返回首页

一周热榜

1作者: oliveskin7 天前原帖
嘿,HN, 我开发了Tinman,因为在生产环境中发现大型语言模型(LLM)的失败实在太麻烦了。传统的测试只检查你已经想到的内容,而Tinman则试图发现你未曾想到的问题。 它是一个自主研究代理,能够: - 生成关于潜在失败模式的假设 - 设计并运行实验来测试这些假设 - 对失败进行分类(推理错误、工具使用、上下文问题等) - 提出干预措施并通过模拟进行验证 核心循环持续运行,每个周期都会为下一个周期提供信息。 为什么是现在:随着像OpenClaw/ClawdBot这样的工具为代理提供真实系统访问,失败的表面远比“糟糕的聊天机器人响应”要大得多。Tinman拥有一个网关适配器,可以连接到OpenClaw的WebSocket流,以便在请求流动时进行实时分析。 三种模式: - LAB:对开发环境进行无限制研究 - SHADOW:观察生产环境,标记问题 - PRODUCTION:需要人工批准 技术: - Python,全面支持异步 - 可扩展的GatewayAdapter ABC,适用于任何代理/网关 - 内存图,用于跟踪何时知道了什么 - 兼容OpenAI、Anthropic、Ollama、Groq、OpenRouter、Together ``` pip install AgentTinman tinman init && tinman tui ``` GitHub: [https://github.com/oliveskin/Agent-Tinman](https://github.com/oliveskin/Agent-Tinman) 文档: [https://oliveskin.github.io/Agent-Tinman/](https://oliveskin.github.io/Agent-Tinman/) OpenClaw适配器: [https://github.com/oliveskin/tinman-openclaw-eval](https://github.com/oliveskin/tinman-openclaw-eval) Apache 2.0。没有遥测,没有付费层。欢迎反馈和贡献。
1作者: eismcc1 天前原帖
你好,HN!<p>我有一些空闲时间,所以我一直在升级 KlongPy。<p>KlongPy 现在支持 PyTorch 作为后端,并在数组语言上实现了自动求导。<p>之前在 HN 上有人提到希望 Klong 能支持自动求导,我也觉得这很不错。<p>在这里有很多示例:<a href="https://github.com/briangu/klongpy/tree/main/examples/autograd" rel="nofollow">https://github.com/briangu/klongpy/tree/main/examples/autograd</a><p>不过这里给你一个简单的示例:<p>f::{x^2} : " 定义 f(x) = x^2<br>f:&gt;3 : " 计算 f'(3) = 6.0<p>期待对新操作符和任何明显不足之处的反馈。谢谢!<p>更新的 GitHub 页面:<a href="https://klongpy.org" rel="nofollow">https://klongpy.org</a><br>代码库:<a href="https://github.com/briangu/klongpy" rel="nofollow">https://github.com/briangu/klongpy</a>