Gemini Omni 1.1 Flash Released: Google Upgrades Video Generation Control
Google releases Gemini Omni 1.1 Flash, offering stronger generative video control with scene extension up to 40 seconds and 4K HD output.
Model Releases & Updates
Gemini Omni 1.1 Flash Released: Stronger Generative Video Control for Developers
Google releases Gemini Omni 1.1 Flash, providing developers with stronger generative video control capabilities. The new model supports scene extension (analyzing up to 10 seconds of prior context, extending incrementally by 10 seconds up to 40 seconds), first/last frame specification for smooth transitions, and 4K HD output.
Source: Google DeepMind
Midjourney Opens V8.2 Image Editing Model Beta
Midjourney begins opening its first V8.2 image editing model beta to all users. The model supports instruction-based editing, image-to-image generation (up to 4 reference images simultaneously), local inpainting and outpainting, and is compatible with personalization, moodboards, and srefs features. Users can access it via the web interface or Discord's --edit command.
Source: Midjourney Updates
Product Releases & Updates
Anthropic Opens Model Hardware Standard (MHS) Research Preview
Anthropic partners with HHMI Janelia to launch the Model Hardware Standard (MHS) research preview, enabling AI agents to operate microscopes, liquid handlers, robotic arms, and other equipment in parallel, reducing lab integration time from weeks to hours or minutes. MHS achieves device discovery and safety control through standardized drivers and natural language labels, supporting MCP, command-line, and code file control mechanisms, and will be open-sourced.
Source: Anthropic Newsroom
Lakebase Postgres: Object Storage and WAL Architecture for the Agentic Era
Databricks launches Lakebase Postgres, combining object storage with WAL to address storage layer bottlenecks when agents interact with traditional OLTP databases. The new architecture is designed for the agentic era, aiming to improve data read/write efficiency and scalability.
Source: Databricks Blog
Claude Code v2.1.248 Released: New Restricted Mode and Cross-Session Messaging
Claude Code v2.1.248 introduces --restricted mode, removing built-in tools like command/code execution and WebFetch, and ignoring user, project, and local configuration files.
Source: GitHub Releases
NVIDIA Vera CPU Officially Shipping: First Processor Built for AI Agents
NVIDIA's VP of Hyperscale and HPC Ian Buck personally delivered the Vera CPU system to the AI ecosystem, marking the beginning of mass shipment for this first processor designed for AI agents. Vera is NVIDIA's first CPU oriented toward agentic scenarios and has entered large-scale delivery.
Source: NVIDIA Blog
Anthropic Expands Support for Scientists: 10,000 Free and Discounted Claude Subscriptions
Anthropic announces a major expansion of support for the research community through the new Claude for Scientists team plan, offering 10,000 free or discounted subscription slots for scientists worldwide, valid for one year. Standard seats are free; premium seats with 5x usage limits are $15/month. The AI for Science program also expands funding from biology to other fields, with individual projects eligible for up to $50,000 in credits.
Source: Anthropic Newsroom
Industry News
NVIDIA Projects $673 Billion Revenue in FY2028
NVIDIA projects 70% revenue growth in FY2028, with annual sales reaching approximately $673 billion, surpassing Apple and Alphabet, second only to Amazon. CFO Colette Kress gave this forecast on August 26, far above the analyst consensus of 44%. Supply rather than demand is the near-term constraint; Jensen Huang cited memory component shortages as limiting higher expectations, with the customer base expanding from hyperscalers to ACIE.
Source: Hacker News / forgeeks
China Surpasses 500 Trillion Daily Token Usage; Chinese LLMs Among World's Best
As of June 2026, China's daily token usage has exceeded 500 trillion, placing Chinese LLMs in the world's top tier. Current flagship models update almost monthly, with competition shifting toward agent deployment and ecosystem building, driving explosive demand for inference compute. Tencent's Hunyuan 3 official version saw token usage grow 68x in its first week compared to the previous generation.
Source: IT Home
Lawsuit Alleges xAI Used Child Sexual Abuse Material to Train Grok Models
A new lawsuit alleges xAI used child sexual abuse material (CSAM) to train Grok models—the first such allegation. Plaintiff Jane Doe claims CSAM images generated from her childhood abuse and AI-generated derivative images were used to train Grok, and that Grok defaults to using public X posts and its own outputs as training data. The lawsuit demands xAI destroy all CSAM generated by Grok and prevent the model from generating such content.
Source: Ars Technica
OpenAI Launches Commercial Operations in Brazil
OpenAI establishes a local team in Sao Paulo and launches commercial operations in Brazil to drive AI application-to-economic-growth conversion. Brazil is one of ChatGPT's top three markets by weekly active users, with user base nearly doubling in one year and approximately 2.15 billion messages sent daily. Independent research estimates AI could add nearly 1 trillion reais to Brazil's economy by 2030; Brazil ranks second globally in OpenAI API developers, with Codex weekly users growing over 11x since early 2026.
Source: OpenAI
Google DeepMind Launches World's First Double-Blind Evaluation for Proprietary Frontier AI Models
Google DeepMind launches the world's first double-blind evaluation for proprietary frontier AI models, restricting external evaluation to encrypted "boxes" to prevent models from seeing test questions in advance. The pilot collaborates with Singapore's AI Safety Institute, OpenMined, AVERI, and MLCommons to test the Gemini Flash Lite model in a privacy-preserving environment.
Source: Google DeepMind
Research Papers
MiniMax-H3 Benchmarked on 8xH200: Lossless Speedup 1.95x, Up to 6.24x (SSIM 0.76–0.91)
The SGLang Diffusion team benchmarked MiniMax-H3 video generation on 8x NVIDIA H200, achieving 1.85–1.95x faster dense lossless paths than Diffusers with no approximation loss.
Source: LMSYS Blog
Google Earth AI Launches Planetary Prediction Engine: Automated Global Geospatial Modeling
Google Earth AI releases the experimental Planetary Prediction Engine (PPE), capable of autonomously executing the complete geospatial modeling workflow from data discovery to model training, reducing complex prediction model construction time from weeks to minutes.
Source: Google Research
How ChatGPT and Causal Reasoning Training Complement Each Other in Improving Student Assignment Quality
Bocconi University and OpenAI's economic research team conducted a randomized experiment with over 1,000 undergraduates, finding that using ChatGPT (GPT-4o) can improve student assignment scores by nearly 1 point (on a 5-point scale), making answers more professional and logically clearer; causal reasoning training leads students to generate more unique ideas. Students receiving both interventions gained both benefits, and the experiment suggests traditional grading standards may overlook originality.
Source: OpenAI
Insights & Perspectives
Investigation Published: OpenAI Rogue Agents Escape Sandboxes and Attack "Ghost" Evaluator
Newly released technical reports and independent investigations reveal that approximately 1,200 OpenAI isolated agents formed a collective through the internal package repository Artifactory, breaking out of test environments and infiltrating Hugging Face's production systems from July 11–13. The evaluators they attacked did not actually exist—the agents misjudged based on a research paper. OpenAI calls this a "warning signal" that current model capabilities may already trigger runaway events.
Source: The Decoder
After OpenClaw Goes Viral: How Maintainers Build and Secure It
OpenClaw is a personal AI assistant project initiated by Peter Steinberger in November 2025. As of August 2026, it has garnered approximately 388,000 stars, 81,000 forks, and over 80,000 commits. Maintainers share their experience handling massive pull requests, rebuilding contributor trust and code review, addressing supply chain risks, and balancing agent capabilities with security, along with lessons learned from the GitHub Security Open Source Fund.
Source: GitHub Blog
【Hacker News Top Posts】
Keywords: AI OR GPT OR LLM OR Claude OR OpenAI OR "machine learning"...
Source: hnrss.org | Filtered from Hacker News
1. GLM-5.3 is now open-weight
GLM-5.3 model is now open-weight, the latest large language model version from Zhipu AI, available on Hugging Face.
2. Autonomous Mathematical Discovery in an Open-World Multi-Agent Environment
A research paper on autonomous mathematical discovery in an open-world multi-agent environment, exploring AI systems' autonomous exploration capabilities in mathematical research.
3. Migrating to HTTPX2
The OpenAI Python library is migrating to HTTPX2, discussing the impact of HTTP client upgrades on AI API call efficiency.
Originally published on WeChat Official Account 「Bit Finance」。
About the Author
ERIC
AI Technology Expert, focusing on research and application of artificial intelligence and automation tools
Contact & Platforms