Jason Says
The headline move today is Cursor's official Plugins repo going public — this isn't a minor update, it's a play to replicate the VSCode ecosystem. Whoever claims the high-traffic plugin slots in Cursor first will lock in developer mindshare early.
Harness is a 'team factory' plugin for Claude Code — describe your project in natural language and it automatically designs an agent team, assigns role responsibilities, and generates corresponding skills. Supporting English, Chinese, and Japanese instructions with six preset team architectures, it offers the closest experience yet to assembling an AI engineering squad inside Claude Code.
Cursor has open-sourced its official plugins repository, cataloging plugins for major dev tools, frameworks, and SaaS products — each stored as a standalone directory with a plugin.json manifest. The move signals Cursor is building a VSCode-style plugin marketplace, with developers now able to submit their own plugins directly to the official channel.
SoftBank announced an investment of up to €75 billion to build data centers in France, targeting 5GW of new compute capacity. The largest single AI infrastructure bet in Europe to date — surpassing anything seen outside the US — it signals that the global AI compute race has reached Europe, with EU AI sovereignty strategy now attracting top-tier capital.
Prominent investor Bindu Reddy notes GPT-5.5 holds a clear coding edge that Opus 4.8 hasn't closed; GPT-5.6 could widen the gap further. With Anthropic's recent releases still failing to overtake OpenAI on coding benchmarks, OpenAI's moat in the coding track deserves a fresh look.
Microsoft's open-source MarkItDown is back on GitHub Trending, converting Office documents, PDFs, images, and more to Markdown in bulk — making it a go-to tool for RAG pipelines and Claude Code knowledge bases. The latest version adds security guidance recommending restricted call scope in untrusted environments, improving readiness for production deployments.
A new 'reflective prompt tuning' technique has LLMs use function calling to self-evaluate and iteratively refine prompts — no large candidate searches or fixed templates required. For developers, system prompt engineering can now be semi-automated: the model identifies issues, rewrites, and validates on its own, significantly cutting the human labor cost of prompt engineering.
Research reveals a systematic 'vertical-distance entanglement' flaw in mainstream vision-language models: they tend to conflate 'distant' with 'above,' a statistical shortcut rather than genuine 3D understanding. This is a critical warning for developers building multimodal agents, robotic vision systems, or spatial reasoning applications — current VLM spatial judgments should not be over-trusted.
研究识别出 LLM「隐性政治偏见」的 7 类表现方式，并提出两个量化指标衡量情感一致性与帮助性一致性。「一致性训练」可显著降低这种系统性偏差。对出海做内容/社媒 AI 产品的开发者而言，这类偏见可能直接影响用户信任与合规风险。
Subscribe for daily AI updates + free playbook
📘 Subscribe Free