Nadella Calls Out AI Labs' "Reverse Information Paradox"

Microsoft CEO Satya Nadella has leveled an interesting critique at OpenAI and Anthropic. He calls it the "reverse information paradox."

At the heart of the issue is a contradiction: AI labs train on publicly available data under fair use, yet prohibit distillation (extracting and transferring knowledge) from their own models — even as they also learn from customer interactions. Nadella argues that enterprises should control their own learning infrastructure, a claim that also comes in the context of Microsoft selling exactly that infrastructure, making the remarks a reflection of the industry's power dynamics.

First End-to-End Ransomware Attack Carried Out by an AI Agent Documented

Security firm Sysdig has documented the first ransomware attack executed entirely by an AI agent, from initiation to completion. It's a concrete example that AI agents can function not just as productivity tools but as attack instruments as well.

In this incident, the AI agent is said to have autonomously carried out everything from reconnaissance to encryption and the ransom demand — suggesting that, even as AI adoption on the defensive side becomes increasingly important, offensive capabilities are becoming a real-world threat too.

Study: AI Agents Cost "136.5x" More Than Traditional AI

According to coverage in The Engineer, a new analysis found that AI agents consume up to 136.5 times more resources than traditional AI approaches. The main driver is compute costs ballooning through autonomous reasoning loops and repeated tool use.

As AI agent adoption accelerates, this efficiency gap could become a major barrier to enterprise deployment. How to resolve the "smart but expensive" trade-off will likely be the key to future architecture design.

Zhipu/Z.ai Founder Defends Open-Source AI Amid Security Debate

The founder of Chinese AI company Zhipu AI (Z.ai) has emphasized the importance of open-source AI as global security debates intensify. Taking the position that open source is essential to democratizing innovation, the stance contrasts with the restrictive approach advocated by some U.S. AI labs.

The remarks once again highlight how the international debate over open-sourcing AI is deeply entangled with geopolitical considerations, going beyond purely technical choices.

Apple M7 Ultra Reportedly Plans Up to 1.5TB of Unified Memory

According to reports in Reddit's LocalLLaMA community, Apple's planned M7 Ultra chip could feature up to 1.5 terabytes of unified memory. Considering that the current M3 Ultra already supports up to 192GB of unified memory, this would be a leap of a different order of magnitude.

For communities pursuing large-scale local LLM inference, unified memory of this size could make it possible to run multiple large models simultaneously, and it is drawing significant attention. That said, note that this is still planning-stage information and not yet confirmed.

Reinforcement Learning Authority Richard Sutton Departs Keen AI to Found Oak Lab

Richard Sutton, known for foundational theory in reinforcement learning, has left Keen AI, where he had been based, to launch a new venture, Oak Lab. Sutton is widely known as co-author of "Reinforcement Learning: An Introduction," the standard textbook in the field, and this move is drawing attention in the AI research community.

Details remain limited at this point about what research approach Oak Lab will take, but it could influence the direction of the reinforcement learning field.

---