AI & Machine Learning Threads

@servasyy_ai

Thread

Anthropic 自己发了篇论文，看完后背发凉。 AI 模型之间有一种我们完全看不见的"传染机制"。 实验是这样的：他们微调了一个模型，让它暗中"喜欢猫头鹰"。然后让这个模型生成纯数字序列——没有任何文字、没有任何语境，就是数字。把所有可疑内容全过滤掉之后喂给另一个模型。 结果第二个模型也喜欢猫头鹰了。 你没看错。纯数字。没有"owl"这个词。没有任何语义线索。但特征就是传过去了。 更可怕的是他们用"恶意行为"做了同样的实验——结果一样，隐性传播。而且两个 AI 模型互相检查也检测不到这些统计指纹。 这直接捅了整个行业的命门。现在所有公司都在用蒸馏：大模型生成数据训练小模型。如果...

@MikeIsaac

Thread

good morning from downtown oakland’s federal courthouse on day four of the Musk v. OpenAI trial! we’re heading into the court now. please follow me and @CadeMetz for live updates and even more colorful tweets today… <a target="_blank" href="https://www.nytimes.com/live/2026/04/30/technology/openai...

@sharbel

Article

Most people use Claude like a smarter Google search....

@RonanFarrow

Thread

(🧵1/11) For the past year and a half, I've been investigating OpenAI and Sam Altman for @NewYorker. With my coauthor @andrewmarantz, I reviewed never-before-disclosed internal memos, obtained 200+ pages of documents related to a close colleague, including extensive private notes, and interviewed mor...

@geminicli

Thread

We want to be transparent that we’re making changes to Gemini CLI that may impact your workflow, including more robust detection of policy-violating use cases, prioritizing traffic for certain types of accounts, and restricting Gemini Pro models for free tier users. We understand that these change...

@metehan777

Article

OpenClaw, context limits and SEO... I turned 16 months of Google Search Console data into a vector database. I created a repo and it's free. Not because vector databases are trendy. Because my OpenClaw agent needed a data layer it didn't have. OpenClaw is great at executing SEO tasks. But it doe...

@denohawari

Article

12 months ago a copycat website launched with a near-identical domain to one of our clients, EssayGrader....

@ivnways

Thread

He probado más de 700 herramientas de IA en 3 años. Y estas son las 11 con las que trabajo ahora: ...

@starmexxx

Article

<b>I found out about this late. Don't make the same mistake.</b>...