AI Pulse · 2026 年 4 月 21 日

📰 今日焦点

StackAdapt 获准在 ChatGPT 内投放广告

StackAdapt 获得 OpenAI 授权，在 ChatGPT 内投放广告。这是小范围试点，不面向所有广告主。

试点广告按 CPM 计费，起拍价为 15 美元，最高 60 美元。单次试点最低投入 5 万美元。

广告嵌入用户提问后的搜索流，不以弹窗形式出现。StackAdapt 将其称为“发现层”接入。

《OpenAI x StackAdapt 有限试点计划》推介材料于 3 月 27 日发送给部分广告采购方。

阅读全文 →

🔥 信号雷达

来自 𝕏 的实时信号，经 AI 聚类分析

AI安全防线全押错了地方？

AI安全AI agent

Google DeepMind发布25页论文，指出当前AI agent威胁模型误判了攻击面——真正风险在互联网本身，而非模型。

6.7K288283

@aakashgupta Google DeepMind just published a 25-page paper arguing the entire AI agent threat model is pointing at the wrong target. Everyone is securing the model. Jailbreak defenses. Prompt injection filters. A

开源路由工具，2毫秒内挑AI模型

AI代理开源工具

MANIFEST是开源路由器，能在2毫秒内为每个请求选择合适AI模型，简单任务用便宜模型，复杂任务用强模型，建AI代理成本最多降70%。

4.9万10134105

@RoundtableSpace MANIFEST IS AN OPEN SOURCE ROUTER THAT PICKS THE RIGHT MODEL FOR EACH REQUEST IN UNDER 2MS. Cheap models for easy tasks, stronger models for harder ones, and up to 70% lower cost for anyone building A

一个插件干掉三个浏览器扩展

浏览器扩展vibe coding

作者用vibe coding方式在18分钟内整合出名为SuperLevels的浏览器扩展，包含标签页自动清理和Cookie批量删除功能。

8.1万9381183

@levelsio Okay it's been 18 minutes and I've now vibe coded and replaced 3 more extensions and made it all into one super extension called 🚀 SuperLevels It has a 🗑️ Tab Cleaner, auto closes tabs after inactiv

一个开源中文模型，刚发布就干掉了Claude Opus

AI开源编程

Kimi K2.6在SWE-Bench Pro编码测试中拿到58.6分，超过Claude Opus 4.6的53.4分。它不是闭源大厂的旗舰模型，而是Moonshot AI发布的开源版本。

11.9万1071.9K364

@bridgemindai Kimi K2.6 just dropped. And it crushed Claude Opus 4.6 on SWE-Bench Pro. Kimi K2.6: 58.6 GPT-5.4 xhigh: 57.7 Gemini 3.1 Pro: 54.2 Claude Opus 4.6: 53.4 An open source Chinese model is now #1 on agenti @kimmonismus Holy smokes, "state-of-the-art results on several coding and tool-use benchmarks" Open Source! Kimi is cooking! -54% HLE w tools -58.6% SWE Bench Pro -66.7% Terminal Bench tl;dr Moonshot AI released K

连AI自己都难搞出真正随机的结果

AI编程内容创作

LLM生成多样输出一直很难，有人发现一个简单提示技巧：让模型先生成再操作一个随机数序列，就能稳定提升输出差异性。视频剪辑师试过之后说‘这下能批量做不重样的B-roll了’。

3.5万8917532

@hardmaru Getting LLMs to simulate “true” randomness or generate diverse outputs is surprisingly difficult. We found a simple prompting trick that solves this by having the model generate and manipulate a rando @nijokestu I just tested the AI tool that has video editors, filmmakers and content creators completely shook… and honestly? The hype is 100% real. 👀 Dreamina Seedance 2.0 just launched globally and I couldn't

Nvidia可能被自己的护城河困住

AI商业硬件

有观点指出，如果智能本身变成可复制的开源商品，Nvidia将失去对AI算力生态的议价权——它卖的不再是不可替代的芯片，而是通用计算管道的一部分。

9.7万31375191

@Rewkang A big insight from this podcast is that intelligence will be commoditized and open source. It is an existential risk to Nvidia for intelligence to be closed source as it would make them captive to a f @gitlawb Don't let Dario control AI's future. Open source must win! Contribute to OpenClaude: open coding agent for any model: https://t.co/R7xasEsXk0

Kimii 2.6 Code让终端写代码像用Claude一样顺手

TOOL编程开源

Kimi 2.6 Code是专为Kimi K2.6设计的终端编码代理，提供类似Claude Code的交互体验。它不依赖云端API，本地运行就能调用工具链完成完整开发任务。

3.9万27488294

@skirano Introducing Kimi 2.6 Code. A Claude Code-like terminal experience built specifically for Kimi K2.6, effectively making it one of the most powerful open-source coding agents on the planet. Simply bring