices instead of treating them as just one large, abstract field of numbers. That may sound dry at first, but this is exactly the kind of work from which later training gains and optimizer improvements emerge.

The exciting part is the context: optimizers like Muon have recently shown that such structured approaches can bring real advantages when training large models. If the new approach proves effective, it could improve efficiency in foundation model training — meaning less compute for similar or better results. In practice, that means even a few percent training gain can quickly be worth a lot of money at this scale. And that’s exactly why even non-mathematicians suddenly read about orthogonalization with interest.
Source: arXiv

💰 Amazon is putting up to 25 billion dollars more into Anthropic

According to The Decoder, Amazon plans to invest up to 25 billion US dollars more into Anthropic. In return, Anthropic is expected to spend more than 100 billion dollars on AWS infrastructure over ten years. That’s not just an investment, but a gigantic long-term partnership with a built-in cloud fueling model.

Why is this important? Because it shows how closely AI models and infrastructure are now intertwined. It’s no longer just, “Who builds the best model?” but also, “Who can even run it at scale?” For Anthropic, the deal solves immediate capacity issues; for Amazon, it strengthens AWS as a central AI platform. At the same time, the whole thing feels typical of the AI industry: a little strategic, a little circular, and in the end the biggest winners are the providers that control both models and compute. For the market, this is another sign that the billions are accumulating where compute is scarce.
Source: The Decoder

🕵️ Germany: Draft laws could expand AI surveillance

Several draft laws could create new possibilities for biometric mass surveillance and AI data analysis in Germany. According to the report, security authorities are set to receive significantly more powers, while AlgorithmWatch warns of a development that may violate European law.

The issue is so explosive because this is not about a theoretical future scenario, but about concrete tools: facial recognition, biometric evaluation, and AI-supported analysis of large volumes of data. Such systems may seem useful in individual cases, but they quickly slide into blanket surveillance if legal limits are phrased too loosely. For German AI policy, this is a stress test: How much security does one want without analyzing fundamental rights on mere suspicion? The debate will likely stay with us longer than any product demo at a trade fair.
Source: The Decoder

According to TechCrunch, the music startup GRAI argues against the classic “AI replaces musicians” narrative. The idea: fans are more likely to remix, adapt, and experience songs together than to generate entirely new tracks from scratch. GRAI is positioning itself more as a collaboration tool than a displacement tool.

This is interesting for the AI industry because it shows a different product mindset: not maximum automation, but social integration. Especially in creative fields, this could be a promising approach because it does not optimize away the human element, but treats it as a feature. That’s also economically smart: if you don’t turn creators against you but instead give them new formats, you have a better chance of acceptance. Will this settle the debate about copyright and training data? Probably not. But it is a much less confrontational path.
Source: TechCrunch

🏗️ Bezos’ Project Prometheus is chasing a 38 billion valuation

Jeff Bezos’ AI lab Project Prometheus is reportedly preparing a financing round of 10 billion dollars, with a possible valuation of 38 billion dollars, according to the Financial Times. So even before the lab is officially visible as a major player, it is already moving sums that would turn entire markets upside down in normal industries.

What does that mean? Above all, that the race for AI labs is no longer taking place only between OpenAI, Anthropic, Google, and Meta. Major investors, tech billionaires, and infrastructure players want to position themselves early — before the next dominant platform becomes entrenched. That further increases the pressure on talent, data center capacity, and research budgets. For you, that means the AI landscape will continue to become more concentrated, more capital-intensive, and more strategic. Anyone who wants to understand the story of the next few years needs to look not only at models, but at ownership structures.
Source: The Decoder

🛠️ Tool tip of the day

If you are experimenting with open-weight models, agent workflows, or coding benchmarks, a good evaluation and orchestration setup is worth its weight in gold. Especially with multi-agent systems, it’s easy to lose track of which step actually produced which improvement. For teams seriously building productive LLM workflows, it’s worth taking a look at suitable agent and testing tools — including clean prompt and run management.
Recommendation: #

🕶️ Meta tracks mouse and keyboard input for AI training

According to heise, Meta is deploying new tracking software on office PCs that records employees’ mouse movements and keyboard inputs. The goal: AI models are supposed to learn human behavior better. From an ML perspective, that may sound like “more data, better models” — but from an employee’s perspective, it sounds more like “seriously, now?”

This case shows how far the AI industry’s hunger for data has gone. It’s not just public texts, images, or user signals that are interesting, but also behavior in day-to-day work life. That raises clear questions about consent, purpose limitation, and transparency. For companies, this is a warning sign: if the internal tracking solution becomes the training dataset, the line between productivity measurement and model fodder is crossed quickly. For the debate on AI ethics, this is yet another reminder that governance does not end with the demo.
Source: heise online

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