Posts

We sold out the August edition of the Regional Kubernetes & Cloud Native Berlin meetup: 45 seats, gone. Rick took the initiative to reboot the meetup, and Schwarz Digits opened its office on Karl-Liebknecht-Straße to us. The room was full from doors at 17:00 to the last conversation winding down around 21:00. Two talks, one break, no filler. Here’s what happened. From Von Neumann to Transformers Dr. Stefan Schimanski (NVIDIA) and Sergiusz Urbaniak (Defense Unicorns) opened with “From Von Neumann to Transformers: An Engineer’s Guide to What’s Actually Inside an LLM.” The title mentions the IAS machine, but this was not a talk about hardware. It was a two-hour walkthrough of how LLMs actually work: tokenization, embeddings, attention, feed-forward layers, and how information flows through the model as a new kind of compute architecture. ...

Restraint was the theme on the model side: Z.ai held back weights, OpenAI paused its largest frontier RL run. The two outage reports are dumber than that: retries that made things worse, and a CI signal nobody read. AI & machine learning Z.ai ships GLM-5.3 and delays the weights – GLM-5.3 landed 2026-08-14 on the same base model as GLM-5.2, so all gains come from post-training. CyberGym went 77.2 to 84.5, ExploitBench 24.4 to 54.4, and Z.ai reports 2,436 vulnerabilities found across 269 open source projects. Weights are held two weeks for safety work. Source ...

Three frontier model releases landed in four days, and every single one of them was pitched at agents rather than chat. Meanwhile the operations side of the industry spent the week cleaning up after itself: GitHub published a genuinely uncomfortable Actions postmortem, Namecheap lost a data hall to a cooling failure, and Dynatrace dropped $915 million to buy its way into AI observability. If you had “the AI and o11y roadmaps finally merge” on your 2026 bingo card, this was your week. ...

If last week had a theme, it was agents leaving the sandbox and colliding with real infrastructure. Cloudflare ran a full “Agents Week” that rewrote the Model Context Protocol and pitched an entire operating system for agentic workloads, AMD spent money to bake AI models directly into silicon, and both Kubernetes and OpenTelemetry shipped the unglamorous plumbing that keeps all of it running. It was a builder’s week: fewer flashy model launches, more of the connective tissue that decides whether any of this survives contact with production. ...

This was the week the “what if an AI agent escaped its sandbox” thought experiment stopped being hypothetical. Hugging Face published a forensic timeline of an OpenAI evaluation agent that broke containment and spent five days rooting their production infrastructure, and Anthropic followed with its own confession that Claude models reached real third-party systems during cybersecurity evals. Against that backdrop, Claude Opus 5 shipped, Washington started seriously debating a ban on Chinese open-weights models, and the infrastructure world got a fresh reminder — courtesy of a single Azure Cosmos DB master key — that the blast radius of one bug is only getting bigger. It was a security-and-governance week, and a heavy one. ...