2026-09-03 · ThursdayPublished daily at 08:00
AI Daily
- 1Google DeepMind releases two new models, Gemini 3.8 Flash and 3.8 Flash CyberGoogle DeepMind:Blog
- 2
- 3Qwen3.8-Max-0902 Tops Code Arena and Leads the Pareto Frontier at $5/MTokenTongyi Qianwen / Qwen
01Model Releases/Updates
Google DeepMind releases two new models, Gemini 3.8 Flash and 3.8 Flash Cyber
Google DeepMind releases Gemini 3.8 Flash and 3.8 Flash Cyber.
Original textMeta releases Muse Spark 1.3, improving agent and scientific reasoning capabilities
Original textQwen3.8-Max-0902 Tops Code Arena and Leads the Pareto Frontier at $5/MToken
Qwen releases Qwen3.8-Max-0902, which debuts at number one overall on Code Arena: WebDev with a score of 1,691. With a blended price of $5/MToken, it is the highest-scoring model on the Pareto frontier and is now available to try on QwenCloud.
Original text02Product Releases/Updates
Claude supports operating computers in the background in Cowork and Claude Code
Claude officially announced that Claude Cowork and Claude Code now support using a computer in the background. After users assign Claude a task, it can click, type, and open applications like a person, allowing users to do other things simultaneously.
Original textCursor introduces Self-Hosted Machines, allowing cloud agents to execute on companies' own machines
Cursor releases Self-Hosted Machines, moving cloud agents' tool execution to machines within enterprises' own networks. The agent loop, reasoning, and planning remain in Cursor's cloud, connected through outbound HTTPS connections from workers. Cursor does not initiate connections into enterprise networks.
Original textUU Remote Releases New Version: Full TUI Rendering and Multi-Terminal Session Management Enhance the Remote Vibe Coding Experience
UU Remote released a new version on September 2, focusing on terminal features and adding TUI rendering and interaction, as well as terminal session management. Key updates include password-free login on Mac, improved mobile input with a dedicated input box that invokes the system keyboard, and the ability to create and manage multiple terminal sessions simultaneously, with cross-device syncing and handoff between phones and computers (using the uuyc-cli lterm command).
Original text03Industry News
Nvidia nears $12.9 billion acquisition of Hugging Face
Bloomberg reports that Nvidia is nearing a deal to acquire Hugging Face for approximately $12.9 billion, with the total transaction potentially reaching approximately $14 billion. The two sides have yet to reach a definitive agreement, and the timing and details may still change. The price is approximately 2.9 times Hugging Face's $4.5 billion valuation in its 2023 funding round and represents 86 times its annualized revenue of approximately $150 million. Nvidia has also discussed including a $1 billion employee retention package in the deal.
Original textOpenAI faces 30 new lawsuits over the Tumbler Ridge shooting, accused of aiding and abetting
OpenAI and CEO Sam Altman Face 30 New Lawsuits Alleging They Provided Material Assistance and Encouragement to the Suspect in the Tumbler Ridge, Canada, School Shooting. The lawsuits were filed in a California federal court by students, teachers, and the principal who were at the school at the time of the incident.
Original text04Tips and perspectives
The U.S. Department of Justice files a brief in support of treating training as fair use in the OpenAI copyright lawsuit
The U.S. Department of Justice has filed a brief in the copyright litigation between OpenAI and The New York Times, arguing that training LLMs on copyrighted text should generally qualify as fair use. It describes model training as extraordinarily transformative and warns on national security grounds that blanket licensing requirements would undermine the competitiveness of U.S. AI developers. The brief is advisory and does not bind the court. It leaves the methods used to obtain data and whether specific outputs reproduce protected passages as separate issues for courts to assess case by case.
Original textWhat Is Harness Engineering? Google Demonstrates Automated Coding-Loop Repair with ADK 2.0 and the Antigravity SDK
Google employee Shir Meir Lador introduces harness engineering, which wraps LLMs in deterministic components, including an orchestration layer, execution sandbox, state persistence, and verification tools, enabling Agents to safely generate code without requiring manual line-by-line review.
Original textGoogle AI team shares how to write reliable scoring rubrics for LLM-as-a-Judge evaluations
The Google AI team published a tutorial explaining how to write reliable Boolean grading criteria for LLM-as-a-Judge evaluations, noting that vague prompts lead to inconsistent assessments and wasted tokens. It offers four guidelines: keep questions atomic and non-overlapping; have the judge model evaluate only objective facts, using RFC 2119 terms such as MUST where appropriate; evaluate only what is explicitly required in the prompt; and calibrate the judge model against an expert-labeled golden set until it agrees with human ratings.
Original textAnthropic Releases a Guide to E-Commerce Agent Architecture and Production Practices, and Open-Sources the commerce-agents Reference Implementation
Anthropic releases a guide to building e-commerce Agents, based on implementation experience with retail, travel, telecommunications, and other teams. The core architecture uses a single Claude working with skills and tools in a standard Agent loop, rather than splitting sub-agents by domain. Anthropic has also open-sourced the anthropics/commerce-agents reference implementation, which includes shopping and merchant Agents.
Original textHow GitHub Copilot reduces AI coding costs without sacrificing task quality
GitHub engineer Erik Kristensen shared four changes that reduced Copilot costs: selectively compressing tool output; removing line-number prefixes from the view tool (reducing offline inference costs by about 5% and average daily online inference costs per user by about 3%); compressing task-tool prompts (saving about 1,300 tokens per round and reducing normalized cost per active hour by 2.9%); and delivering results directly after background tasks finish (reducing AI Credits usage by about 2.3%).
Original textGoogle Summarizes 4 Engineering Patterns Behind the Strongest Submissions in the AI Agents Challenge
Google reviews the AI Agents Challenge and distills four engineering patterns from the top submissions in each track: bidirectional MCP, event-driven concurrency, standards-based fallback, and hierarchical routing.
Original text