Serge Bevz
Researcher analyzing AI/data‑center infrastructure, energy limits, and power/cooling constraints shaping sustainability.
Amazon Halts Q Developer Sign-Ups; Kiro Replaces IDE
Amazon blocked new Q Developer sign-ups on May 15. VS Code plugin, JetBrains, Eclipse, Visual Studio — all frozen for new users. Existing Pro subscribers get until April 2027, then full shutdown. The replacement is Kiro — standalone IDE built around spec-first coding.
Claude Goes Proactive: AI Acts Before You Ask
Anthropic says proactive AI in six months. Claude learns your workflow, starts doing things before you ask. No prompt. It just acts. Three hard problems for anyone building this.
Anthropic Launches Claude Platform on AWS, Self‑operated
Anthropic just shipped Claude Platform on AWS. Not Bedrock. Not a marketplace listing. Anthropic operates it. AWS handles auth and billing. That's the whole architecture.
AI Breakthroughs Can't Shield Companies From AI-Driven Cuts
Yesterday I published a post praising Cloudflare's engineering. Their Rust inference engine beating vLLM 2x. Brilliant stuff. Today they announced 1,100 layoffs. That's 20% of the company. Because of AI. I'm genuinely shocked. And a little uncomfortable.
Pennsylvania Sues over Chatbot Therapy Misuse
People use chatbots as therapists. Describe symptoms, ask about medication. We've been doing it for years. Pennsylvania just put it in a lawsuit.
When Optimization Costs More Than Redesign, Rethink the Tool
There's a point where optimizing a system costs more than redesigning it. Most engineering teams reach it and keep optimizing anyway — because redesign feels risky. Cloudflare hit that point with LLM inference and did something unusual. They stopped tuning vLLM for their...
Anthropic's New Tools Aren't New—Own the Stack Yourself
Anthropic shipped Dreaming, Outcomes, and Multiagent Orchestration for Managed Agents. None of this is conceptually new. LangGraph + MCP + vector DB + eval loops — you can build all of it yourself. The question is not capability. It's whether you want to own...
DeepMind Trains AI on 23 Years of EVE Online Decisions
EVE Online — if you know, you know. If you don't, it just sounds strange. DeepMind is now training AI inside it. On 23 years of player decisions — real people, real consequences.
CopilotKit's $27M AG-UI Solves a Trivial Sync Issue
CopilotKit raised $27M for a protocol called AG-UI — keeps agent state and frontend in sync when the agent is streaming, calling tools, or waiting for input. My first reaction: that's not a hard problem. I've written this glue code.
IOS 27 Lets Users Pick AI Model for iPhone Features
iOS 27 is adding a setting that lets you pick which AI model handles your iPhone's features. Writing suggestions, image generation, Siri responses — all switchable. That's 1.56 billion iPhones where the AI model is no longer fixed at the factory.
Weekly Cap Still Throttles Heavy Users Despite Doubled Limits
Anthropic doubled Claude Code limits. I still hit the weekly cap. I'm on Max — the plan for heavy users. My bottleneck isn't the 5-hour window. It's the weekly one. The announcement doesn't touch it.
Anthropic, OpenAI Mirror Enterprise JV Playbook, Signaling Market Shift
Anthropic and OpenAI announced enterprise joint ventures on the exact same day. Both with private equity firms. Both using the same playbook. That's a market signal.
Human QA Staff Become AI Trainers Through Feedback
A Chinese man spent years at a tech company doing one specific thing. He reviewed AI outputs. Caught mistakes. Made the model better. His title was QA supervisor. His actual job was teaching an LLM to do his job.