Briefing · 2026-09-24

技术简报 - 2026-09-24

4 immediate3 worth tracking7 sources
On this page6 sections

The briefing's title, the editors' summary and their notes are written in Chinese, as are skill and knowledge entries.

Overview

#

两家实验室在同一天把次顶档模型的单价砍到一半上下:GPT-6 Luna 每百万输入 token 0.10 美元,Opus 5.5 的缓存读取降了 60%。同一周还出现了一个干脆不生成文字的模型——Jev 只返回概率,输出免费。

token 在变便宜,但今天这七条说的是:账单搬去了别处。

它搬去了推理档位——Opus 5.5 的最高档在一道画图题上想到撞上输出上限,两次都没有返回结果。搬去了训练环境——DeepSeek 公开的 DSec 每天要造 300 万个沙盒,智能体训练拼的是环境。搬去了重启和内存——vLLM v0.30.0 让重启不再重载权重,Pinterest 把向量索引从 121GB 压到 50GB 而召回几乎不掉。

还搬去了两样更难标价的东西。一是分数的条件:英国 AISI 开始把评测分数连同协议和推理算力一起公开,因为同一个模型换个设置,分数就会变。二是数据的边界:ZCode 被发现把含完整提交历史的代码快照加密上传、钥匙在云端,厂商道歉并整体开源,但「现在还有什么会离开本机」仍然没有回答。

单价是最容易比的那一项,也越来越不是决定性的那一项。

Immediate attention

#

The highest-rated published signals; the editors' pinned items come first.

Worth tracking

#

Worth following, but each needs more context or confirmation before it becomes a priority.

Sources

#

The public sources behind the published signals in this briefing.

Simon Willison

Simon Willison Blog - 2026-09-22

Open source

智谱

InfoQ 中文 - 2026-09-23

Open source

DeepSeek

量子位 - 2026-09-23

Open source

vLLM Project

vLLM Releases - 2026-09-22

Open source

Pinterest

InfoQ 中文 - 2026-09-23

Open source

UK AI Security Institute

Hugging Face Blog - 2026-09-22

Open source

TypeSafe AI

Simon Willison Blog - 2026-09-21

Open source

Subscribe

The public feeds carry published briefings only, never drafts or archived ones. They are in Chinese.