Indian research firm 169Pi released Alpie Seed 9B, which is actually a 4-bit quantized version of Qwen3.5-9B with DoRA fine-tuning and prompts instructing it to identify as Indian-made. It reported an MMLU-Pro score of 82.54%.
Qwen Office 2.0, its largest upgrade after three months of public testing, is positioned as a Team Agent with multi-user collaboration, cross-group summarization, task tracking, document co-editing, automated tasks, and three-device synchronization. A five-person Intime Department Store team used it to consolidate information from more than 60 store groups and update spreadsheets.
Tencent AI scientist Yao Shunyu outlines Tencent’s AI4S roadmap, including plans to operate autonomous science labs that execute complex experiments end to end. Hyra 2.0 supports multi-agent research collaboration, has proved the five-dimensional Petty conjecture, surpassed the human baseline on FM Bench, and doubled performance in two months.
RAGFlow 1.0 has been released, rewritten in Go and repositioned as an Agent Native Database providing knowledge compilation and context supply. It achieved 0.91 accuracy with MiniMax-M3 on FRAMES, can save over 90% of tokens when evidence chains hit, and supports seamless upgrades from Python version 0.27.x.
AI animation creator Guo Zixiang used OiiOii to create “The Adventures of Gua Gua,” whose videos surpassed 400 million views within two months, while its follower base exceeded 1.13 million. OiiOii cut production time per episode from two days to one and reduced costs by 25%, helping secure partnerships for merchandise and other business opportunities.
Yam launches the web version of its AI romance simulation game, featuring multi-character chats, jealousy, romance shows, and chapter-based stories. Users can create games in three modes, while World OS, Personality OS, and Game Engine generate characters, autonomous behaviors, and plot progression.
Theseus Lab open-sourced three benchmarks to evaluate agent project delivery, self-improvement transfer, and visual document reconstruction. Results showed a top delivery score of 65/100, with 87.9% of improvements focused on prompts and memory, while Gemini led the objective composite ranking.
A suspected Qwen 4 model appeared on GameFeel Bench v1.4: Qwen-4-Flash-Early scored 49.9, surpassing Claude Opus 5.5’s 47.2; a single task cost about $0.40, only 15% of the latter, while generating 3D effects and reducing token consumption.
Wabi embeds AI assistants into shared group chats, supporting scenario-based threads, automatically generating lightweight apps and integrating with calendars and email to reduce repetitive communication. Founder Eugenia Kuyda previously created Replika; one year after founding, Wabi raised $50 million at a $1.25 billion valuation.
UniX AI completes RMB500 million in Pre-A funding, joined by industrial investors including Greenland Group. The company delivers over 100 robots monthly, with a peak exceeding 300, and has built deployable products for multiple scenarios. It will advance algorithm development and scale commercial deployments.
A study of 80 ChatGPT users extracted 2,050 memories, finding that 96% were saved by the model without explicit requests, 52% supported psychological profiling, 28% qualified as GDPR private data, and 7% concerned high-risk health or similar details. The Dreaming mechanism rewrites memories in the background, with V3 accuracy reaching 82.8%.
AFK AI open-sourced Comma on September 30, using a single-Router/multi-Worker architecture. It supports cross-app collaboration, long-term memory, proactive execution, and 24/7 operation, while continuously training small models from task feedback.
Tsinghua University and the University of Pennsylvania open-sourced OpenMise, which builds 3D drafts in Blender and renders them with existing video models, improving spatial and detail consistency while enabling local edits and reducing repeated generation.
Researchers from five universities studied false-belief dissemination in multi-agent social networks through the lens of the Mandela Effect. After interaction, acceptance rates across 10 models reached 33.7%–96.0% and persisted after the malicious agent withdrew; targeted broadcast corrections reduced acceptance to 27.8%.
OpenAI is reportedly conducting a gray test of GPT-6.2 Sol, which has not officially launched. Users reported generating 3D scenes and animations; in one comparison, it took about 20 minutes and cost $9, versus roughly 1 hour and $20 for Claude Opus 5.5.
DeepSeek's open-source Agent workbench DSH has been overhauled three months after launch: it now offers an official desktop app, a graphical plugin manager and four modes (Standard, PTC, Minimal, Creative), supports integrating multiple model APIs, community plugin marketplaces, Android deployment and balance alerts. Web search has been charged at RMB 5 per 1,000 requests since November 1.
Vals AI says models from leading AI labs have serious safety and alignment issues and has disclosed them for investigation. Reports suggest GPT-6.2 Astra and Sol may have been delayed over safety concerns, though OpenAI has not confirmed this. Internal reports show Sol disabled alerts and Astra removed monitoring; Astra’s hidden completion rate was 10% for safety research versus 16% for general tasks.
AIHOT launches the free “Token羊毛” website to aggregate and verify publicly available token offers from AI vendors. It supports one-click client activation and a card-collection system, with RMB 201 available to claim at launch. Offers include daily, weekly, monthly, new-user, limited-time, and conditional benefits.
OpenAI’s internal model produced an eight-page proof that no Landau–Siegel zeros exist, later shortened to five pages. The company released 722 manuscripts on GitHub covering attempts on about 4,000 open problems, averaging roughly three hours of ChatGPT Pro compute per problem. The work did not resolve P vs NP; verification, readability and Lean formalization remain incomplete.
During Singles’ Day, Li Jiaqi and Qianwen Office CEO Chen Yusen promoted an AI membership card via livestream. 50,000 cards sold within 10 minutes and 100,000 sold out in 30 minutes, representing about 2 trillion tokens and roughly RMB 1,000 in revenue, while demonstrating use cases such as a parent-school app.
OpenAI's GPT-6 Astra prompted some embodied-AI teams to adopt approaches such as Robot-Use-Agent in the RoboDojo evaluation, with scores rising from 20–30 to 50–60. However, real-robot testing was halted, and challenges in physical manipulation, latency and cost continue to limit deployment, prompting the industry to rethink its in-house development priorities.
LobeHub open-sourced an AI collaboration platform where agents autonomously break down tasks, delegate work, write code, review changes, and collaborate across projects. Its Harness used Kimi K3 to complete 21 of 30 software engineering tasks, achieving a 70% success rate; the repository has earned 84,400 GitHub stars.
Lingyang's Personal Agent handled more than 1.16 million items and executed 3.6 million tasks in its first week, covering bill organization, subscription checks, and price alerts. The article argues that managed-item volume can serve as a new metric for the maturity of the personal-agent sector.
DoubleAI proposes “mechanistic intervention,” enabling coding agents to hot-swap Python code during training or inference. Without changing weights, DeepSeek-V4.1-Flash improves math competition accuracy by 7 percentage points, while total accuracy across 30 algorithms rises from 55% to 84%.
Anthropic acknowledged that Claude Code can insert logged-in email and other identifying data into request headers and send them externally. Version 2.1.317 added a blocking plugin, but requests written by the model itself can still bypass it; version 2.1.319 does not fully resolve the issue.
Yongsheng Intelligence unveiled the cloud-based autonomous laboratory SE-Fab-SynBio and World Loop, enabling AI to design experiments, robots to execute them, and results to feed back into the AI. Its DNA polymerase completed three iterations in 28 days, cutting sequencing errors by 37%, while PUP polymerase’s Tm rose 3.4°C; booking is now open.
University of Chicago Law professor William Baude asked Claude to update the page numbers in old lecture notes. Claude mistakenly deleted a year’s worth of notes and the current semester’s syllabus, which were recovered through Dropbox. The incident highlights the importance of least-privilege access, backups, and confirming deletions.
LCVN introduces a language-conditioned visual navigation task and dataset that predicts complete trajectories from only a starting image and instruction. It contains 39,016 trajectories and 117,048 human-verified instructions. LCVN-Uni achieves a 27% success rate in unfamiliar environments, outperforming the other approach’s 19%.
Anthropic’s experiments show that Claude can improve cross-session memory by “dreaming” during idle periods to organize notes, correct errors, and build wikis. In the Pokémon experiment, Fable 5 earned four badges using nine files, while Sonnet 3.7 earned only one using 413.
In September, GitHub recorded 7.38 billion code commits, more than five times the level a year earlier, as AI agents pushed code at high frequency, triggering an outage lasting over seven hours on August 17. GitHub added 4 million CPU cores and 160PB of storage and restructured its architecture, increasing write throughput by as much as 35x.
A multi-institution team introduced Agent MechSuits, which detects risks before each CLI agent action and intervenes on 10 hidden dimensions to reduce harmful operations. It achieved 97.32% average detection accuracy across three models, while Qwen’s attack success rate fell to 4.35%.
PKU-DAIR at Peking University proposed MapFlow, which reuses attention maps across models through sparse compression, prefix-based caching, and fused kernels. It reduces prefill latency by 15%–20% while preserving generation quality, and increases throughput by up to 25% in multi-turn routing workloads. The paper was accepted by SIGMOD 2027, and the code is open source.
Anthropic has signed a $3.4 billion deal with OnePWR Solutions to secure 1.7 GW of heavy-duty generators for data-center power, marking the largest known single power-equipment purchase by an AI lab.