OpenAI’s GPT-6 Astra, Meta Muse & ByteDance Enters the World Model Race | Lightwork

Claire Zau and Josh Machiz are joined by Lightspeed partner Bucky Moore for a week about how much control we are handing over to AI. OpenAI released GPT-6 Astra, its most capable model and, by OpenAI's own account, a harder one to monitor. Bucky explains why looping inference compresses the chain of thought that used to make these models readable, and why the near term answer may be tighter permissions and access controls rather than interpretability research.

Then to the agents people are actually running. Bucky has given Instinct his personal email and credit card and let it book flights and hotels, and he argues the ChatGPT moment for personal assistants came from a startup rather than an incumbent. Meta launched Muse the same week. Claire and Josh weigh where they still want a person reviewing the work, and the group gets into what the web does when booking sites cannot tell an agent from a customer. Bloomberg also reported that ByteDance is building a real time world model, and Bucky explains why serving those models is a hard problem of its own. Reactor is a Lightspeed portfolio company. Lightspeed co-led the seed round and led the Series A.

Chapters

0:00 Welcome, Bucky Moore!

1:12 OpenAI's Astra

3:50 The New AI Model’s Architecture

4:40 AI Safety and Guardrails

8:01 Trusting AI Agents

11:09 The Future of AI Models

12:06 The AI Agent Land Grab

17:09 AI, Trust, and Personal Data

20:12 Improving AI Agent Capabilities

22:11 Enterprise AI Assistants and Personal Context

24:05 AI Agents and the Future of the Web

25:50 ByteDance and World Models

28:19 World Models and Consumer Content

30:56 Reactor and World Model Infrastructure

34:13 Real-Time Video Models and Scaling