🤖 AI 治理
20min
Sam Altman @ Stanford CS183
by Sam Altman · 2024 回炉访谈
Scale 哲学 + ChatGPT 故事 + AI 是新 Utility + 3 个未来 Fork（80% 民主化概率）。Fable 5 现象：intelligence 移到 harness。
"Even though we would be one of these companies, the risk of keeping this concentrated is not something we should tolerate."
🤖 AI 治理
12.5min
我们可能真的需要停下 AI
by Nate Herk · theMITmonk
OpenAI + Anthropic 同周呼吁放慢 AI 发展。Nate 关键洞察："我们要求暂停 ≠ 我们愿意停"。3 个未来 fork + Cloud Mythos 6/12 被 US 政府要求暂停。
"The break they're asking for isn't a break they're willing to pull. It's a break that they want somebody else to create and hold."
🤖 AI 路线
35.5min
Hybrid LLM 路由 + 决策框架
综合 8 篇文章 + 3 视频 · knowledge base
2026 Hybrid LLM 工作流综合：本地 vs Cloud 决策矩阵 + 5 维评分（context/complexity/privacy/latency/quality）+ 4 大 Router 工具（LiteLLM / OpenRouter / Inworld / RouteLLM）+ 5 维优化叠加 ~80% 成本下降。
"The goal is not replacement, it is orchestration." — Cloud.in
🛠️ 工具实战
16min
Claude Code 接管 Unreal Engine 5
by Stefan 3D AI
免费 MCP 插件让 Claude Code 直接驱动 UE 5.7。Unreal Claude + Web UE（Vibe UE）。Milestone + commit 节奏 + 截图自检循环。实战做出可玩 endless runner。
"Less money less expensive" — 4B 专用 extraction + SOTA reasoning
🛠️ 工具实战
39min
M5 Max + MLX 本地栈
by IndyDevDan
M5 Max 64GB 跑 4 个模型（Qwen 3.5/Gemma 4 × GGUF/MLX）3 套 benchmark。MLX 碾压 GGUF（2x decode）。Context 8K-16K 是分水岭。Agentic 本地能跑但限 8K。
"Local models are a work in progress. For 90% of tasks, use a model provider. But this is a space I'm keeping a sharp eye on."
🛠️ 工具实战
17.7min
本地 Agentic Coding 完整工作流
by Tech With Tim
Qwen Coder + LM Studio + VS Code 完整本地 agentic coding。VRAM 决定一切 + Q4 量化甜点 + 5 步安装（Manage Language Models + Continue 扩展）。
"Model selection is by far the most important part. Pick the lowest quantization you can run."
📄 论文
17.7min
Self-Harness Optimization
by Discover AI · Shanghai AI Lab 论文
让 LLM 自己优化自己的 harness（不改权重）。Fable 5 现象：intelligence 移到 harness。Boris 6/8 公告："stop prompting, build loops that prompt themselves"。4 步 self-harness loop + 4 元素（prompts/tools/memory/policy）。
"An LLM-based agent improves its own operating harness without any humans in the loop... boiling in your own soup."
📄 论文
35.5min
K1: Agent Native Knowledge Orchestration
by Discover AI · Shanghai AI Lab 2026-06-11
GraphRAG 5 大问题 → 5 schema 多模态 KG + 4B 专用 extraction LLM + 3 源 retrieval（不用 vector DB）+ 17 MCP tools。MIT license。性能 41.8% → 66.3%。
"Separation of concerns is one of the strongest engineering decision. 4B extraction + SOTA reasoning = less money less expensive."
📄 学习路径
15.5min
AI Engineer 学习路径：DataCamp 实地评测
by Tech With Tim
DataCamp "Associate AI Engineer" 9 门课 + 3 项目 + 26 小时实地评测。$13/月。OpenAI + HF + LangChain + RAG。Tim 亲自跑完 20+ 小时。
"AI engineering is applying things already built. You don't need math."
🧠 思维
20min
Systems Thinking 4 大系统
by Sandeep Swadia (theMITmonk)
Clear / Complicated / Complex / Chaotic + DART 诊断框架。Van Halen M&M's 协议 · Cobra 错激励 · Tylenol 1982 召回 3,100 万瓶。3 个外部视角（Mentor / Data / Time）。
"It is not difficult to be consistently better than yourself. The world will meet you at your level of audacity and hope."