AI in English, Ep.31: A Haiku from Space, a Trillion-Parameter Model, and Tighter Teen Rules

From AI in English · 4:22

Show by Lan

Alex 和 Jordan 用自然英文聊 3 条 AI 新闻。本期话题：谷歌 Project Suncatcher 在轨卫星上跑出第一条完整 AI 回复（一首俳句）；Mistral 发布 1 万亿参数的 Large 4 模型权重；OpenAI 收紧 ChatGPT 青少年账号限制，暴力话题下可主动结束对话。 【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 Sunday, November first, and today's stories go from orbit to the open-weight world to the rules for teenagers. Let's get started. Alex: Story one literally comes from space. Google says it has run an AI model on a satellite in orbit and received its first complete response back on Earth. Jordan: The project is called Project Suncatcher, and Google's chief executive, Sundar Pichai, called it a "hello world" milestone. The team sent a command up to the satellite, booted up its Tensor Processing Unit, or TPU, ran a prompt, and got an answer back. Alex: The prompt was simple: write a haiku about space. The model on board was Gemma 4, Google's open model, and the answer came down during a pass over Svalbard, Norway. Jordan: The satellite itself launched on October first. Since then, Google says it has finished its early checks, powered up its host and TPU trays, and run early machine learning tests. It has also tried space-themed limericks, and prompts asking for the satellite's own technical status. Alex: So why put AI in space at all? Google's long-term idea is to test whether space can support large-scale machine learning. Satellites in low Earth orbit get near-constant sunlight, and Google says they could generate up to eight times more solar power than systems on Earth. Jordan: The next tests are less poetic and more practical: how the TPUs handle heat and radiation in orbit. If they hold up, Google plans two interconnected satellites in twenty twenty-seven, a first step toward constellations that could share larger workloads. Alex: Story two stays on Earth, but it is also about scale. Mistral has released the weights for Large 4, a multimodal model with one trillion parameters. Jordan: Releasing the weights means other organisations can download the model and run it on their own infrastructure, rather than only using it through Mistral's service. The model offers three reasoning modes, so users can choose how much thinking time to give it. Alex: In Mistral's own evaluations, its coding agent benchmark scored forty-eight percent, comparable to GLM 5.3, while using forty-seven percent fewer tokens per task on average. That is a claim about efficiency, not just size: similar results, less work per task. Jordan: There is a licence catch worth reading. The release uses a modified MIT licence, and businesses that offer the model as a service need a separate commercial licence. So "open weights" does not mean "no strings attached." Alex: Still, for teams that want control over where a very large model runs, this is a serious new option. A trillion-parameter model you can host yourself used to sound like a fantasy. Now it is a download. Jordan: Story three is about rules, and it affects teenagers first. On Friday, OpenAI announced tighter limits for teen accounts in ChatGPT. Alex: Teen accounts will no longer get practical information about weapons and security systems that could be misused, even when the question sounds harmless, like homework or curiosity. The account's age band decides, not the wording of the question. Jordan: The changes go further. For every user, ChatGPT will be able to end a conversation on its own when someone seems to be fixating on violence, and OpenAI may temporarily pause an account after several such chats. Details about past mass attacks will also be limited for everyone, including which weapons were used and how they were obtained. Alex: OpenAI says it has also strengthened the system that estimates a user's age, so more users it believes are under eighteen will land in the teen experience. And it plans to tell parents when a teen account is paused for repeatedly discussing violence. Jordan: The rollout starts over the next few weeks, so families will not see all of it this weekend. OpenAI also promised a first Community Safety Report in early twenty twenty-seven, covering the second half of twenty twenty-six, with updates every six months after that. Alex: So today: a haiku from orbit, a trillion-parameter model you can run yourself, and a chatbot that is learning when to stop a conversation. Three very different stories, one shared question: how far should AI go, and who gets to set the limit? Jordan: That's all for today. Thanks for listening, and we'll see you tomorrow with three more stories. Bye for now. 【生词表】 milestone /ˈmaɪl.stoʊn/ 里程碑，重要进展 orbit /ˈɔːr.bɪt/ 轨道；在轨道上运行 constellation /ˌkɑːn.stəˈleɪ.ʃən/ 卫星群；星座 multimodal /ˌmʌl.tiˈmoʊ.dəl/ 多模态的 parameter /pəˈræm.ə.tɚ/ 参数 infrastructure /ˈɪn.frəˌstrʌk.tʃɚ/ 基础设施 benchmark /ˈbentʃ.mɑːrk/ 基准测试；衡量标准 licence /ˈlaɪ.səns/ 许可证 fixate /ˈfɪk.seɪt/ 执着于，念念不忘 rollout /ˈroʊl.aʊt/ 推广上线，逐步推出

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.