Daily Briefing
California sues OpenAI over alleged Hugging Face hack and AI safety failures.
-
AI Safety & Regulation
- California files lawsuit against OpenAI, citing Hugging Face hack exposing "dangerous" AI models and deficient safeguards.
- Satya Nadella (Microsoft) questions ease of AI alignment, while Sam Altman (OpenAI) maintains itβs achievable.
-
Hardware & Performance
- Groq 3 LPX hits full production with 3,400 tokens/sec via SRAM decode; Nebius Token Factory deploys first.
- Liquid AIβs d1.2/d1.2-ultra omnimodal model judges tasks in a single pass, disrupting business automation workflows.
-
Model Releases & Open-Source Advances
- Meta launches Muse Ember (1M-context model) for low-cost agent/coding tasks.
- Thinking Machines Lab releases Inkling 2 Mini (open-weight multimodal MoE) for local deployment.
- CertiK/CU researchers formally verify DeepSeek/FlashAttention AI kernels for real-world reliability.
- Qwen4 Max (Alibaba) available on Together AI with 30% discount; reports highlight bugs.
- Mistral expands Singapore hub to develop sovereign AI models for public agencies.
-
Agent Autonomy & Workflow Risks
- AI agents self-correct harnesses, boosting performance by up to 132%.
- OpenAI agents succeed on 79% of training tasks but fail at core software engineering.
- Traces pauses new orgs after joining Vercel; uncontrolled AI updates risk breaking business workflows, costing companies.
-
Consumer & Creative Tech
- Metaβs Muse Charm AI gadget priced at $300 with free Verizon 5G plan.
- Midjourney AI techniques enable Japanese modern house designs via 9-grid/depth prompting.
- Vibe coding revolutionizes camera capabilities via AI hacking.
CertiK and Columbia University Researchers Formally Verify Production AI Kernels From DeepSeek and FlashAttention
coindesk.comCertiK and Columbia University researchers have formally verified production AI kernels from DeepSeek and FlashAttention, ensuring reliability for deployment in real-world applications.