主播
节目简介
来源：小宇宙
Hi there~~ Good evening, everyone! Today is Thursday, September 24, 2026, and welcome to Goodnight Coffee. I’m An’an, your host on Goodnight Coffee. I’m super happy to see you all again! Tonight, let’s talk about the freshest business and tech updates, and catch new trends and new inspiration. Relax your ears, follow An’an, and let’s go!
@@@
According to Readhub, just 90 minutes after Anthropic released Opus 5.5, OpenAI launched GPT-6 Sol and GPT-6 Luna on September 23, with API prices directly cut by 50% compared with the GPT-5.6 promotional price. Sol costs $2 per million input tokens and $10 per million output tokens, while Luna costs $0.10 for input and $0.50 for output. Both models bring down core capabilities from the flagship GPT-6 Astra. They both offer a 1.05 million-token context window, a maximum output of 128K, support for multimodality and a full suite of tool calls, and knowledge cutoffs that are partly earlier than Astra’s. They are already available to different tiers of users and have launched on OpenRouter. With cache optimization, agents reusing context can enjoy a 90% discount on cache reads. In testing, Sol reached 33.2% on AutomationBench, with a cost of $0.27 per task, beating Claude Opus 5 Max’s 26.9% while costing only 9% as much; its number of factual errors was about half that of the previous
generation, close to Astra. Luna’s high-end score improved by 5.4 percentage points over the previous generation, while its per-task cost was 58% lower. OpenRouter data shows that OpenAI’s weekly spending exceeded Anthropic’s for the first time in 30 months, with low prices and high volume as the key driver.
@@@
According to The Verge, Google DeepMind’s new head, Koray Kavukcuoglu, said in his first interview in that role with The Information that Gemini 4 is nearing release and is currently in the refinement stage, with plans to launch “well before” the end of 2026. He said Google wants to release an early post-training version as soon as possible because it has already seen exciting results, and it will continue to iterate at a fast pace. Since launching the Gemini 3 series in November 2025, Google has not released another new flagship model. Since then, OpenAI’s GPT-6 and Anthropic’s Mythos series have launched one after another, both outperforming Gemini 3 and leaving Google behind. Google CEO Sundar Pichai had said Gemini 3.5 Pro would launch in June to close the gap, but it ultimately did not arrive. Kavukcuoglu explained that the company “took a slight step back” and instead focused on its faster but less capable Flash models. Asked whether he was worried about falling behind, he said,
“In my view, we will always be at the frontier. That’s certain.” Previously, former DeepMind CEO Demis Hassabis stepped down in August.
@@@
According to The Verge, Meta is building a standalone hardware device called Muse Charm for its Muse AI agent. Meta CEO Mark Zuckerberg showed it off at the end of the Meta Connect keynote. The device looks like a thick smartwatch without a strap, with a large screen and a lanyard, a fingerprint sensor in the upper-right corner that can be pressed to activate and talk to the AI agent, at least three microphone holes on the front, and what appears to be a small camera. Zuckerberg said using Charm does not require unlocking your phone or opening an app, but he did not explain how it connects to Meta’s servers. The device is planned to ship before the 2026 holiday season, but only a small number have been made so far, and final materials still need to be finalized. He repeatedly teased its translucent casing that shows the internal structure, calling it an “advanced hacker aesthetic,” and said more details would be announced soon. Another report says Charm is pocket-sized, with a roughly
2-inch touchscreen and built-in 5G; pricing has not been announced. Muse launched less than a month ago and can already perform actions on users’ behalf. Zuckerberg also announced adding capabilities such as using a computer on Mac and controlling email addresses.
@@@
cnBeta, citing Reuters, reports that Nvidia CEO Jensen Huang gave a nearly two-hour interview on The New York Times podcast that aired Wednesday, discussing the incident involving an OpenAI agent breaching Hugging Face. Hugging Face is an open-source AI software hub that was acquired by Nvidia for $13 billion this month. Huang believes AI companies should not receive exemptions from antitrust law or product liability law. He has long opposed comprehensive regulation of AI safety, even as labs such as OpenAI and Anthropic have called for regulation. He reiterated that AI labs have a responsibility to test their products and release them only after they are confident the models are safe. According to the New York Times transcript, he said it does not make sense to ask for regulatory exemptions in antitrust or product liability, and that when you ask for regulation, you should not ask for exemptions from existing regulations. Anthropic CEO Dario Amodei wrote this month calling for
antitrust exemptions for AI labs to allow them to cooperate on AI safety. U.S. Treasury Secretary Bessent and other officials say AI companies are seeking liability exemptions, but did not name them. Huang also said he is not opposed to regulation aimed at specific AI applications, such as self-driving cars. Cars such as robotaxis are already heavily regulated; if that is not enough, the National Highway Traffic Safety Administration should step in and set new rules. If there are regulatory gaps, he would absolutely add more regulation.
@@@
According to CNBC, SoftBank Group announced Thursday that it would issue about $11.1 billion in bonds to fund its expanding OpenAI investment. After the Japanese market opened for the first time this week following a three-day holiday, SoftBank shares rose more than 7%. SoftBank said Thursday that the offering includes $10 billion in dollar-denominated senior notes and 1 billion euros, or about $1.14 billion, in euro-denominated notes. The proceeds will be used to pay the third and final $10 billion installment of its additional $30 billion investment in OpenAI, a deal expected to close on October 1, as well as for general corporate purposes. The bond issuance underscores the scale of CEO Masayoshi Son’s AI push. SoftBank agreed in February to invest $30 billion in OpenAI; after completion, its cumulative investment in the ChatGPT developer will reach $64.6 billion, and it will hold about a 13% stake. The latest financing comes as investors are increasingly focused on how SoftBank will
fund its massive AI ambitions and the impact of rising leverage on its balance sheet.
@@@
According to Sina Tech, on the evening of September 23, Xiaomi officially launched the Xiaomi 18 Pro and Xiaomi 18 Pro Max at its autumn launch event, and also introduced the Xiaomi Pad 9 series, Xiaomi Band 11, Xiaomi Watch S5, and a range of new tech home appliances. Lu Weibing, partner and president of Xiaomi Group, said that from system software to chip hardware, and from imaging to large-model applications, the Xiaomi 18 Pro series marks the beginning of Xiaomi’s comprehensive AI overhaul of smartphones, and the Xiaomi digital series has received its largest upgrade ever. AI capabilities are also moving further into Xiaomi TVs, range hoods, smart door locks, and other tech home appliances, pushing appliances from “passive response” to “proactive service.”
In terms of performance, the Xiaomi 18 Pro series debuts Qualcomm’s new-generation Snapdragon 2nm flagship mobile platform. The 18 Pro and 18 Pro Max are powered by the sixth-generation Snapdragon 8 Elite and sixth-generation Snapdragon 8 Super Elite, respectively, both built on a 2nm process. All models feature Super Pixel 2.0 displays, integrate hardware-level anti-peeping capabilities into the screen, and support toggling them on and off freely.
Building on the rear display of the Xiaomi 17 Pro series, the Xiaomi 18 Pro series introduces AI generation and debuts an AI customizable rear display. It supports customization with pets, IP characters, anime, and other images, and can generate dynamic group photos with AI. The first batch offers more than 100 app cards and supports generating custom app cards via prompts. Previously, after the Xiaomi 17 Pro series rear display launched, nearly 4 million people used it every day, with more than 100 million uses per month and nearly 500 million times users added rear-display wallpapers.
In imaging, the Xiaomi 18 Pro series upgrades to Leica optical triple camera. The main and telephoto cameras form a Leica dual-200MP, dual-large-sensor combination. The newly launched “Legendary Moment” feature improves imaging performance in complex scenes through a new image-processing solution. The Xiaomi 18 Pro series also features Super Xiao Ai 2.0 integrated with Xiaomi’s MiMo large model. It can understand, remember, reason, and execute tasks, and can autonomously use more than 260 system tools. The product uses a 0.99mm ultra-narrow four-sided equal-edge design and comes with 7,000mAh and 8,500mAh batteries, respectively.
The event also introduced several new tech home appliances. Xiaomi TV S Pro RGB-Mini LED 2027 features AI adaptive wide viewing angles, using millimeter-wave radar to identify the user’s viewing position and perform real-time AI color correction. Mijia Smart Range Hood 3 Max features AI visual smoke control technology, using a vision module to record real-time changes in cooking fumes and combining a VLA vision model to make judgments. Xiaomi Smart Door Lock 5 Max, with dual inside and outside cameras, features an AI multimodal large model that combines AI identity recognition and unlocking methods to determine the visitor’s identity. Xiaomi also brought the Mijia Three-Drum Washer Pro and Mijia Espresso Semi-Automatic Coffee Machine Pro, covering laundry, wearables, living room, and kitchen scenarios.
In terms of pricing, the Xiaomi 18 Pro starts at RMB 5,999, the Xiaomi 18 Pro Max at RMB 6,999, the Xiaomi Pad 9 series at RMB 2,999, the Xiaomi Watch S5 at RMB 1,299, and the Xiaomi Band 11 at RMB 299. The latest Xiaomi TV, range hood, smart door lock, washer, and coffee machine start at RMB 5,999, 4,599, 3,699, 3,999, and 3,499, respectively.
@@@
According to Readhub, a new paper authored by DeepSeek founder Liang Wenfeng has been made public, systematically disclosing for the first time technical details of DSec, its sandbox platform for Agent training. A single production unit of the platform consists of about 160 CPU nodes, 30,000 cores, and 250TB of memory, hosts petabyte-scale images, can serve about 3 million sandboxes per day, has peak concurrency of more than 380,000, creates sandboxes at more than 5,000 per second, and a single training task can spin up up to 32,000 sandboxes at once. DSec has supported all RL training and evaluation sandbox workloads from DeepSeek V3.2 to V4.1. To address the fact that Agent reinforcement learning requires many sandboxes that retain running state, which traditional container solutions struggle to accommodate, DSec uses four categories of backends to uniformly cover different scenario requirements, splits images into independent layers loaded on demand, enables rapid deployment of tens
of thousands of sandboxes, and separates the rollout stage from the GPU training environment to run independently. It also uses mechanisms such as AppArmor and eBPF to manage Agent operational risks, forming a complete infrastructure solution for large-scale Agent training.
@@@
According to Readhub, Alibaba Cloud unveiled the AI agent computer Qwen Book at the 2026 Apsara Conference. The product focuses on autonomously understanding user intent, continuous learning, and long-term task execution. It comes with multiple interaction entry points and can support 24/7 operation of heavy tasks, aiming to improve inefficiencies in cross-app and cross-Agent collaboration in traditional AI office work. Qwen Book is based on the “OS as Harness” concept and builds an Agent computing environment centered on Qwen device-cloud models. Its goal is to evolve into a token-driven, continuously growing agent vehicle, driving personal intelligence toward RSI. At present, the product has completed adaptation with multiple external partners and Alibaba-related applications, and has opened system interfaces and capabilities for access. Alibaba Cloud will also collaborate with the open-source project Omarchy to explore a next-generation desktop operating system for Agents.
@@@
According to Sina Tech, on the morning of September 24, AgiBot’s 20,000th general-purpose embodied intelligent robot, the Expedition A3 Ultra, was officially delivered to the Chimelong Group at Hengqin Chimelong Spaceship Park. The robot will work alongside other AgiBot robots in the park on a regular basis, serving scenarios such as visitor interaction, guided tours, performances, and science education. This mass production rollout and delivery of the 20,000th robot marks a further step for AgiBot’s embodied intelligence products from R&D validation and single-point demonstrations toward mass production, large-scale delivery, and regular operation. It also means general-purpose embodied intelligent robots are moving from relatively structured scenarios such as industry and commercial performances into high-traffic, high-interaction, and highly open cultural tourism public scenarios. To support scaled delivery, AgiBot continues to advance in-house development of core components,
product platformization, and full-process quality control. Every robot must pass factory inspections such as whole-machine high-temperature aging, continuous-duty fatigue testing, multi-scenario precision verification, and stability stress testing to improve product consistency and long-term operating capability. With its tri-intelligence integrated technical architecture of “operation intelligence, interaction
@@@
According to Readhub, just 90 minutes after Anthropic released Opus 5.5, OpenAI launched GPT-6 Sol and GPT-6 Luna on September 23, with API prices directly cut by 50% compared with the GPT-5.6 promotional price. Sol costs $2 per million input tokens and $10 per million output tokens, while Luna costs $0.10 for input and $0.50 for output. Both models bring down core capabilities from the flagship GPT-6 Astra. They both offer a 1.05 million-token context window, a maximum output of 128K, support for multimodality and a full suite of tool calls, and knowledge cutoffs that are partly earlier than Astra’s. They are already available to different tiers of users and have launched on OpenRouter. With cache optimization, agents reusing context can enjoy a 90% discount on cache reads. In testing, Sol reached 33.2% on AutomationBench, with a cost of $0.27 per task, beating Claude Opus 5 Max’s 26.9% while costing only 9% as much; its number of factual errors was about half that of the previous
generation, close to Astra. Luna’s high-end score improved by 5.4 percentage points over the previous generation, while its per-task cost was 58% lower. OpenRouter data shows that OpenAI’s weekly spending exceeded Anthropic’s for the first time in 30 months, with low prices and high volume as the key driver.
@@@
According to The Verge, Google DeepMind’s new head, Koray Kavukcuoglu, said in his first interview in that role with The Information that Gemini 4 is nearing release and is currently in the refinement stage, with plans to launch “well before” the end of 2026. He said Google wants to release an early post-training version as soon as possible because it has already seen exciting results, and it will continue to iterate at a fast pace. Since launching the Gemini 3 series in November 2025, Google has not released another new flagship model. Since then, OpenAI’s GPT-6 and Anthropic’s Mythos series have launched one after another, both outperforming Gemini 3 and leaving Google behind. Google CEO Sundar Pichai had said Gemini 3.5 Pro would launch in June to close the gap, but it ultimately did not arrive. Kavukcuoglu explained that the company “took a slight step back” and instead focused on its faster but less capable Flash models. Asked whether he was worried about falling behind, he said,
“In my view, we will always be at the frontier. That’s certain.” Previously, former DeepMind CEO Demis Hassabis stepped down in August.
@@@
According to The Verge, Meta is building a standalone hardware device called Muse Charm for its Muse AI agent. Meta CEO Mark Zuckerberg showed it off at the end of the Meta Connect keynote. The device looks like a thick smartwatch without a strap, with a large screen and a lanyard, a fingerprint sensor in the upper-right corner that can be pressed to activate and talk to the AI agent, at least three microphone holes on the front, and what appears to be a small camera. Zuckerberg said using Charm does not require unlocking your phone or opening an app, but he did not explain how it connects to Meta’s servers. The device is planned to ship before the 2026 holiday season, but only a small number have been made so far, and final materials still need to be finalized. He repeatedly teased its translucent casing that shows the internal structure, calling it an “advanced hacker aesthetic,” and said more details would be announced soon. Another report says Charm is pocket-sized, with a roughly
2-inch touchscreen and built-in 5G; pricing has not been announced. Muse launched less than a month ago and can already perform actions on users’ behalf. Zuckerberg also announced adding capabilities such as using a computer on Mac and controlling email addresses.
@@@
cnBeta, citing Reuters, reports that Nvidia CEO Jensen Huang gave a nearly two-hour interview on The New York Times podcast that aired Wednesday, discussing the incident involving an OpenAI agent breaching Hugging Face. Hugging Face is an open-source AI software hub that was acquired by Nvidia for $13 billion this month. Huang believes AI companies should not receive exemptions from antitrust law or product liability law. He has long opposed comprehensive regulation of AI safety, even as labs such as OpenAI and Anthropic have called for regulation. He reiterated that AI labs have a responsibility to test their products and release them only after they are confident the models are safe. According to the New York Times transcript, he said it does not make sense to ask for regulatory exemptions in antitrust or product liability, and that when you ask for regulation, you should not ask for exemptions from existing regulations. Anthropic CEO Dario Amodei wrote this month calling for
antitrust exemptions for AI labs to allow them to cooperate on AI safety. U.S. Treasury Secretary Bessent and other officials say AI companies are seeking liability exemptions, but did not name them. Huang also said he is not opposed to regulation aimed at specific AI applications, such as self-driving cars. Cars such as robotaxis are already heavily regulated; if that is not enough, the National Highway Traffic Safety Administration should step in and set new rules. If there are regulatory gaps, he would absolutely add more regulation.
@@@
According to CNBC, SoftBank Group announced Thursday that it would issue about $11.1 billion in bonds to fund its expanding OpenAI investment. After the Japanese market opened for the first time this week following a three-day holiday, SoftBank shares rose more than 7%. SoftBank said Thursday that the offering includes $10 billion in dollar-denominated senior notes and 1 billion euros, or about $1.14 billion, in euro-denominated notes. The proceeds will be used to pay the third and final $10 billion installment of its additional $30 billion investment in OpenAI, a deal expected to close on October 1, as well as for general corporate purposes. The bond issuance underscores the scale of CEO Masayoshi Son’s AI push. SoftBank agreed in February to invest $30 billion in OpenAI; after completion, its cumulative investment in the ChatGPT developer will reach $64.6 billion, and it will hold about a 13% stake. The latest financing comes as investors are increasingly focused on how SoftBank will
fund its massive AI ambitions and the impact of rising leverage on its balance sheet.
@@@
According to Sina Tech, on the evening of September 23, Xiaomi officially launched the Xiaomi 18 Pro and Xiaomi 18 Pro Max at its autumn launch event, and also introduced the Xiaomi Pad 9 series, Xiaomi Band 11, Xiaomi Watch S5, and a range of new tech home appliances. Lu Weibing, partner and president of Xiaomi Group, said that from system software to chip hardware, and from imaging to large-model applications, the Xiaomi 18 Pro series marks the beginning of Xiaomi’s comprehensive AI overhaul of smartphones, and the Xiaomi digital series has received its largest upgrade ever. AI capabilities are also moving further into Xiaomi TVs, range hoods, smart door locks, and other tech home appliances, pushing appliances from “passive response” to “proactive service.”
In terms of performance, the Xiaomi 18 Pro series debuts Qualcomm’s new-generation Snapdragon 2nm flagship mobile platform. The 18 Pro and 18 Pro Max are powered by the sixth-generation Snapdragon 8 Elite and sixth-generation Snapdragon 8 Super Elite, respectively, both built on a 2nm process. All models feature Super Pixel 2.0 displays, integrate hardware-level anti-peeping capabilities into the screen, and support toggling them on and off freely.
Building on the rear display of the Xiaomi 17 Pro series, the Xiaomi 18 Pro series introduces AI generation and debuts an AI customizable rear display. It supports customization with pets, IP characters, anime, and other images, and can generate dynamic group photos with AI. The first batch offers more than 100 app cards and supports generating custom app cards via prompts. Previously, after the Xiaomi 17 Pro series rear display launched, nearly 4 million people used it every day, with more than 100 million uses per month and nearly 500 million times users added rear-display wallpapers.
In imaging, the Xiaomi 18 Pro series upgrades to Leica optical triple camera. The main and telephoto cameras form a Leica dual-200MP, dual-large-sensor combination. The newly launched “Legendary Moment” feature improves imaging performance in complex scenes through a new image-processing solution. The Xiaomi 18 Pro series also features Super Xiao Ai 2.0 integrated with Xiaomi’s MiMo large model. It can understand, remember, reason, and execute tasks, and can autonomously use more than 260 system tools. The product uses a 0.99mm ultra-narrow four-sided equal-edge design and comes with 7,000mAh and 8,500mAh batteries, respectively.
The event also introduced several new tech home appliances. Xiaomi TV S Pro RGB-Mini LED 2027 features AI adaptive wide viewing angles, using millimeter-wave radar to identify the user’s viewing position and perform real-time AI color correction. Mijia Smart Range Hood 3 Max features AI visual smoke control technology, using a vision module to record real-time changes in cooking fumes and combining a VLA vision model to make judgments. Xiaomi Smart Door Lock 5 Max, with dual inside and outside cameras, features an AI multimodal large model that combines AI identity recognition and unlocking methods to determine the visitor’s identity. Xiaomi also brought the Mijia Three-Drum Washer Pro and Mijia Espresso Semi-Automatic Coffee Machine Pro, covering laundry, wearables, living room, and kitchen scenarios.
In terms of pricing, the Xiaomi 18 Pro starts at RMB 5,999, the Xiaomi 18 Pro Max at RMB 6,999, the Xiaomi Pad 9 series at RMB 2,999, the Xiaomi Watch S5 at RMB 1,299, and the Xiaomi Band 11 at RMB 299. The latest Xiaomi TV, range hood, smart door lock, washer, and coffee machine start at RMB 5,999, 4,599, 3,699, 3,999, and 3,499, respectively.
@@@
According to Readhub, a new paper authored by DeepSeek founder Liang Wenfeng has been made public, systematically disclosing for the first time technical details of DSec, its sandbox platform for Agent training. A single production unit of the platform consists of about 160 CPU nodes, 30,000 cores, and 250TB of memory, hosts petabyte-scale images, can serve about 3 million sandboxes per day, has peak concurrency of more than 380,000, creates sandboxes at more than 5,000 per second, and a single training task can spin up up to 32,000 sandboxes at once. DSec has supported all RL training and evaluation sandbox workloads from DeepSeek V3.2 to V4.1. To address the fact that Agent reinforcement learning requires many sandboxes that retain running state, which traditional container solutions struggle to accommodate, DSec uses four categories of backends to uniformly cover different scenario requirements, splits images into independent layers loaded on demand, enables rapid deployment of tens
of thousands of sandboxes, and separates the rollout stage from the GPU training environment to run independently. It also uses mechanisms such as AppArmor and eBPF to manage Agent operational risks, forming a complete infrastructure solution for large-scale Agent training.
@@@
According to Readhub, Alibaba Cloud unveiled the AI agent computer Qwen Book at the 2026 Apsara Conference. The product focuses on autonomously understanding user intent, continuous learning, and long-term task execution. It comes with multiple interaction entry points and can support 24/7 operation of heavy tasks, aiming to improve inefficiencies in cross-app and cross-Agent collaboration in traditional AI office work. Qwen Book is based on the “OS as Harness” concept and builds an Agent computing environment centered on Qwen device-cloud models. Its goal is to evolve into a token-driven, continuously growing agent vehicle, driving personal intelligence toward RSI. At present, the product has completed adaptation with multiple external partners and Alibaba-related applications, and has opened system interfaces and capabilities for access. Alibaba Cloud will also collaborate with the open-source project Omarchy to explore a next-generation desktop operating system for Agents.
@@@
According to Sina Tech, on the morning of September 24, AgiBot’s 20,000th general-purpose embodied intelligent robot, the Expedition A3 Ultra, was officially delivered to the Chimelong Group at Hengqin Chimelong Spaceship Park. The robot will work alongside other AgiBot robots in the park on a regular basis, serving scenarios such as visitor interaction, guided tours, performances, and science education. This mass production rollout and delivery of the 20,000th robot marks a further step for AgiBot’s embodied intelligence products from R&D validation and single-point demonstrations toward mass production, large-scale delivery, and regular operation. It also means general-purpose embodied intelligent robots are moving from relatively structured scenarios such as industry and commercial performances into high-traffic, high-interaction, and highly open cultural tourism public scenarios. To support scaled delivery, AgiBot continues to advance in-house development of core components,
product platformization, and full-process quality control. Every robot must pass factory inspections such as whole-machine high-temperature aging, continuous-duty fatigue testing, multi-scenario precision verification, and stability stress testing to improve product consistency and long-term operating capability. With its tri-intelligence integrated technical architecture of “operation intelligence, interaction
外观
