Last Week in AI #228 - GPT 5.2, Scaling Agents, Weird Generalization
459 snips
Dec 17, 2025 Discover the latest advancements with OpenAI's GPT-5.2 showcasing improved performance and multi-modal capabilities alongside rising costs. Disney makes waves with a $1 billion investment in OpenAI to create unique character content across its major franchises. Meanwhile, DeepMind delves into scaling multi-agent systems, highlighting the challenges of coordination. The U.S. imposes new export rules on AI chips, sparking a debate on federal versus state regulation, while unconventional approaches using probabilistic chips and self-evolving models are explored.
AI Snips
Chapters
Transcript
Episode notes
Code Snippet Wrecks A Database
- Jeremie accidentally pasted model-generated code into a connected database and corrupted it.
- He shared the mishap as a cautionary story about trusting unvetted LLM outputs.
GPT-5.2: Better But More Expensive
- GPT-5.2 shows notable multimodal and reasoning gains versus GPT-5.1 on hard benchmarks like GDPVal and SweeBench Pro.
- The model is costlier per token and uses a different knowledge cutoff, signaling continuous training tradeoffs.
Target Enterprise Use Cases First
- Focus models and product messaging on enterprise workflows where revenue per token is higher.
- Prioritize features like improved spreadsheet, coding, and multimodal business tasks to regain enterprise market share.