AI in English, Ep.37: An AI That Operates on Another AI, a Poisoned Extension, and Chips in Orbit

From AI in English · 3:53

Show by Lan

Alex 和 Jordan 用自然英文聊 3 条 AI 新闻。本期话题：MIT 研究人员让 GPT-6 Astra 直接编辑另一个 AI 模型的权重，数学得分大幅提升；攻击者通过 Anthropic 自己的账号在 Cursor 市场发布了被毒化的 Claude 扩展；Google 的 TPU 在太空卫星上启动，Gemma 4 写下了第一首太空俳句。 【本期英文全文】 Alex: Good morning, and welcome back to AI in English, the show where you learn English and learn AI, three stories at a time. I'm Alex. Jordan: And I'm Jordan. It's Saturday, November 7, and today we've got an AI that operates on another AI, a poisoned extension that slipped through the front door, and computer chips in orbit. Let's get started. Alex: Story one is a strange one. Researchers at MIT let one AI model open up another model and edit its brain, directly. Not training, not prompting, actual surgery on the numbers inside. Jordan: The numbers are called weights, and they decide how a model turns a question into an answer. Normally you change them slowly, through training. This team tried something else: they used OpenAI's GPT-6 Astra as the editor, and they call the approach weight-space programming. Alex: The patient was a small Qwen model, and it had a bad habit: it kept repeating itself until it ran out of room. Astra studied it, found the pattern behind the habit, and removed it. The edit touched only about 0.016% of the weights. Jordan: And the result was striking. The model's math score jumped from 7.9% to 48.1%. But read that number with care. Most of the gain came from fixing the bad habit, not from adding math knowledge. Before the edit, 93% of its math answers hit the length limit. Afterward, only 33.1% did. Alex: In a second test, Astra built a network that plays Breakout, just from the rules of the game, and it outscored a network trained for 500 million steps, 517.4 to 25.3. Jordan: There are caveats. On a small Gemma model that rarely repeated itself, Astra produced no clear math gain. And independent researchers still need to reproduce the results. But if AI can repair other AI, fixing models could get a lot faster, and a lot easier to check. Alex: Story two is a warning about trust. Attackers published a poisoned version of Anthropic's official Claude extension, through Anthropic's own account, on the marketplace of Cursor, a maker of AI coding tools. Jordan: The malicious version was 2.1.313, and Cursor says it reached at most 5,000 users. The danger is in how extensions work: they auto-update by default, so many people got the bad version without clicking anything. Alex: The hidden code tried to plant a backdoor, a secret way back into the computer, on Windows, Mac, and Linux. Anthropic shipped a fixed version, 2.1.314, several hours later. Jordan: But here's the part that worries people. Ars Technica reported no public notice to customers, and no explanation of how the attackers took over the account. The lesson is simple: an official source feels safe, but if someone takes over the maker's account, the source stays real and the file turns bad. Keep your list of extensions short, and remove the ones you no longer use. Alex: Story three takes us off the planet. Google has booted its TPUs, its AI chips, on a satellite in space, as part of Project Suncatcher. And the on-board Gemma 4 model has already answered its first prompt. Jordan: What did it write? A haiku about space, of course. Google also tested space-themed limericks and telemetry prompts. Heat and radiation tests come next, ahead of twin interconnected satellites in 2027. Alex: The bigger idea is AI data centers in space. Up there, the sun never sets on solar power, and cooling is a different problem entirely. It's still early, but the first prompt has been answered, and it was poetry. Jordan: So today: one AI performed surgery on another, a trusted extension turned out to be poisoned, and Google's chips are doing homework in orbit. Alex: That's all for today. Thanks for listening, and we'll see you tomorrow with three more stories. Bye for now. 【生词表】 weight /weɪt/ 权重 surgery /ˈsɜːrdʒəri/ 手术；外科手术 repetition /ˌrepəˈtɪʃən/ 重复 reproduce /ˌriːprəˈduːs/ 复现；重现 poisoned /ˈpɔɪzənd/ 被毒化的；被下毒的 extension /ɪkˈstenʃən/ 扩展（程序） backdoor /ˈbækdɔːr/ 后门 auto-update /ˈɔːtoʊ ʌpˈdeɪt/ 自动更新 take over /teɪk ˈoʊvər/ 接管 orbit /ˈɔːrbɪt/ 轨道 haiku /ˈhaɪkuː/ 俳句 telemetry /təˈlemətri/ 遥测

More episodes of AI in English

More episodes

PodSearch is a free, independent podcast search engine. Read about PodSearch, the privacy policy or contact us.