爱窝啦 AI 日报 AI Daily 2026/10/28: Google Genie 4 Brings World Models into the Age of Audio Interaction, Jimeng AI Launches Short-Drama Studio for End-to-End Production
每日 AI 行业动态、产品更新、前沿研究、开源项目与社媒信号
Today’s Summary
Google Genie 4 natively generates ambient sound and dialogue; OpenAI released GPT-6.1 Terra on the same day, focused on computer control — two major labs launching new products simultaneously.
AI is shifting from executing single instructions to actively managing tasks and verifying tens of thousands of pages of mathematical proofs, with capability boundaries expanding in multiple directions at once.
Start with item 4 on math verification and item 8 on workload reflection, then decide whether to follow up on hands-on testing of Genie 4 or Terra.🔥 Today’s Top 10 Highlights
1. Google Genie 4 Brings World Models into the Age of Audio Interaction
The world model has been upgraded. Google DeepMind officially released the next-generation general-purpose world model Genie 4 , which for the first time natively generates ambient sound and spontaneous character dialogue 🔊. A single continuous generation runs up to approximately 2 minutes, with interactive visuals at around 24 frames per second. Three Gemini models collaborate with distinct roles: Gemini 3.8 Flash handles scene monitoring, and Gemini 4 Argon breaks down fine-grained actions — though Ultra subscribers still cannot access it directly.
2. Jimeng AI Launches Short-Drama Studio for End-to-End Production
The barrier to short-drama production has dropped. Jimeng AI has integrated script diagnosis, storyboard generation, video production, and post-production into a single platform. As reported in Jimeng AI launches short-drama studio for end-to-end production , its DreamScript model supports long texts of up to 100,000 characters and generation of up to 100 episodes. The team edition supports multi-account and credit unified management, with assets automatically parsed and bound to the canvas. Consistency across long scripts still needs to be verified in practice; production teams with scaled short-drama needs are encouraged to test it first.
3. AI Spends 24 Days Auditing the Longest Proof in Math History, Finds One Gap
The longest proof in mathematical history has been verified by computer for the first time. Math AI company Axiom announced on October 26 that they used dozens of AI agents over 24 days to complete a formal verification of the entire Classification of Finite Simple Groups theorem. The proof is scattered across more than 500 papers totaling approximately 15,000 pages; one paper’s proof contained a gap, which the AI patched by substituting a new proof. This is the largest-scale mathematical formalization project ever completed, and is a meaningful reference for researchers working in mathematics or logic verification.
4. Developer Uses 629 Articles to Measure the Telltale “Europeanized” Traits of AI Writing
AI writing has measurable stylistic markers. Jike user Jiangzhi Hou pushed a developer’s study using 629 articles to identify the typical Europeanized traits of AI writing to GitHub. The corpus was expanded to approximately 5.2 million characters, covering 270 pairs of articles on the same topic with the same facts. The evaluation identified key features including Europeanized balanced phrasing, dense parallel structures, and low aptness of metaphors, and used a semantic model to make approximately 430,000 judgments. Writers looking to reduce the detectability of AI-generated content can refer to this open-source evaluation framework.
5. Anthropic Engineering Lead Announces Candidates Will No Longer Write Code by Hand in Interviews
The format of coding interviews has changed. Anthropic’s engineering lead publicly stated that they have stopped asking candidates to solve algorithm problems during interviews. The specific strategies in the original post now emphasize pursuing business impact and rapid prototype validation. He believes that with the help of AI, new graduates have the opportunity to quickly surpass engineers with far more experience. Recent graduates preparing for interviews at major tech companies may want to reassess how much time they allocate to grinding algorithm problems.
6. GPT-6.1 Terra Leads with Computer Control Capabilities and a Low-Price Strategy
OpenAI releases a new-tier model. Sam Altman announced on Twitter that GPT-6.1 Terra is now live, positioned to deliver fast and practical computer control capabilities 💻 at a lower price. The model can understand screen content, execute actions across applications, and write naturally, targeting full-workday use cases. Some users on V2EX have reported that its performance falls short of previous versions, though OpenAI has not officially responded to those claims.
7. A Tech Lead’s Firsthand Account: AI Increased Workload Rather Than Reducing It
AI didn’t reduce work — it created more tasks. Bold.org VP of Engineering O Ioannou reflected on his personal blog that in August 2026 his GitHub contribution count reached 427, surpassing his previous single-month record. His conclusion: AI lowered the friction of execution, but also led him to kick off more tasks simultaneously, with his attention fragmented by approval requests from multiple agents. The core frustration in this reflection is that his capacity for planning and judgment has not kept pace with the growth of his coding capacity.
8. Wikipedia Volunteers Identify a New AI Content Threat: Distorting Real Sources
AI contamination has shifted from fabricating citations to misrepresenting real ones. IEEE Spectrum interviewed a Wikipedia volunteer about this new AI content threat: distorting real sources . She noted that the problem is escalating: the papers being cited are real, but they do not support what the article claims 🔍. She estimates that around 20% of new articles added in late 2025 were flagged as AI-generated, dropping to approximately 15%–16% by early 2026. These figures are based on specific detection methods and do not represent the overall proportion across all of Wikipedia.
9. Roblox Accused of Scraping Over 145 GB of 3D Assets from Open3DLab
A 3D asset site hit by large-scale unauthorized crawling. Open3DLab disclosed in a site announcement that a crawler identifying itself as serving Roblox research read nearly 19,000 project pages and downloaded 145.5 GB of files over approximately 34 hours, including 345 adult-content items and 423 items licensed for non-commercial use only. The crawler continued scraping after reading robots.txt, which the site considers a violation of its terms of service. The attribution is based on user agent and email domain; Roblox has not officially responded, and the purpose of the scraping has not been independently verified.
10. China’s First Fully AI-Generated Film Earns Only About $180,000 at the Box Office
Fully AI-generated film delivers a disappointing box office result. According to the Telegram channel AI探索指南 , China’s first fully AI-generated commercial film earned only approximately $181,000 at the box office after release. This figure comes from a Telegram channel relay and is not backed by official box office platform data, so it should be treated with uncertainty. For practitioners evaluating the return on investment of AI in film and television, this may be an early market signal worth watching.
◎ Industry Shifts and Personal Impact
DeepSeek Opens Campus Recruiting for HR Roles, Seeking Recruiters to Support Algorithm Teams
DeepSeek is expanding its internal recruiting team. A job posting from Jike user Deepthink shows that the DeepSeek HR team is recruiting full-time permanent HR staff from recent graduates, focused on supporting algorithm and core systems infrastructure, based in Beijing. This information comes from a social media post; the application email is the address provided in that post. Recent graduates interested in non-technical roles at AI fundamental research companies may want to verify the details independently before applying.
⌘ Top Open Source Projects
nightly-labs/openbot: Local Multi-Agent Desktop Workspace
A local-first AI team management tool topped today’s daily rankings. nightly-labs/openbot is a TypeScript desktop application that can run multiple models simultaneously — including Codex, Claude, and Grok — and assigns each agent its own workspace and task queue. It gained 777 new stars today, bringing its total to 2,663. It supports file sharing, browser control, and inter-agent collaboration, making it worth trying for developers who need to manage multi-model workflows locally.
obra/superpowers: Agent Skills Framework and Software Development Methodology
A skills framework for AI-assisted development topped today’s daily rankings. obra/superpowers is implemented in Shell and provides a practical agent skills structure and software development methodology. It gained 617 new stars today, bringing its total to 306,666. It is a useful reference for engineers looking to standardize AI-assisted development workflows and improve agent reusability.
😄 AI Fun
AI Helps Dad Make a Slideshow — Father Gets Digitally Managed in Real Time
A developer set up an AI WeChat assistant called “Lele” for their elderly parent, running on DeepSeek Harness. When dad was preparing to give a talk at the community center, he sent his materials to Lele: “Help me put together a ten-minute presentation to share with the neighbors.” Lele produced a PPTX file. Dad thought the font was too small and the jargon too dense, so he said so in WeChat — and Lele handed back a revised version. The Jike post ends with: the son just needs to keep upgrading the assistant, while the assistant is also iterating based on the elderly user’s needs. So at this point, exactly who is taking care of whom remains an open question.
❓ Related Questions
GPT-6.1 Terra Is Out — How Can Regular Users Access the Computer Control Features?
GPT-6.1 Terra is now live; specific access depends on OpenAI’s official rollout schedule. According to Sam Altman’s announcement , Terra is focused on computer control capabilities and lower usage costs, targeting full-workday use cases. The specific scope of Terra’s available features should be checked on OpenAI’s official pages; having an account does not automatically unlock all new capabilities.