Explorar

每个 AI 模型都能学习使用 Skills 吗？ @Letta_AI 发布 Context-Bench Skills 评测基准，来测试 AI 模型能否像人类一样"按需学习技能"。核心问题 AI 智能体在现实应用中不可能预先掌握所有知识。团队提出的解决方案是让智能体能够动态加载专业技能——就像你在需要时查阅操作手册，而不是把所有手册都背下来。什么是 Skills（技能）技能本质上是可挂载的知识包，包含一个描述文件（SKILL. md）和相关资源（数据集、脚本、示例等）。关键在于智能体不应该在对话开始时就加载所有技能，而应该只在需要时才挂载相关技能。比如，智能体可能有一个"公司风格指南"技能，只在需要撰写营销内容时才加载；或者有一个"人口普查数据模式"技能，只在分析人口统计信息时使用。评测方法 Context-Bench Skills 使用 Anthropic 的开源技能库，通过 LLM 为每个技能生成合适的任务。评测分三个场景： 1. 基线：智能体没有任何技能 2. 技能使用：提供所需技能的元数据，智能体需要加载技能内容 3. 技能选择与使用：智能体需要从完整技能库中找到并使用正确的技能主要发现对于擅长使用技能的 Claude 模型，提供相关技能可使任务完成率平均提高 14.1%。更有意思的是，GPT-5 和 GLM-4.6（开源权重）等非 Anthropic 模型也能获得类似的性能提升，这表明技能获取是一种通用能力而非 Claude 特有的功能。不过也有局限：从技能库中选择正确技能比使用已识别的技能更难——当模型需要先找到正确技能时，性能下降约 6.5%。而且能力较弱的模型（如 GPT-5 Mini）即使提供了技能也无法正确使用，说明存在一个能力门槛。实际意义 Letta 团队开发了 Letta Code，这是一个模型无关的工具，让任何 LLM（GPT-5、Gemini、GLM-4.6 等）都能使用最初为 Claude 设计的技能库。这意味着技能成为了一种可移植的知识单元，支持智能体的持续学习——当一个智能体开发出解决方案时，可以将其打包成技能供其他智能体使用。

专注 - Context Engineering, AI(Coding)Agents. 分享 - AI papers, apps and OSS. ex Microsoft MVP 合作 - 私信/邮箱：shaomeng@outlook.com 📢 公众号/小红书: AI 启蒙小伙伴 🔗 信息卡提示词 🔽

meng shao

Sat Nov 08 08:20:16

Funny conspiracy theory. Noam is very serious about his trad Judaism, "divine benevolence" is not just a cute joke. I don't much believe in his singular Gemini-saving genius, CharacterAI was ≈high Chinese penny-pinching lab level. But why would he need any fake drama to get out?

We're in a race. It's not USA vs China but humans and AGIs vs ape power centralization. @deepseek_ai stan #1, 2023–Deep Time «C’est la guerre.» ®1

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)

Sat Nov 08 08:12:22

Cool

🔖 Building · Sharing ✨ Zipic · TimeGo · KeygenGo · HiPixel · Orchard · Prism

十里

Sat Nov 08 08:07:03

RT @indie_maker_fox: 持续开发迭代就会持续有流量，也就持续有收入 ✌️

🚀 The best AI SaaS boilerplate - https://t.co/VyNtTs0jSX 🔥 The best directory boilerplate with AI - https://t.co/wEvJ1Dd8aR 🎉 https://t.co/zubXJCoY92 & https://t.co/tfQf8T7gGF & https://t.co/TqRkfQj41f

Fox@MkSaaS.com

Sat Nov 08 08:01:10

RT @zephyr_z9: Bro, this has to be fake bullshit

We're in a race. It's not USA vs China but humans and AGIs vs ape power centralization. @deepseek_ai stan #1, 2023–Deep Time «C’est la guerre.» ®1

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)

Sat Nov 08 07:57:33

I'm sorry, noise and all, but o1-high was so much stronger than o1-preview that "the best benchmark for common sense" ought to have picked up on that. Yes these is a way in which proprietary models have an edge. I think it's a matter of a couple good human-written datasets.

We're in a race. It's not USA vs China but humans and AGIs vs ape power centralization. @deepseek_ai stan #1, 2023–Deep Time «C’est la guerre.» ®1

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)

Sat Nov 08 07:56:45

Newest first — browse tweet threads

Explorar

Newest first — browse tweet threads

Funny conspiracy theory. Noam is very serious about his trad Judaism, "divine benevolence" is not just a cute joke. I don't much believe in his singular Gemini-saving genius, CharacterAI was ≈high Chinese penny-pinching lab level. But why would he need any fake drama to get out?

Cool

RT @indie_maker_fox: 持续开发迭代就会持续有流量，也就持续有收入 ✌️

RT @zephyr_z9: Bro, this has to be fake bullshit

I'm sorry, noise and all, but o1-high was so much stronger than o1-preview that "the best benchmark for common sense" ought to have picked up on that. Yes these is a way in which proprietary models have an edge. I think it's a matter of a couple good human-written datasets.