X’s algorithm has changed. Time to start publishing high-quality articles, everyone!
Pay-for-results sounds great until someone has to write the attribution logic
Did the ring move come from Claude or from the chemist who asked the right question
My guess: the dispute queue ships bigger than the billing system
Dario Amodei: some tokens Claude produces are worth a few cents, others could be worth tens of millions of dollars.
His examples:
1. Telling someone to restart their Mac: the model has said it 10 million times, worth a dollar or less
2. Telling a pharma company to move an aromatic ring to the other end of a molecule: potentially tens of millions
3. Expect pay-for-results pricing, or labor-like pricing by the hour
4. It's a new industry, so many models will be tried
Today pricing is simpler: per the S-1, usage-based pricing was 80% of H1 2026 revenue and subscriptions 20%.
Full breakdown here:
theaiopportunities.com/p/ant…
NVIDIA 给 Meta 投 30 亿美元,一起调 Muse
还说要在 OpenShell 参考架构和 Sentry 监控上继续合作
芯片、推理栈、沙箱、监控,眼看要全是一家的
监控跟着供应商走,换卡那天告警规则也得重写
被监控的和做监控的签同一份合同,这不叫可观测,叫自证
蚂蚁开源 5M 参数安全探针
读模型自己的隐藏状态,逐 token 打分
有点像让服务自己写自己的审计日志
模型被带偏,探针读的也是那份被带偏的脑子
而且它管的是说了什么
agent 出事大概率在调了什么工具
护栏装嘴上还是手上?
huggingface.co/inclusionAI/Q…
百度开源 Unlimited-OCR 2,今天刚发
几百页的 PDF 不用再先切块了
它把跨页的注意力压成固定一组参照,比如 512 个 token
文档越长,每步延迟和显存几乎不涨
像解析 XML,从 DOM 换成 SAX
以前长文档 OCR 是先切页、再拼接、再修表格跨页
现在切块那层胶水代码可能能删了
你的 RAG 管线里,切块逻辑写了多少行?
github.com/baidu/Unlimited-O…
猫王起诉 Anthropic
猫王 274 首歌,每首 15 万美元,索赔约 4100 万
代码依赖有 SBOM,一扫就知道哪个包是 GPL
训练数据大概率没这张表,哪首歌进了哪版权重说不清
依赖还能 revert,权重里的歌只能重训
开源 agent 同时接管浏览器、电脑和手机,
恭喜GitHub 破 5000 星
界面交互打磨精致,多端都支持
Cindy just crossed 5,000 ⭐ on GitHub! Thank you to everyone building with us. Open-source AI agent that drives your browser, computer and phone. Free 👉 github.com/makecindy/cindy
自定义皮肤,缺个分享皮肤配置的。
Reminder you can customize the appearance of chatgpt on web and make it soothing
chatgpt.com/settings/appeara…
Claude 写 90% 代码,回退的 PR 还少了四成
先别急着鼓掌
回退率只算「被发现、还有人愿意回滚」的那部分
就像告警变少,可能是服务稳了,也可能是阈值被调高了
PR 数翻倍,评审的人没翻倍,大概率每个 PR 看得更快了
我更想看上线三个月后才冒出来的 bug 降没降
评审到底是在把关,还是在盖章?
teapawt.net/addyosmani/status/2114…
公司级 Agent 最难的是「这步算谁干的」
LangChain 的做法:私聊让它开 PR 用你的身份,后台 review 用它自己的
代人发信的工具只在私聊开,公共频道关掉
不然就是全组共用 root 的老毛病,出事日志里只有一个名字
你的 Agent 拿的是谁的钥匙?
langchain.com/blog/component…















