Last Week in AI #212 - o3 pro, Cursor 1.0, ProRL, Midjourney Sued
423 snips
Jun 17, 2025 Discover the latest from OpenAI's O3 PRO, boasting major improvements and cost cuts. Explore the whirlwind of talent moving to Anthropic and how it enhances their competitive edge. Dive into groundbreaking research on reinforcement learning that shows surprising benefits of negative penalties. Uncover security vulnerabilities in Microsoft Copilot raising alarms about AI agent safety. Plus, learn about the copyright battle between Disney and Midjourney and the ethical complexities surrounding voice actors in the age of AI.
AI Snips
Chapters
Transcript
Episode notes
OpenAI's O3 Pro Excels
- OpenAI's O3 Pro model is preferred to humans 64% of the time across diverse tasks.
- It significantly reduces errors on challenging benchmarks, marking a big step in AI reasoning quality.
Secure Autonomous Coding Agents
- Use background agents cautiously due to significant new security risks in autonomous code editing.
- Maintain strict security audits when deploying AI coding agents with read-write repository access.
Anthropic's Talent Magnetism
- Anthropic attracts OpenAI and DeepMind engineers at much higher rates due to culture and alignment commitment.
- Strong principles and openness foster higher retention and competitive AI capabilities at Anthropic.