Stabilizing RL for LLMs: Objectives & Conditions

Our 228th episode with a summary and discussion of last week's big AI news!

Recorded on 12/12/2025

Hosted by Andrey Kurenkov and Jeremie Harris

Feel free to email us your questions and feedback at contact@lastweekinai.com and/or hello@gladstone.ai

Read out our text newsletter and comment on the podcast at https://lastweekin.ai/

In this episode:

• OpenAI's latest model GPT-5.2 demonstrates improved performance and enhanced multi-modal capabilities but comes with increased costs and a different knowledge cutoff date.
• Disney invests $1 billion in OpenAI to generate Disney character content, creating unique licensing agreements across characters from Marvel, Pixar, and Star Wars franchises.
• The U.S. government imposes new AI chip export rules involving security reviews, while simultaneously moving to prevent states from independently regulating AI.
• DeepMind releases a paper outlining the challenges and findings in scaling multi-agent systems, highlighting the complexities of tool coordination and task performance.

Timestamps:

• (00:00:00) Intro / Banter
• (00:01:19) News Preview
• Tools & Apps
• (00:01:58) GPT-5.2 is OpenAI’s latest move in the agentic AI battle | The Verge
• (00:08:48) Runway releases its first world model, adds native audio to latest video model | TechCrunch
• (00:11:51) Google says it will link to more sources in AI Mode | The Verge
• (00:12:24) ChatGPT can now use Adobe apps to edit your photos and PDFs for free | The Verge
• (00:13:05) Tencent releases Hunyuan 2.0 with 406B parameters
• Applications & Business
• (00:16:15) China set to limit access to Nvidia’s H200 chips despite Trump export approval
• (00:21:02) Disney investing $1 billion in OpenAI, will allow characters on Sora
• (00:24:48) Unconventional AI confirms its massive $475M seed round
• (00:29:06) Slack CEO Denise Dresser to join OpenAI as chief revenue officer | TechCrunch
• (00:31:18) The state of enterprise AI
• Projects & Open Source
• (00:33:49) [2512.10791] The FACTS Leaderboard: A Comprehensive Benchmark for Large Language Model Factuality
• (00:36:27) Claude 4.5 Opus' Soul Document
• Research & Advancements
• (00:43:49) [2512.08296] Towards a Science of Scaling Agent Systems
• (00:48:43) Evaluating Gemini Robotics Policies in a Veo World Simulator
• (00:52:10) Guided Self-Evolving LLMs with Minimal Human Supervision
• (00:56:08) Martingale Score: An Unsupervised Metric for Bayesian Rationality in LLM Reasoning
• (01:00:39) [2512.07783] On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models
• (01:04:42) Stabilizing Reinforcement Learning with LLMs: Formulation and Practices
• (01:09:42) Google’s AI unit DeepMind announces UK 'automated research lab'
• Policy & Safety
• (01:10:28) Trump Moves to Stop States From Regulating AI With a New Executive Order - The New York Times
• (01:13:54) [2512.09742] Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs
• (01:17:57) Forecasting AI Time Horizon Under Compute Slowdowns
• (01:20:46) AI Security Institute focuses on AI measurements and evaluations
• (01:21:16) Nvidia AI Chips to Undergo Unusual U.S. Security Review Before Export to China
• (01:22:01) U.S. Authorities Shut Down Major China-Linked AI Tech Smuggling Network
• Synthetic Media & Art
• (01:24:01) RSL 1.0 has arrived, allowing publishers to ask AI companies pay to scrape content | The Verge

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.