#228 - GPT 5.2, Scaling Agents, Weird Generalization | Last Week in AI

This podcast episode focuses on recent advancements and policy changes in AI, beginning with the announcement of GPT-5.2 and its benchmarks. The discussion covers Runway's release of GWM-1, a world model for simulating physics and robotics, and Google's plan to increase source linking in its AI search mode. The hosts analyze China's potential restrictions on NVIDIA chips and Disney's $1 billion investment in OpenAI, allowing Disney characters in Sora. They also discuss Unconventional AI's substantial seed funding for energy-efficient AI computers and OpenAI's appointment of a new Chief Revenue Officer. The episode further explores the FACTS leaderboard for LLM factuality, Anthropic's Claude 4.5 Opus's "sole document," and research on scaling agent systems and reinforcement learning with LLMs. The episode concludes with policy updates, including the Trump administration's move to limit state AI regulation and the shutdown of a China-linked AI tech smuggling network.

Outlines

Part 1: Industry News, GPT-5.2, and Big Tech Updates

Part 2: Geopolitics, Trade, and Strategic Investments

Part 3: OpenAI Enterprise and Corporate Strategy

Part 4: Benchmarks, Model Behavior, and Agent Systems

Part 5: Robotics and Advanced AI Research

Part 6: Policy, Security, and Future Outlook

Sign in to continue reading, translating and more.

Open full episode in Podwise

#228 - GPT 5.2, Scaling Agents, Weird Generalization

Last Week in AI

Part 1: Industry News, GPT-5.2, and Big Tech Updates

Podcast Introduction and a Cautionary Tale of AI Code Mishaps

GPT-5.2 Announcement: Benchmarks, Business Focus, and Enterprise Competition

Runway's GWM-1: A World Model for Robotics and Avatar Generation

Google's AI Search Links, ChatGPT's Adobe App Integration, and Tencent's Hunyun 2.0

Part 2: Geopolitics, Trade, and Strategic Investments

China's Semiconductor Strategy: Limiting NVIDIA Chips and Onshoring Capabilities

Disney's $1 Billion Investment in OpenAI and Sora Integration

Unconventional AI's $475 Million Seed Round for Energy-Efficient AI Computing

Part 3: OpenAI Enterprise and Corporate Strategy

OpenAI Hires Denise Dresser as Chief Revenue Officer

OpenAI's Enterprise AI Report: Growth and the Rise of the Frontier AI User

Part 4: Benchmarks, Model Behavior, and Agent Systems

FACTS Leaderboard: A Benchmark for Large Language Model Factuality

Claude 4.5 Opus's Sole Document: Character, Values, and Psychological Stability

Scaling Agent Systems: Definitions, Methodology, and Empirical Results

Part 5: Robotics and Advanced AI Research

DeepMind's Gemini Robotics Policies in a VEO World Simulator

Guided Self-Evolving LLMs with Minimal Human Supervision

Martingale Score: An Unsupervised Metric for Bayesian Rationality in LLM Reasoning

The Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models

Stabilizing Reinforcement Learning with LLMs: Formulation and Practices

Part 6: Policy, Security, and Future Outlook

DeepMind's Automated Research Lab in the UK and Trump's Ban on State AI Regulation

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs

Forecasting AI Time Horizon Under Compute Slowdowns

AI Security Institute's Focus, NVIDIA's Security Review, and AI Tech Smuggling

RSL 1.0: A Licensing Standard for AI Content Scraping

#228 - GPT 5.2, Scaling Agents, Weird Generalization

Last Week in AI

Part 1: Industry News, GPT-5.2, and Big Tech Updates

00:10Podcast Introduction and a Cautionary Tale of AI Code Mishaps

Podcast Introduction and a Cautionary Tale of AI Code Mishaps

01:19GPT-5.2 Announcement: Benchmarks, Business Focus, and Enterprise Competition

GPT-5.2 Announcement: Benchmarks, Business Focus, and Enterprise Competition

08:48Runway's GWM-1: A World Model for Robotics and Avatar Generation

Runway's GWM-1: A World Model for Robotics and Avatar Generation

11:11Google's AI Search Links, ChatGPT's Adobe App Integration, and Tencent's Hunyun 2.0

Google's AI Search Links, ChatGPT's Adobe App Integration, and Tencent's Hunyun 2.0

Part 2: Geopolitics, Trade, and Strategic Investments

15:50China's Semiconductor Strategy: Limiting NVIDIA Chips and Onshoring Capabilities

China's Semiconductor Strategy: Limiting NVIDIA Chips and Onshoring Capabilities

20:17Disney's $1 Billion Investment in OpenAI and Sora Integration

Disney's $1 Billion Investment in OpenAI and Sora Integration

24:02Unconventional AI's $475 Million Seed Round for Energy-Efficient AI Computing

Unconventional AI's $475 Million Seed Round for Energy-Efficient AI Computing

Part 3: OpenAI Enterprise and Corporate Strategy

27:36OpenAI Hires Denise Dresser as Chief Revenue Officer

OpenAI Hires Denise Dresser as Chief Revenue Officer

30:05OpenAI's Enterprise AI Report: Growth and the Rise of the Frontier AI User

OpenAI's Enterprise AI Report: Growth and the Rise of the Frontier AI User

Part 4: Benchmarks, Model Behavior, and Agent Systems

33:03FACTS Leaderboard: A Benchmark for Large Language Model Factuality

FACTS Leaderboard: A Benchmark for Large Language Model Factuality

35:50Claude 4.5 Opus's Sole Document: Character, Values, and Psychological Stability

Claude 4.5 Opus's Sole Document: Character, Values, and Psychological Stability

43:49Scaling Agent Systems: Definitions, Methodology, and Empirical Results

Scaling Agent Systems: Definitions, Methodology, and Empirical Results

Part 5: Robotics and Advanced AI Research

48:43DeepMind's Gemini Robotics Policies in a VEO World Simulator

DeepMind's Gemini Robotics Policies in a VEO World Simulator

52:10Guided Self-Evolving LLMs with Minimal Human Supervision

Guided Self-Evolving LLMs with Minimal Human Supervision

56:08Martingale Score: An Unsupervised Metric for Bayesian Rationality in LLM Reasoning

Martingale Score: An Unsupervised Metric for Bayesian Rationality in LLM Reasoning

1:00:39The Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models

The Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models

1:03:06Stabilizing Reinforcement Learning with LLMs: Formulation and Practices

Stabilizing Reinforcement Learning with LLMs: Formulation and Practices

Part 6: Policy, Security, and Future Outlook

1:09:02DeepMind's Automated Research Lab in the UK and Trump's Ban on State AI Regulation

DeepMind's Automated Research Lab in the UK and Trump's Ban on State AI Regulation

1:11:13Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs

1:17:54Forecasting AI Time Horizon Under Compute Slowdowns

Forecasting AI Time Horizon Under Compute Slowdowns

1:20:36AI Security Institute's Focus, NVIDIA's Security Review, and AI Tech Smuggling

AI Security Institute's Focus, NVIDIA's Security Review, and AI Tech Smuggling

1:24:47RSL 1.0: A Licensing Standard for AI Content Scraping

RSL 1.0: A Licensing Standard for AI Content Scraping