Alibaba's Metis agent cuts redundant AI tool calls from 98% to 2% — and gets more ...

Alibaba's HDPO framework trains AI agents to skip unnecessary tool calls, cutting redundant invocations from 98% to 2% while ...

Why OpenAI's 'goblin' problem matters — and how you can release the goblins on your own

If OpenAI can accidentally train its flagship model to obsess over goblins, what other more subtle and potentially harmful ...
来自MSN

Local LLM experiments reveal hardware, model choice matter most

Months of hands-on testing with locally run large language models (LLMs) show that raw parameter count is less important than architecture, context window, and memory bandwidth. Advances in ...

不可思议！400B大模型在iPhone上跑起来了

有时候看到一些大模型项目，总会怀疑是不是真的有外星人在干预地球科技。 就比如今天这个。 刚看到这个 Demo 的时候着实有些想笑，很久没有见过吐词如此之慢的大模型了。观感上就像「闪电」老师。 尽管只有每秒 0.6 个 tokens 的输出速率，这依旧是一个令人不可思议的工作。因为这是一个跑在 iPhone 17 Pro 上的 400B 大模型！ 准确的来说，这是在 iPhone 17 Pro 的 ...
Security Boulevard

Copy Fail (CVE-2026-31431): Frequently asked questions about Linux kernel privilege ...

A flaw in the Linux kernel present since 2017 allows a local user to gain root access on virtually every major Linux distribution. A public exploit is available and reported to work reliably.Key ...