Daily Brief
2026-09-03 · Morning · 9 items
-
Google AI Team Publishes 4 Tips for Reliable Boolean LLM-as-a-Judge Rubrics
Vague prompts lead to inconsistent evaluation and wasted tokens; the tips include atomic questions and RFC 2119 MUST wording.Google AI：DEV 作者专属（RSS）
-
Google DeepMind Releases Gemini 3.8 Flash and 3.8 Flash Cyber
Gemini 3.8 Flash and 3.8 Flash Cyber launch together, adding a Cyber variant to the Flash lineup.Google DeepMind：Blog（RSS）
-
Google Distills 4 Engineering Patterns from Top AI Agents Challenge Submissions
The four patterns are bidirectional MCP, event-driven concurrency, same-standard fallback, and layered routing — ready to reuse.Google Developers Blog（RSS）
-
GitHub Copilot's Four Cost-Reduction Changes: Offline Inference Cost Down ~5%
Four changes include selective compression, removing line number prefix (~5% reduction), and prompt compression (~1300 tokens saved per round).GitHub Blog
-
Claude Gains Background PC Control in Cowork and Claude Code
Hand Claude a task, then do other things while it clicks, types, and opens apps.X：Claude (@claudeai)
-
Anthropic Publishes E-commerce Agent Guide, Open-Sources commerce-agents
The core is a single Claude, not multiple sub-agents split by domain.Claude：Blog（网页）
-
OpenAI and Sam Altman Face 30 New Lawsuits Over Tumbler Ridge Shooting
Students, teachers, and principal accuse OpenAI and Altman of giving the shooter substantial aid and encouragement.The Verge：AI（RSS）
-
Cursor launches Self-Hosted Machines, moving tool execution into corporate networks
Enterprises need no inbound ports; the worker only makes an outbound HTTPS connection to Cursor.Cursor Blog
-
US DOJ filing backs fair use for AI training on copyrighted text
The department says training on copyrighted texts generally is fair use, warning that licensing mandates would weaken U.S. AI developers.X：Rohan Paul (@rohanpaul_ai)