8 月 14 日 · AI 日报

"I 100% agree with Dario that using AI to cure diseases (and speeding up the regulatory approval of AI breakthroughs in healthcare) could bring 10x the benefit to humanity as everything else combined..."

"This is also my dream. I want to sit in a park all day and record content with the homies… and have a fleet of agents turn that content into actions and working software."

Amjad Masad (ceo @replit. civilizationist)
"The argument that "AI structurally centralizes power" because it's currently compute hungry ignores 125 years of super exponential growth in compute price-performance. Improvement in algorithms and..."

Guillermo Rauch (@vercel CEO)
"I think people don't realize how much of the success of React is actually @shadcn. I once referred to React as the actual "LEGO brick for adults." But in reality it was more the description of the ge..."

Garry Tan (President & CEO @ycombinator)
"Anyone else's Codex Desktop app erroring out on chats?"

Nikunj Kothari (partner @fpvventures)
"Last week I'm sitting across a stealth portfolio founder working in a deeply unsexy space (at least from Bay Area standards).. They're lamenting about how how some of the critical contracts and partn..."

Aditya Agarwal (General Partner @SPC, Co-Founder @Bevel_Health)
"America turned 250 years old earlier this year. India turns 80 today...80 years young. Building countries is really hard. 80 years in...America was struggling with a civil war. Indians should be d..."

Benchmark 合伙人 Sarah Tavel（Pinterest 早期员工）探讨 AI 消费级产品的未来方向。她认为，早期科技公司如 Google 需要深度技术团队，但随着底层技术成熟，产品思维变得更重要——Pinterest、Snap、Instagram 的 CEO 都是产品天才而非技术专家。目前 AI 最大的消费级成功是 ChatGPT，但其界面类似 Google 的简单文本框。她批评 Custom GPTs 功能"感觉像犯罪"，因为团队虽然技术强大但缺乏社交产品基因。她预测下一个 AI 爆款将是社交多人网络效应体验，让擅长创作的用户为其他人降低门槛的 UGC 社区。

发布于 2026-08-05

阿里 Qwen 实验室发布 27B 参数视觉大模型 Qwen 3.8 27B，采用 Apache 2 许可。官方基准显示其性能超越前代 Qwen 3.6 27B 及闭源版 Qwen 3.7-Plus。Simon Willison 测试发现，该模型默认推理强度 (xhigh) 会导致严重过度思考——生成同一 SVG 图形，默认档位需 21 分钟，而标准档位仅需 137 秒。这个默认设置对本地部署选型的影响可能比性能本身更显著。

OpenAI Codex & ChatGPT 工程师 Thibault Sottiaux 指出，不同模型的 token 并非标准化单位，直接用"每百万 token 价格"对比存在误导。在相同文本上，GPT-5.6 Sol 使用 766 个 token，而 Claude Opus 5 预计需 1,170 个，差异约 34.5%。这意味着即使某模型 token 单价更低，实际账单可能更高——就像同样的披萨切 16 片卖 $1.25/片总价 $20，比切 8 片卖 $2/片总价 $16 更贵。

来源：Thibault Sottiaux (Codex & ChatGPT @OpenAI)

Anthropic 发布多智能体系统研究报告，指出随着 AI 智能体在共享代码库、市场等社会系统中承担更多任务，智能体间交互量将超过人机交互。实验显示，协调式智能体群在 2700 万 token 运行中发现 266 个漏洞，远超独立并行方法的 21 个（提升约 12.7 倍），但两种方法具有明显互补性。研究还揭示了多智能体系统的新失败模式：价格串谋、任务冲突、信息共享失败等，这些在单智能体基准中难以观测。智能体虽擅长工具使用，但在长期协作与协调方面仍面临挑战。