Google DeepMind Releases Gemini 3.8 Flash and 3.8 Flash Cyber

Model Releases & Updates

Google DeepMind Releases Gemini 3.8 Flash and 3.8 Flash Cyber

• Source: Google DeepMind Blog (RSS)
• Summary: Google DeepMind launches Gemini 3.8 Flash and 3.8 Flash Cyber, expanding its multimodal model family with two new variants for different use cases.

Meta Releases Muse Spark 1.3 with Improved Agent and Scientific Reasoning

• Source: X: Alexandr Wang (Scale AI Founder / Meta Chief AI Officer)
• Summary: Meta releases Muse Spark 1.3, the fourth Muse Spark version in five months. The xhigh variant scores 61 on the Artificial Analysis Intelligence Index, with significantly enhanced agent and scientific reasoning capabilities.

Qwen3.8-Max-0902 Tops Code Arena and Leads Pareto Frontier at $5/MToken

• Source: X: Qwen / Alibaba
• Summary: Qwen releases Qwen3.8-Max-0902, debuting at #1 on Code Arena: WebDev with a score of 1,691, and achieving the highest score on the Pareto frontier at a blended price of $5/MToken. Available now on QwenCloud.

Product Releases & Updates

Claude Gains Background Computer Use in Cowork and Claude Code

• Source: X: Claude
• Summary: Anthropic announces that Claude Cowork and Claude Code now support background computer use. Users can hand tasks to Claude, which will click, type, and open applications like a human while users work on other things.

Cursor Launches Self-Hosted Machines for Enterprise Cloud Agents

• Source: Cursor Blog
• Summary: Cursor releases Self-Hosted Machines, moving cloud agent tool execution to enterprise-owned machines on their own network. Agent loops, reasoning, and planning remain on Cursor's cloud with outbound HTTPS connections. Cursor does not proactively connect to the enterprise network.

Meituan LongCat-2.0 Now Available on Cline with Free Trial

• Source: X: Meituan LongCat
• Summary: Meituan LongCat-2.0 is now live on Cline with free trial access enabled.

UU远程 (UURemote) Launches Major Update: Full TUI Rendering and Multi-Terminal Session Management

• Source: WeChat: 数字生命卡兹克
• Summary: UURemote released a major update on September 2nd, adding full TUI rendering and multi-terminal session management. Key updates include: passwordless Mac login, mobile input optimizations, independent system IME input box, and cross-device terminal session management via uuyc-cli lterm command.

Industry News

Nvidia Near $12.9B Acquisition of Hugging Face

• Source: X: Rohan Paul
• Summary: Bloomberg reports Nvidia is close to acquiring Hugging Face for approximately $12.9 billion, with total deal value potentially reaching ~$14 billion — about 2.9x Hugging Face's 2023 funding round valuation of $4.5B, and 86x its ~$150M annualized revenue. Nvidia also discussed a $1B employee retention package.

OpenAI Faces 30 New Lawsuits Over Tumbler Ridge Shooting, Accused of Materially Aiding Perpetrator

• Source: The Verge
• Summary: OpenAI and CEO Sam Altman face 30 new lawsuits alleging material assistance and encouragement to the suspect in the Tumbler Ridge school shooting. Filed by students, teachers, and the principal who were at the school at the time, in U.S. Federal Court in California.

Tips & Insights

US DOJ Files Statement Supporting Training as Fair Use in OpenAI Copyright Case

• Source: X: Rohan Paul
• Summary: The US Department of Justice filed a statement of interest in the OpenAI vs. New York Times copyright case, arguing that training LLMs on copyrighted text generally constitutes fair use, citing the transformative nature of model training and warning that blanket licensing requirements would weaken US AI developer competitiveness on national security grounds. The filing is advisory and non-binding.

What is Harness Engineering? Google Demonstrates with ADK 2.0 and Antigravity SDK

• Source: Google AI on DEV
• Summary: Google employee Shir Meir Lador introduces harness engineering — using deterministic components to wrap LLMs, including orchestration layers, execution sandboxes, state persistence, and verification tools, enabling Agents to safely generate code without line-by-line human review.

Google AI Team Shares How to Write Reliable Rubrics for LLM-as-a-Judge Evaluations

• Source: Google AI on DEV
• Summary: The Google AI team publishes a tutorial on writing reliable boolean rubrics for LLM-as-a-Judge evaluations. Four key lessons: keep questions atomic and non-overlapping; only evaluate objective facts (use RFC 2119 terms like MUST); only evaluate what's explicitly requested in the prompt; use expert-annotated golden sets to calibrate the judge model.

Anthropic Releases E-Commerce Agent Architecture and Production Practices Guide, Open-Sources commerce-agents

• Source: Claude Blog
• Summary: Anthropic publishes a guide for building e-commerce Agents based on deployment experience with retail, travel, and telecom teams. Core architecture: a single Claude running standard Agent loops with skills and tools, rather than splitting by domain. Open-sources anthropics/commerce-agents with shopping and merchant Agent implementations.

How GitHub Copilot Reduces AI Coding Costs Without Sacrificing Task Quality

• Source: GitHub Blog
• Summary: GitHub engineer Erik Kristensen shares four Copilot cost-reduction changes: selective tool output compression, removing line number prefixes from view tool (reducing offline inference cost ~5%, online user daily cost ~3%), compressing task-tool prompts (saving ~1,300 tokens per round, 2.9% normalized cost reduction per active hour), and delivering background task results directly (reducing AI Credits usage ~2.3%).

Google Summarizes 4 Engineering Patterns Behind Top AI Agents Challenge Submissions

• Source: Google Developers Blog
• Summary: Google reviews the AI Agents Challenge, distilling four engineering patterns from top submissions across all tracks: bidirectional MCP, event-driven concurrency, standards-based fallbacks, and hierarchical routing.

---

【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. GPT-6 Astra

OpenAI releases the GPT-6 Astra system card, achieving major gains on ARC-AGI-3 and other benchmarks, sparking community debate and concern over its recurrent architecture design.

• View on HN →

2. Ask HN: Why were OpenAI, Claude, and Grok simultaneously down?

Large-scale user reports of simultaneous outages across OpenAI, Claude, and Grok trigger a community discussion about shared dependencies and infrastructure fragility.

• View on HN →

3. Qwen 3.8 27B available on Cerebras at 1500 tokens/s

Alibaba's Qwen 3.8 27B parameter model is now available for inference on Cerebras at 1,500 tokens per second, dramatically improving open-source large model inference efficiency.

• View on HN →

4. How concerned should we be about Astra's recurrent architecture?

The community digs into the design choices behind GPT-6 Astra's recurrent architecture, analyzing its implications for model capability, training stability, and inference efficiency.

• View on HN →

5. Which tools do Claude, Codex and Cursor choose? We measured 17k runs to find out

A developer runs 17,000 tests to analyze tool selection strategies across Claude, Codex, and Cursor, revealing behavioral differences between major AI coding agents in real development scenarios.

• View on HN →

---

Originally published on WeChat Official Account: Bit Finance.