Saturday 23 May 2026

LLMs are scraping and summarising web content in ways that undermine site owners' intent, creating friction around how sites should structure or restrict content for AI consumption. The post explores the emerging tension between LLM crawling behaviour and content owners' control, including the llms.txt convention as a partial workaround.

yt-dlp is deprecating and limiting Bun runtime support, forcing users who rely on Bun-based JavaScript interpreters in the tool to migrate or lose functionality. High comment engagement signals significant user disruption.

LLMs performing 3D CAD generation in OpenSCAD expose a gap: existing models struggle with structured, constraint-based CAD languages, requiring a dedicated benchmark to measure architectural code generation quality. Results highlight that current frontier models have meaningful accuracy gaps on structured geometric code.

Microsoft is cancelling Claude Code licenses for employees, abruptly removing access to an AI coding tool that developers had been relying on, with no clear replacement path signalled.

LangChain's `convert_to_openai_function` raises a `TypeError` when a `TypedDict` contains `NotRequired` fields, because the internal conversion logic incorrectly re-wraps already-unwrapped type arguments through `NotRequired` again. Standard Python typing patterns fail silently or crash at schema conversion time.

LangChain tool definitions are model-agnostic, causing small local models (1.5B) to waste thousands of prompt tokens on tool descriptions and parameters they can't effectively use, resulting in ~50% tool selection accuracy with large toolsets. A tier-aware tool metadata system is proposed to adapt descriptions and parameter sets based on model capability.

When binding tools with deeply nested Pydantic v2 models to a LangChain chat model, the tool invocation fails to correctly parse or honour the Pydantic v2 schema, causing `AIMessage.tool_calls` args to be arbitrarily or incorrectly generated rather than following the defined model structure.

CI/CD workflow permissions on GitHub can be exploited at scale to backdoor repositories, highlighting a systemic security gap in how CI pipeline configurations are trusted and reviewed across the ecosystem.

An upcoming Secure Boot Certificate Authority rollover risks breaking boot chains for Linux distributions that haven't updated their signed bootloaders in time, creating operational friction and potential bricking risk for distributions and users relying on Secure Boot.