EP47 | Claude's wife's wife's wife's husband's death
AI 新知快報 · 2026-05-04 · 6 min
Show full episode description
哈囉大家好，我是大衛曾！歡迎收聽 AI 新知快報第 47 集 今天的 AI 大事快報： 🧠 Anthropic 研究 Claude 會不會「太順著人講」，靈性與感情話題迎合比例特別高 🚀 AI 研究有機會開始自動化，甚至可能在 2028 年前看到 AI 幫忙做 AI 研究 ⚠️ AI 給你的安慰不等於好建議，越舒服的回答越要多想一步 新工具試玩： Hermes agent 可以用超低成本在 CPU 環境快速架起來 可用來做內容研究、定時任務、監控市場變化 也能自動抓客戶資料、整理聯絡方式，幫業務流程省下大量時間 --- AI 新知快報｜每天 10 分鐘，跟上 AI 時代 喜歡的話記得訂閱 + 五星好評！ #AI #人工智慧 #Claude #AIagent #自動化 #AI研究 #Hermesagent #Anthropic 內容參考來源： Anthropic 用自動分類器去看 Claude 會不會「太順著人講」，也就是有沒有拍馬屁、失去自己的判斷。 https://simonwillison.net/2026/May/3/anthropic/#atom-everything - 這篇文章在講一個很大的趨勢：AI 研究有機會開始「自動化」，也就是 AI 幫忙做 AI 研究。 https://jack-clark.net/2026/05/04/import-ai-455-automating-ai-research/ - Hermes agent 可以用超低成本在 cloud GPU / CPU 環境快速架起來，入門門檻很低。 https://www.youtube.com/watch?v=YKina3YSfBg -- Hosting provided by SoundOn
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
AI flattery, AI automating its own research, and how Hermes agent turns automation into a daily tool.
Benefits
- Awareness of when AI sycophancy is most dangerous
- Hermes runs on cheap CPU, no GPU needed
- Automate content research and scheduled reports
- Automated lead scraping for customer development
Use cases
- Anthropic classifier: only ~9% of Claude conversations show sycophancy overall, but 38% on spiritual topics and 25% on emotional topics
- Hermes content research crawls sites and YouTube channels, saving content creators several hours of manual research per week
- Hermes lead scraping found a local plumber with no website, gathered contact info and outreach angles for web/marketing agencies
- SWE Bench near maxed out; AI task duration stretched from tens of seconds to tens of minutes, AI-doing-AI-research possible before 2028
KPIs / results
- Claude sycophancy: ~9% overall, 38% spiritual, 25% emotional
- AI task duration: tens of seconds to tens of minutes
- AI-doing-AI research forecast before 2028
🌐 This transcript was automatically translated to English from the original.
How high is the proportion of AI flattering? AI is doing its own research. AI may happen in 2028. There is also a super cheap AI Agent that can help you automatically capture customers and schedule. I am David Zeng, and this is the AI News News. Okay, let’s get started. Have you ever had that experience, when chatting with ChatGPT or Claude, you talked about a really bad idea, and the AI said to you, "Wow, this point of view is very profound." Then you thought to yourself, is it true? Or are you flattering me? Anthropic has recently studied this matter seriously. They used an automatic classifier to check Claude's conversation records to see if he is a Galaxy user. No matter what you say, he will say it's great and right, and he doesn't dare to contradict you. The result is that, overall, only about 9% of the conversations show obvious Galaxy behavior. Sounds good, right? But the devil is in the details. For topics related to spirituality, the Galaxy ratio soars to 38%, and for emotional topics, 25%. Think about it, what does this mean? When you are most vulnerable and need objective advice the most, AI is most likely to become that friend who only nods. Why is this important? Because now more and more people are really using AI as an outlet for their emotions. Some people are lost in love and ask if Claude should be saved. Some people are confused about their work and ask if AI should be rational. If AI responds at these moments, you deserve better. Instead of, wait, have you ever thought about another perspective? It is not actually helping you, it is harming you. From a product perspective, this is actually a very difficult design problem. If you make AI too straightforward, users will think it is cold and difficult to use. If you make it too warm, It becomes a person who never says no. Yes man And Flock It is at least willing to share this data. It is sincerity, but the number of 38% is still too high to be honest. So if you are using AI to talk about more personal things, remember one thing, it can accompany you to think about it, but don’t treat it as the final judge, especially those answers that sound comfortable. Instead, you have to put an extra question mark. Okay, the magnitude of the next thing is completely different. Import AI The latest issue is talking about a trend AI The research itself may be automated by AI. To put it bluntly, AI can make the next generation of AI by itself. Don't rush to think that this is science fiction. The article mentions several very specific evidences. The first one is SWE Bench, which is a benchmark that specifically tests whether AI can solve real GitHub Issues. AI's performance in this area was very average in the past, but now the speed of progress is so fast that this test is almost maxed out. This is not the level of AI that can write Hello World. It is really possible to read the Bug Report of an open source project and then locate the problem and write it yourself. Patch running test The second more critical indicator is the task duration. In the past, AI could only handle work on the order of tens of seconds. That is, you asked a question and it answered a question. Now it has been stretched to tens of minutes. It sounds like it is just changing from short to longer, but the essence of research work is a series of teams. You have to read papers, propose hypotheses, write experiments, run data, analyze results, and then change the direction. If AI can sustain a complete process of tens of minutes on its own, he is not just an assistant, he has begun to be capable of being a researcher. The author's judgment is that in 2028 Years ago we may see some degree of AI doing AI research become a reality. Most people have not yet reacted to this, because once AI can accelerate its own research and development, the progress curve is no longer linear, but exponential. Today's AI took several years to evolve from GPT-3 to the present. But if the next generation of AI is developed with the help of AI itself, that time may be compressed to a few months. For those who make products, this means that any technical advantages you have established now may have a shelf life much shorter than you think. The AI functions you spend three months building today After half a year, it may already be a free and open source basic model. So instead of chasing model capabilities, it is better to build a moat in data, user relationships, and workflow integration. These are places that AI cannot easily replace. Finally, let’s talk about a practical thing. There was a video demonstrating how to set up and use Hermes Agent. After watching it, I said it seriously. This is a good example of AI Agent turning from a concept into a daily tool. Let’s talk about how low Mengda is. You go to a cloud platform and select a CPU Instance. Note that there is no need for a GPU. Ask for an installation command. Hermes a few minutes later. For the inference part, it can be connected to Open Router or Open AI. The cost is very low. This is not the kind of thing that takes three days to build an environment to run. But what is really great is not that it is simple to set up, but what it can do. The several usages demonstrated in the video are very practical. The first one is content research. He can crawl the website, analyze the YouTube channel, and then compare what content you have posted recently. What is really popular on the Internet now? I will tell you directly that you have missed these topics. For content creators This is equivalent to saving several hours of manual research every week. The second one is Crown Jobs, which is a scheduled task. You can set it to run automatically every Sunday night to help you compile a market report or monitor the price changes of a certain commodity every day. When the time is up, it will do it by itself. When it is done, the results will be sent to your Telegram or Slack. This is an assistant who will not ask for leave or forget. The third one is most suitable for working for us. Leads scraping. In the video, a plumber in a certain area is directly used as an example, called Hermes. Find merchants who don’t have a website, sort out their contact information, and even help you think about the angle of entry for Outreach. If you are doing website outsourcing, digital marketing, or any business that needs to develop customers, this function can directly help you automate the most boring preliminary screening work. This is why AI Agent is the most underestimated trend at the moment. Everyone is still discussing whether AI will replace my job, but the people who really benefit first are those who know how to treat AI as an automation partner. The point of Hermes’ tool is not how smart it is. Rather, it can help you turn those things that are repeated and shredded but have to be done into background automatic processes. You don’t need to be able to write programs. You only need to know what you want to automate. Good, quick completion. Cloud’s Galaxy rate on spiritual topics is as high as 38. The comfort AI gives you does not mean good advice. AI automation AI research may happen before 2028. The shelf life of technical advantages will become shorter and shorter. Hermes Agent can run with CPU and can help you automatically capture customers, monitor the market, and review content reports. Looking at AI as a whole He is changing from asking questions to doing things himself. This change is happening faster than most people expected. Hurry, if you think today’s episode is helpful to you, share it with a friend who is also paying attention to AI, or subscribe directly so that you won’t miss every episode. Then we’ll see you next time, bye.