Workers teach robots. Anthropic leaks itself. Builders get tools.

Thursday brought three distinct stories worth your attention, each revealing something about how AI moves from research into the real world.

Teaching robots through demonstration

A Czech startup called RoboTwin is solving a specific manufacturing problem: how to make robot programming accessible to people without code. Instead of writing instructions, factory workers demonstrate a task once. The system captures their movements and converts them into robot programmes-spray painting, surface treatment, that kind of work. The result is real. They've already deployed with companies across Central Europe, and just secured €2.3 million in EU funding to scale. What matters here isn't the novelty-imitation learning is decades old. What matters is that small manufacturers can now automate tasks that were previously too expensive to program. Labour shortages in surface treatment are real. The cost barrier was real. This removes it.

The Claude Code leak and what it revealed

Anthropic accidentally exposed 512,000 lines of Claude Code source code through a missing .npmignore file in their npm registry. The incident itself is embarrassing. The aftermath is interesting. The leak showed a deceptively simple architecture: a single event loop coordinating multiple tools, with sophistication pushed into context management (a four-layer compression stack for handling long conversations), telemetry, and product polish. Several developers immediately re-implemented parts of it as open-source frameworks. Most importantly, Anthropic's response was measured-DMCA takedowns initially overstretched, then corrected quickly. The lesson here isn't that their code was exposed. It's that once code is exposed, open-source clones follow within hours, and the competitive advantage shifts entirely to product, infrastructure, and trust. The code itself becomes commodity.

Building with live web data

Meanwhile, builders are getting better tools for what's become essential infrastructure: turning webpages into structured data for LLMs. A tool called webclaw strips 67% of token waste from typical webpage extraction-navigation, footers, cookie banners all removed before it reaches your model. It includes MCP integration (the USB standard for AI tools), so Claude Code and other agents can call it directly during conversations. The practical value is straightforward: your RAG pipelines get cleaner embeddings. Your agents conserve context window. Your API bills drop. This is infrastructure work-unsexy, necessary, and increasingly competitive.

Three stories, three different distances from hype. RoboTwin is solving a real manufacturing problem today. The Claude leak is already reshaping the competitive landscape for coding agents. webclaw is quietly becoming essential plumbing for anyone building LLM applications. None of them announced anything significant. All of them moved the work forward.