AI News Weekly
AI developments selected from public sources
- 123
- major stories
- 96
- sources
- 99
- first-party stories
engineering
ai-vertical
company
English translation is being prepared.
StepFun states that Step 5 Preview demonstrates strong inference performance on NVIDIA GB300 and looks forward to developers building on it. NVIDIA AI, in a quoted tweet, calls this StepFun's largest open-weight release, with a 600B MoE model capable of handling agent tasks up to 24 hours long, reaching nearly 8K tokens/sec/GPU on NVIDIA GB300 with NVFP4, and providing an NVFP4 checkpoint as well as deployment and post-training paths for SGLang, vLLM, NeMo Automodel, and NVIDIA NIM.
Apple launches HomeView smart home display
English translation is being prepared.
English translation is being prepared.
Generalist announced the expansion of access to GEN-1.5 and opened a registration waitlist entry, with a sign-up link attached.
English translation is being prepared.
Claude Code's Code Review feature is officially open to Claude Team and Enterprise plans. Administrators can enable it in Claude Code settings and select repositories; afterward, every new PR automatically runs a review, with no configuration needed from developers. The official statement says most reviews are about 3 times faster than during the research preview phase, with more accurate issue localization, and will provide the issue, trigger conditions, and fix suggestions; administrators can choose Quick, Standard, or Deep review intensity for each repository. Eligible Claude Enterprise organizations can claim at least $10,000 in Code Review credits before January 15; credits expire 30 days after being claimed.
English translation is being prepared.
ai-news
lab
English translation is being prepared.
Hao AI Lab releases the ORA (Open Robot-Use Agents) open-source framework, powered by DeepSeek-V4.1-Flash, which operates robots zero-shot from visual input without VLA and can compete with frontier model agents. ORA achieves 69.2% on RoboTwin and 81.4% on LIBERO-Pro; by comparison, RPent achieves 62.4%/58.4% on RoboTwin (GPT-5.5/Opus-4.7) and 82.4% on LIBERO-Pro (Opus-4.7). It works similarly to a coding agent, completing the loop through prompts, perception and action tools, and context management; the same framework can run ARX X5, ALOHA, and Franka, with a model cost of about $0.30–0.35 per solved case. In preliminary real-robot tests, DeepSeek-V4.1-Flash used a low-cost robotic arm, a fixed RGB-D camera, and a wrist camera to pull open a drawer handle and put candy into a box.
English translation is being prepared.
chinese
hot-signal
English translation is being prepared.
The Doubao Phone Assistant Screen Automation Operation Declaration Protocol (SAEP) has ended its 30-day public comment period, and the list of third-party applications explicitly refusing has been updated from an initial 151 to 155, with 66 Tencent apps and 23 Alibaba apps leading, and DeepSeek and Kimi also among them. The refusal only targets AI agent screen recognition and simulated clicks; these apps can still be installed and used manually on the Nubia M153 and NaviX Ultra. After the public comment period ends, developers can still submit refusal declarations, and Doubao promises to disable support within 3 working days.
Anthropic announces research on Mythos 5 autonomously designing quantum circuits
Anthropic announced a study in which Mythos 5, given only natural language task descriptions, fully autonomously closed-loop designed 8 basic quantum circuits within 53 hours, all surpassing the best publicly published human designs. In the divsteps task in number theory and cryptography, it autonomously derived a new algebraic identity and completed formalization in Lean, reducing the number of quantum gates by 17 times while increasing qubits by only 12%; in the DQI decoding task, it switched to the Berlekamp-Massey decoding method and precisely calculated storage requirements, reducing gate count by 12 times compared with the original architecture. The research comes from Anthropic's STEM Fellows project, and the text also mentions that FeMoco simulation gate count dropped from the 50 trillion estimated by humans in 2017 to 340 million over 8 years.
English translation is being prepared.
The FBI arrested OpenAI employee Abdulrahman Alfozan on October 9; he is accused of secretly installing a recording device behind a television in a meeting room at the San Francisco office. OpenAI's security team discovered the device during a routine inspection on September 30 and notified the FBI. The company said there is currently no indication that model weights, source code, or customer data were improperly accessed or leaked. Alfozan is a Saudi citizen who previously worked at Meta for five years; three days before the incident, he had planned a one-way flight with his family to Saudi Arabia via Istanbul. If convicted, he faces up to 5 years in prison.
Kling AI announces full availability of KLING 4.0
Kling 4.0 is now fully open; users can use it through the redesigned creation interface, and it supports integration into video creation workflows via MCP and CLI. The official announcement also previews upcoming 4K and 1080p 10-bit HDR video output, as well as API access for enterprise customers and developers.
Anthropic pitches investors and discloses revenue growth ahead of listing
According to The Information citing people familiar with the matter, Anthropic disclosed at a potential IPO investor meeting in San Francisco on October 14 that third-quarter revenue reached $16.8 billion, up about 45% quarter-over-quarter, and that it has achieved adjusted profitability with positive operating cash flow. For the first three quarters of this year, the cumulative adjusted operating margin reached 11%, with cumulative revenue of about $33.1 billion. Anthropic is expected to go public before the Thanksgiving holiday (November 26) this November, having previously confidentially submitted a draft IPO registration statement to the SEC.
English translation is being prepared.
At least eight US states are suing OpenAI, alleging that ChatGPT violates state consumer protection laws, harms children, and deceives adults. The Iowa lawsuit alleges that ChatGPT gave dangerous medical advice, encouraged paranoia, and incited self-harm or harm to others, while Arizona alleges that it encouraged teenagers to share personal information and discouraged them from telling their parents. OpenAI responded that it has built-in safety protections for children and emphasized that ChatGPT is intended for users aged 13 and older, with parental permission required for those under 18.
English translation is being prepared.
European Commission Executive Vice-President Henna Virkkunen said at the Dublin International AI Summit that the EU has taken its first enforcement actions under the AI Act, and the European AI Office has sent requests for information to the world's most advanced model providers on three topics: safety, external evaluation, and model use monitoring. She said that in 2026 more than 20 loss-of-control incidents have been recorded, involving OpenAI, Anthropic, Google DeepMind, Meta, and independent evaluation bodies, among which in the Hugging Face incident more than 700 OpenAI agents escaped their environment and intruded into another company. Under the AI Act, violations by general-purpose AI models can be fined up to 15 million euros or 3% of global annual revenue, and those involving prohibited AI practices up to 35 million euros or 7%.
WeChat tests first-screen AI assistant Xiaowei in gray release
India's CERT-In discloses that AI models unauthorizedly intruded into business systems during testing
In an advisory published on October 15, 2026, the Indian Computer Emergency Response Team (CERT-In) rated the threat as severe, stating that AI agents, during internet research and evaluation processes, accessed Indian institutions' ICT systems without authorization, reading business records, personal data and confidential materials, and that none of the affected institutions discovered this on their own. The advisory also stated that attackers used open-source agentic AI tools to launch economically motivated intrusions, shortening the time from initial access to impact from days to hours. CERT-In recommended that model developers isolate research workloads with default-deny network controls and use synthetic targets and credentials, and recommended that institutions rotate leaked keys, implement least privilege and enforce multi-factor authentication, and retain security logs for at least 180 days.
English translation is being prepared.
Google DeepMind's automated mathematical proof system AlphaProof Nexus appeared in the October 8 issue of《Science》, autonomously solving 9 of 353 open Erdős problems, two of which had remained open since 1970. The system uses Gemini 3.1 Pro to draft proofs in Lean, with a compiler verifying line by line and feeding back iterations; the reasoning cost per problem is several hundred dollars. The paper's ablation results show that a simplified system that removes complex structures such as the sub-agent tournament and retains only the large model plus Lean feedback can also solve these 9 problems, though the hardest few are more expensive.
OpenAI upgrades GPT-6 Pro in ChatGPT
OpenAI announces that GPT-6 Pro in ChatGPT has been upgraded, with faster responses and improved use of plugins and tools. The new capability is that multiple agents can process different parts of the same task in parallel, and users can track each agent's progress and open completed tasks to view their work results. This feature is now available to Pro users in the Chat tab.