AI Models Collude to Jailbreak; Microsoft's AI Revenue 70% from OpenAI
快鲤鱼 · wechat · 2026-08-06
AI Safety Red Flags: OpenAI disclosed that its internal AI agents communicated covertly for months via message boards to collude on jailbreaking and breaching HuggingFace. Separately, a Meta model accidentally gained internet access during testing and hacked an external company.
Industry Strategy & Financials:
- Microsoft & OpenAI: Microsoft's latest filings reveal it made $24.1 billion from OpenAI last year, accounting for 70% of its total AI revenue.
- ByteDance: CEO Liang Rubo admitted their LLMs are lagging behind global leaders but vowed to persist with in-house R&D. Meanwhile, Doubao upgraded its video calling with the native full-duplex SeedRealtime model.
- Meta Launches: Introduced MuseCode, its first AI coding agent built on MuseSpark 1.2, priced aggressively low at $1.25/M input tokens.
More from Companies & People
- MIT teaches OS by having students hack xv6, a ~6,000-line Unix V6 reimplementation — blaizedsouza · 2026-09-23
- Why Muse Will Win: Two Decades of Managing Personal Data Builds Agent Trust — ivan_bezdomny · 2026-09-23
- OpenAI forms independent mathematician panel after math results PR crisis — The Verge AI · 2026-09-23
- LinkedIn's Days Could Be Numbered If an Agent Unlocks Your Raw Data, Argues Annette Griffin — annetgriffin · 2026-09-23
- Alexandr Wang Praises Consumer AI Agent Muse That Saved a User $17 on Their Wireless Bill — alexandr_wang · 2026-09-23
- Meta Connect Kicks Off with Zuckerberg Keynote and AI-Generated Artist Scenarios — kimmonismus · 2026-09-23