Peter H. Diamandis

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

1h 57mJul 17, 2026

Key Themes

AI governanceOpen-weight modelsRecursive self-improvementOn-device AIDigital avatarsAI healthcarePost-transformer AILongevity biotech

Summary

A wide-ranging AI episode on regulation, open-weight models, self-improving systems, Liquid AI’s edge strategy, AI avatars, healthcare, patents, and longevity biotech.

The episode centers on the tension between accelerating AI capability and the governance, business, and social structures needed to absorb it. The panel debates frontier AI standards bodies, US-China open-model policy, Mira Murati’s Thinking Machine Labs and its Inkling open-weight model, the meaning of recursive self-improvement, and Ramin Hasani’s Liquid AI approach to small, post-transformer, on-device models. Later sections examine official AI avatars for leaders, Liquid AI’s Mercedes-Benz deployment, patent secrecy in national security, AI medical intelligence, and enzyme-based longevity research.

AI governance is shifting from abstract principles to concrete release controls

The panel repeatedly returns to the design problem of regulating frontier AI: who sets standards, what capabilities should trigger review, and whether rules should target model releases, harmful actions, or compute infrastructure. The discussion suggests that governance structures may emerge soon, but their legitimacy will depend on avoiding incumbent capture and adapting faster than conventional bureaucracy.

Open-weight AI is becoming a geopolitical and strategic fault line

The episode frames open models not just as a technical choice but as a national-strategy issue. A policy pegging U.S. open releases to Chinese capabilities could create strange incentives, while models already released online are practically impossible to recall. The panel argues that openness has historically strengthened the U.S. innovation ecosystem, even as it complicates safety and security.

Customization may matter as much as raw model rankings

The discussion of Thinking Machine Labs’ Inkling highlights a broader point: many organizations may care less about having the single best general model and more about adapting a capable model to their own workflows, data, privacy requirements, and deployment constraints. This makes fine-tuning, on-premise use, and data sovereignty central to how AI will be adopted in real organizations.

The term recursive self-improvement needs careful definition

The panel separates lighter forms of AI-assisted optimization from deeper systems that modify model weights, architectures, or training algorithms. That distinction matters because claims of recursive self-improvement can sound like hard-takeoff scenarios even when the underlying system is mainly improving prompts, code, or workflows rather than redesigning foundation models themselves.

AI avatars could reshape public communication, but verification becomes essential

The segment on Malaysia’s prime minister and broader digital twins shows how AI avatars could make leaders and organizations more accessible, multilingual, and interactive. At the same time, the panel stresses that official avatars could blur authenticity unless supported by watermarking, verification, and clear norms around what is official, synthetic, or manipulated.

Edge AI is framed as a practical alternative to cloud-only intelligence

Liquid AI’s Mercedes-Benz example illustrates why some AI systems need to run locally: connectivity can fail, latency matters, private spaces require stronger data control, and many devices have tight memory limits. The chapter uses cars as a concrete case for a broader move toward intelligence outside data centers.

Post-transformer AI may be less about one replacement architecture and more about automated design search

Ramin Hasani describes Liquid AI as drawing on liquid neural network ideas while avoiding a single architectural dogma. The approach is presented as an automated search across transformers, dynamical systems, attention variants, and hybrids, optimized for practical constraints such as memory, compute, latency, and accuracy.

AI will pressure existing systems for intellectual property, healthcare access, and longevity science

The closing chapter shows AI’s spillover into institutional domains beyond model labs: patent secrecy debates, medical diagnostics, mass distribution through consumer apps, and biotechnology research into aging damage. Across these topics, the common pattern is that AI accelerates capability faster than existing legal, healthcare, and scientific systems are used to handling.

Select any chapter text to Deep Dive with AI

01AI Regulation and Standards Bodies

The panel examines whether frontier AI needs a new standards or regulatory body, drawing on proposals associated with Sam Altman, Elon Musk, and Demis Hassabis. Speakers debate whether a FINRA-like AI body could create useful testing and auditing or instead entrench incumbent labs, disadvantage open-weight and academic work, and become a form of regulatory capture.

Demis Hassabis’s proposed frontier AI standards body is framed as a possible way to test major models before release.

Panelists worry that incumbent-designed rules could raise barriers for smaller labs, open-weight models, universities, and non-incumbent players.

Ramin Hasani argues that AI regulation should account for capability thresholds and sector-specific deployment contexts.

The discussion contrasts ex ante regulation of model capabilities with ex post liability for harmful model actions.

Compute supply chains, chips, foundries, and data centers are raised as possible but controversial chokepoints for AI governance.

02US-China AI Capability Framework

The panel critiques a proposed framework that would let U.S. companies release open models only up to a capability ceiling pegged to Chinese open-weight models already available. Speakers argue that this would create perverse incentives, make China an indirect pace-setter for U.S. openness, and potentially push talent or technical progress toward less constrained jurisdictions.

Open-weight releases are described as hard to ban or recall once widely downloaded.

A China-pegged release ceiling is criticized for letting Chinese model releases indirectly determine what U.S. labs can publish.

Speakers warn that researchers and companies could respond to restrictive policy by moving talent, projects, or incentives abroad.

The panel emphasizes that open ecosystems and permissionless innovation have historically been U.S. strengths.

Fixed capability benchmarks are presented as risky because they can freeze flawed measurement and distort model development.

03Thinking Machine Labs’ Inkling and the Open-Weight AI Race

The hosts discuss Thinking Machine Labs’ Inkling, a 975B-parameter open-weight model led by former OpenAI CTO Mira Murati. Rather than presenting Inkling mainly as a leaderboard champion, the panel frames it as a customization-first model for enterprise fine-tuning, on-prem deployment, and data-sovereign workflows in sectors such as defense, banking, biotech, and other proprietary-data-heavy environments.

Inkling is introduced as an open-weight model with 975B total parameters and 41B active at a time.

The model is positioned as a Western response to Chinese open-weight models such as DeepSeek and Qwen.

The panel sees customization, fine-tuning, and proprietary workflows as the core strategic bet.

Speakers contrast Western closed-API incentives with China’s stronger incentives around open-weight releases and integration.

On-premise deployment is linked to privacy, proprietary data, and data sovereignty.

The segment notes Mira Murati’s prominence amid a shortage of women leaders in frontier AI.

04Recursive Self-Improvement, Model Development, and AI Safety

The panel debates claims that a startup has demonstrated an early form of recursive self-improvement. Alex sees importance in a system where an outer AI loop improves and polices an inner AI loop against reward hacking, while Ramin Hasani argues that true recursive self-improvement must alter weights, architectures, learning algorithms, or core model capabilities rather than merely optimize prompts or code workflows.

Recursive self-improvement is defined as AI improving itself and then using the improved system to create still stronger AI.

A startup’s AI-driven exploration system is described as an early or limited self-improvement loop.

The idea of defensive co-scaling is introduced: stronger AI systems policing weaker or riskier systems as capabilities grow.

Ramin Hasani distinguishes workflow or prompt optimization from deeper model self-improvement involving weights, architectures, and training methods.

The panel emphasizes that true RSI is likely compute-intensive and probably being pursued by major foundation-model labs.

Cybersecurity risks and release controls are treated as serious issues around increasingly autonomous improvement pipelines.

05Foundation Model Self-Improvement, AI Avatars, and Liquid AI

The discussion moves from near-term expectations for rapidly improving foundation models into different depths of customization and self-improvement. It then explores AI-generated digital doubles for political leaders, CEOs, organizations, religious figures, and personal legacy use cases before introducing Liquid AI’s MIT/CSAIL roots and bio-inspired approach to efficient post-transformer models.

Speakers expect very rapid improvement in foundation models if compute and memory supply constraints do not intervene.

Recursive improvement is reframed as self-accelerating experimentation that improves the process of innovation itself.

Malaysia’s prime minister is cited as preparing an official AI digital double for multilingual public communication.

The panel highlights risks around authenticity, deepfakes, elections, watermarking, and verification.

Digital twins are predicted to become a bidirectional successor to social media for leaders and institutions.

Liquid AI’s origin story is linked to bio-inspired research on the C. elegans nervous system and a goal of running high intelligence outside data centers.

06Small Language Models, On-Device AI, and Liquid AI’s Mercedes Deployment

Ramin Hasani explains Liquid AI’s approach to small, customizable foundation models designed for on-device use rather than cloud-only deployment. Automotive is presented as an ideal use case because cars have constrained hardware, intermittent connectivity, privacy expectations, and safety-critical needs. The chapter details Liquid AI’s Mercedes-Benz rollout and broadens into its enterprise platform and automated architecture-search strategy.

Small language models are framed as general-purpose but often specialized systems that can be adapted to dedicated tasks.

Hasani describes “small” as generally below 100B parameters but emphasizes that on-device viability depends on hardware constraints.

Automotive AI is highlighted because in-car chips may have only 2GB to 8GB of RAM and cannot always rely on cloud connectivity.

Liquid AI’s Mercedes-Benz model is described as a multimodal foundation model under 1GB, delivered through an approximately 600MB over-the-air update.

The in-car AI is said to access hundreds of vehicle functions and support private, offline interaction.

Liquid AI’s enterprise offering is described as a model plus customization platform, not just static model weights.

Hasani says Liquid AI uses automated architecture search across transformers, dynamical systems, attention variants, and hybrids to optimize memory, latency, efficiency, and accuracy.

07Patent Secrecy, AI Healthcare Abundance, and Longevity Enzymes

The final chapter begins with a Fountain Life sponsor segment on preventive brain health, then turns to Palmer Luckey’s proposal to expand classified national-security patents. The panel debates whether secrecy protects strategic inventions or undermines the patent bargain and open innovation. The episode closes with claims about AI medical models becoming broadly accessible through Meta products and a longevity segment on engineered enzymes that may reverse glycation-related tissue damage.

The sponsor segment emphasizes dementia prevention, brain-age testing, and lifestyle interventions.

Palmer Luckey’s position is framed as concern that patent disclosures can function as instruction manuals for adversaries.

Panelists warn that secret patents could create hidden monopolies and suppress civilian innovation.

The patent discussion shifts toward trade secrets, proprietary learning loops, and continuous innovation in an AI era.

The hosts discuss AI medical models performing at or above specialist physicians on certain benchmarks and becoming broadly distributed through consumer platforms.

Revel Pharmaceuticals and Calico are discussed for engineered enzyme work targeting advanced glycation end products in human tissue samples.