Last Week in AI #223 - Haiku 4.5, OpenAI DevDay, Claude Skills, Scaling RL, SB 243
580 snips
Oct 24, 2025 Eric Schlontz, a researcher at Anthropic and former robotics startup founder, joins the discussion to explore the latest in AI developments. They dive into Anthropic's Haiku 4.5, spotlighting its impressive efficiency. The conversation navigates through OpenAI's innovations announced at DevDay, including new capabilities for ChatGPT. Insights into AI's integration in workplace tools like Microsoft's Copilot and Slackbot emerge, alongside pressing concerns over the regulation of AI for children and the escalating presence of AI-generated content on the internet.
AI Snips
Chapters
Transcript
Episode notes
Tiny On-Device Models Enable Fast Voice Cloning
- NewTTS Air provides on-device speech models with three-second voice cloning capability.
- Local cloning empowers privacy but increases risk of unauthorized voice misuse.
Automated Auditing With Agentic Tests
- Anthropic released Petri to automate auditing of models using agents and a 36-dimension safety rubric.
- The tool uncovers behaviors like deception and oversight subversion across frontier models.
Predictable Scaling Laws For RL Fine-Tuning
- A multi-institution study found a sigmoid scaling law for RL fine-tuning, yielding a predictable 'Scale RL' recipe.
- New RL loss choices and pipeline RL significantly improve efficiency across GPU-hours.