# Mira Murati's 975B Open Model, Ramin Hasani on Post-Transformer AI, and Demis' AI FINRA \| EP \#271

- Channel: [Peter H. Diamandis](https://www.youtube.com/channel/UCvxm0qTrGN_1LMYgUaftWyQ)
- Published: 2026-07-17
- Source video: [Watch on YouTube](https://youtube.com/watch?v=bAoXVyibE6Q)
- Read on ReadTube: https://www.read.tube/p/videos/bAoXVyibE6Q
- Language: English

## Summary

**AI Giants Push for Rules**

Top AI leaders are openly calling for regulation, but the episode argues that their proposals could harden barriers for smaller labs and open-model competition. The discussion spans Demis Hassabis’s FINRA-style standards body, the White House’s reported capability ceiling tied to Chinese open models, Mira Murati’s new open model focused on customization, and the idea that recursive self-improvement is already beginning in practical systems.

The episode centers on a series of fast-moving AI governance and capability stories. Sam Altman, Elon Musk, and Demis Hassabis are all portrayed as pushing for stronger oversight, including a FINRA-like standards body for frontier models. The hosts agree that some form of oversight is likely, but they warn that static regulation could become obsolete quickly, create regulatory capture, and favor incumbents over open-weight labs and new entrants.

That concern carries into the report that the White House may use China’s best open-weight model as a ceiling for what U.S. models can openly release. The panel sees this as strategically perverse because it could reward China’s pace, discourage Western open releases, and incentivize talent flight or benchmark manipulation. They argue that the bigger issue is not just model release policy, but how to define capabilities in a way that does not freeze the entire field into a distorted standard.

A major bright spot is Mira Murati’s Thinking Machines release of Inkling, a 975B-parameter open-weight model designed for customization rather than leaderboard domination. The guests frame it as a Western answer to Chinese open models and a sign that enterprise fine-tuning, on-prem deployment, and multimodal customization may be the real commercial path. They connect this to broader trends: open models are becoming more useful for private data, and business value may now come from adaptation rather than raw model access.

The episode then turns to recursive self-improvement, after a startup called Wo AI claimed early evidence that an outer AI loop can improve an inner AI loop. Alex treats it as an important but not first-in-kind result, emphasizing defensive co-scaling, while Ramine Hassani argues that true recursive self-improvement would require changes to weights, architectures, and training processes, not just code or prompt-level automation. The consensus is that the work is exciting but still far from full autonomous self-improving AI.

Other notable threads include the rise of AI digital twins for heads of state and companies, with Malaysia’s prime minister preparing an AI likeness for public outreach, and the broader idea that AI avatars could become a new interface for democracy, corporate communication, and even religious institutions. The episode also covers a patent-system debate, with Palmer Luckey arguing patents are now a national-security liability, though the panel largely rejects expanding secrecy-based patenting. Finally, the show highlights several abundance stories: AI medical models outperforming doctors, and a new aging-reversal enzyme from Revel Pharmaceuticals that may make a key class of protein damage reversible for the first time.