Peter H. Diamandis
July 17, 2026
AI Regulation, Open Models, and Recursive Self-Improvement
TL;DR
CEOs of major AI labs push for regulation modeled on FINRA while Mira Murati launches Inkling, an open-weight model emphasizing customization, and researchers demonstrate early recursive self-improvement breakthroughs.
Key Takeaways
- 1Incumbent AI labs are advocating for FINRA-style regulation, raising concerns about regulatory capture and barriers to entry for smaller labs
- 2Mira Murati's Inkling challenges the leaderboard-dominance approach by focusing on customization and fine-tuning capabilities for enterprises
- 3Recursive self-improvement is advancing with systems like AID² showing outer loops policing inner loops while improving capabilities through defensive co-scaling
- 4Fine-tuning is becoming economically viable as a service, with models like Fable 5 making it accessible at scale beyond expert-only operations
- 5Digital twins of political leaders are emerging as tools for civic engagement and scaled communication across diverse populations
- 6Liquid AI's liquid neural networks, inspired by C. elegans worm neurobiology, offer an alternative to transformer architecture
Notable Quotes
“When the incumbents ask for the rules and they set the standards, they set up a barrier for all the entry level labs coming in.”
— Podcast host
“AI moves way way too fast for any kind of traditional government bureaucracy”
— Alex Wissner-Gross
“She's built exactly the thing hitting the market that exactly what everybody needs right now”
— Panel member
“Customization over leaderboard dominance is what's going to win her the day”
— Podcast host
Chapters
1. AI Regulation and Frontier Lab Consolidation
Discussion of CEOs from major AI labs (Sam Altman, Elon Musk, Demis Hassabis) calling for regulation. Demis proposes a FINRA-like standards body. Concerns raised about regulatory capture, moats against competitors, and whether incumbents are seeking liability protection rather than genuine safety measures.
2. Open-Weight Model Competition and Policy Framework
China dominates the open-weight model race due to different incentive structures. US White House proposes capability framework tied to China's open releases. Discussion of how this creates perverse incentives and why Western labs have been poorly incentivized to release strong open models versus closed API models.
3. Mira Murati's Inkling: Customization Over Leaderboards
Former OpenAI CTO launches Inkling, a 975B parameter mixture-of-experts model. Strategy focuses on fine-tuning and customization for enterprises rather than leaderboard dominance. Model enables on-premises deployment and proprietary data handling, addressing concerns about data privacy with closed models.
4. Fine-Tuning as a Service and Business Model Shift
Fine-tuning historically wasn't capability-increasing until reinforcement fine-tuning (RFT) emerged. Fable 5 makes fine-tuning accessible via prompt rather than engineering expertise. Thinking Machines bets on fine-tuning as their revenue engine while providing strong base models.
5. Recursive Self-Improvement and Defensive Co-Scaling
Wico AI publishes evidence of recursive self-improvement using outer and inner loops. Outer loop discovers reward hacking prevention, demonstrating emergent defensive co-scaling. Discussion of depths of customization from prompt engineering to pre-training automation as the true holy grail.
6. Liquid AI and Liquid Neural Networks
Ramin Hasani presents Liquid AI, which reimagines neural networks from first principles using C. elegans neurobiology. Liquid neural networks offer continuous-time recurrent processing alternatives to transformer attention mechanisms, enabling efficient, customizable models for specific domains.
7. Digital Twins and AI Avatars in Governance
Malaysia's PM prepares AI digital double for public communications across 135 languages. Discussion of how avatars enable scaled civic engagement, directional interactivity, and potential pathways for organizations to transition leadership to digital representatives.
Key People & Entities
- Mira Murati
- Former OpenAI CTO; founded Thinking Machines Labs; released Inkling open-weight model
- Ramin Hasani
- Co-founder and CEO of Liquid AI; pioneered liquid neural networks inspired by C. elegans neurobiology
- Demis Hassabis
- CEO of DeepMind; proposed FINRA-like frontier AI standards body
- Sam Altman
- CEO of OpenAI; published op-ed proposing US-led international AI standards forum
- Elon Musk
- CEO of Tesla and xAI; called for standalone AI regulator similar to FAA or FCC
- Daniela Ruoso
- Director of MIT Computer Science and AI Lab (CSAIL); recruited Liquid AI founders to MIT
- Anwar Ibrahim
- Prime Minister of Malaysia; deploying AI digital double for public communications
- Alex Wissner-Gross
- Podcast panelist; AI researcher discussing defensive co-scaling and regulatory challenges
Glossary
- FINRA
- Financial Industry Regulatory Authority; industry-funded self-regulatory organization that polices Wall Street under SEC oversight, proposed as a model for AI governance
- Mixture of Experts (MoE)
- Neural network architecture where different specialized sub-networks (experts) handle different types of inputs, improving efficiency and capability
- Fine-tuning
- Process of making small adjustments to a pre-trained model's weights to customize it for specific tasks or domains
- Reinforcement Fine-Tuning (RFT)
- Fine-tuning approach using synthetic data and full weight access to increase model capabilities beyond original training
- Recursive Self-Improvement (RSI)
- AI system that improves its own capabilities through automated processes, enabling self-accelerating development
- Defensive Co-Scaling
- Strategy of maintaining proportional alignment capabilities to match AI capabilities growth, analogous to police forces scaling with population
- Liquid Neural Networks
- Neural network architecture inspired by C. elegans neurobiology that uses continuous-time recurrent processes as alternative to transformers
- Chinchilla Law
- Scaling law stating that for optimal compute usage, model parameters should be matched with training tokens at approximately 20:1 ratio
- Digital Twin
- AI-powered virtual representation of a person trained on their communication patterns and knowledge to enable scaled interaction