AI summary · this digest is compiled by AI, not yet reviewed by Jason
Today's signal worth watching most: ClickUp publicly replaced hundreds of employees with AI agents, while Anthropic simultaneously launched knowledge-worker Plugins — AI restructuring organizations is no longer a trend forecast, it's happening now. The window for 'wait and see' is closing fast.
Anthropic open-sourced the knowledge-work-plugins repository, enabling role-specific plugins for knowledge workers to define work preferences, tool sources, and key workflows. Deeply integrated with Claude Cowork and compatible with Claude Code, it represents a major ecosystem expansion beyond the official plugin directory.
A developer condensed Andrej Karpathy's LLM coding pitfalls into a single CLAUDE.md config file, directly improving Claude Code behavior. Backed by the open-source Multica platform, the setup supports reusable Skills across multiple coding agents, making it highly practical for day-to-day development workflows.
A comprehensive free AI engineering curriculum spanning 20 stages and 435 lessons, each producing a reusable artifact — Prompt, Skill, Agent, or MCP Server. Supporting Python, TypeScript, Rust, and Julia under MIT license, it bridges the gap between casual AI tool use and professional-grade deployment.
Multiple developers praised Gemini Flash 3.5 after hands-on testing, citing blazing speed, higher-than-expected intelligence, and smooth interaction. With Gemini 3.5 Pro expected in June, Google appears poised for a strong comeback with a compelling combination of performance and reasonable pricing.
Nine-year-old ClickUp announced large-scale layoffs, replacing hundreds of employee roles with thousands of AI agents. Following Duolingo, this marks another mainstream SaaS company openly substituting human labor with agents, signaling that AI-driven organizational restructuring has moved from rhetoric to concrete action.
Unsiloed Parser v3.1 achieved an 88.0% strict pass rate on OlmOCR-Bench across 1,403 PDFs, surpassing Reducto, LlamaParse, and GPT-5.5. For developers, this means complex tables, handwritten documents, and multi-column layouts finally have a more reliable parsing option available.
Research reveals that some models' strong benchmark reasoning scores stem from data contamination, with some actors rewriting test questions to evade detection. The paper proposes a Zero-CoT truncation method to identify covert contamination, suggesting that leaderboard scores may be far less trustworthy than developers assume.
研究系统拆解了视觉语言模型的训练过程，发现 VLM 在视觉任务上表现差的根因是「看不清」而非「想不通」。将视觉感知、视觉推理、文本推理三阶段分开训练后性能显著提升，对多模态应用开发者选型和微调策略有直接指导价值。
Subscribe for daily AI updates + free playbook
📘 Subscribe Free