Peter H. Diamandis's Mira Murati's 975B Open Model, Ramin Hasani on Post-Transformer AI, and Demis' AI FINRA | EP #271: skim's analysis identifies 14 key moments, with 2 potential conflicts of interest flagged. Liquid AI CEO Ramin Hasani joins Peter Diamandis, Dave Blundin, Salim Ismail, and Alexander Wissner-Gross to dissect frontier AI regulation, Mira Murati's 975-billion-parameter Inkling model, and post-transformer architectures.
Category: Tech. Format: Panel Discussion. YouTube video analyzed by skim.
Key Points (14)
1. Debating Demis Hassabis's FINRA Model for AI
Timestamp: 00:04:17 to 00:12:43
Peter H. Diamandis introduces calls from Sam Altman, the speaker, and Demis Hassabis for frontier AI oversight, highlighting Hassabis's proposal for a FINRA-style industry standards body. Ramin Hasani suggests modeling regulation as a Stackelberg game across vertical applications, while Salim Ismail and Dave Blundin argue static laws and multi-year staff rotations cannot keep pace with AI.
Significance (High): Exposes the structural mismatch between traditional self-regulatory bodies that rely on slow human committees and an AI industry iterating on weekly cycles.
Sources in support: Peter H. Diamandis (Host; Founder of XPRIZE, Singularity University, ZeroG, and A360), YouTube: Peter H. Diamandis
Sources against: Salim Ismail (Founder of Open ExO; GP at Exponential Venture Capital), Dave Blundin (Founder & GP of Link Ventures)
Neutral sources: Ramin Hasani (Co-founder and CEO of Liquid AI)
2. Alexander Wissner-Gross Warns of AI Cartels
Timestamp: 00:12:43 to 00:20:03
Dr. Alexander Wissner-Gross argues that Demis Hassabis's proposed FINRA-style oversight smells like regulatory capture designed to form a cartel of frontier labs that boxes out open-weight and academic competitors. He contends that regulators should police harmful model actions through legal liability rather than thought-policing raw model intelligence at construction time, while noting past proposals to monitor data center supply chains.
Significance (High): Distinguishes between regulating model capabilities versus downstream actions, warning that input caps protect incumbents at the expense of open innovation.
Sources in support: Dr. Alexander Wissner-Gross (Computer scientist and founder of Reified), Salim Ismail (Founder of Open ExO; GP at Exponential Venture Capital)
Sources against: Peter H. Diamandis (Host; Founder of XPRIZE, Singularity University, ZeroG, and A360)
3. Pegging US AI Limits to Chinese Open Models
Timestamp: 00:20:04 to 00:25:55
Peter H. Diamandis cites a Washington Post report that the White House is weighing a framework clearing US models as long as their capabilities stay at or below China's best open-weight model. Dr. Alexander Wissner-Gross, Dave Blundin, and Salim Ismail condemn the proposal, warning it creates a perverse incentive to let China lead, drives researchers abroad, and risks freezing flawed evaluation benchmarks.
Significance (High): Highlights how pegging domestic technology ceilings to a competitor's public releases cedes strategic pacing and risks accelerating brain drain.
Sources in support: Peter H. Diamandis (Host; Founder of XPRIZE, Singularity University, ZeroG, and A360), Dr. Alexander Wissner-Gross (Computer scientist and founder of Reified), Salim Ismail (Founder of Open ExO; GP at Exponential Venture Capital), Dave Blundin (Founder & GP of Link Ventures), Ramin Hasani (Co-founder and CEO of Liquid AI)
4. Mira Murati Ships 975B Open Model Inkling
Timestamp: 00:26:12 to 00:33:30
Peter H. Diamandis reports that former OpenAI CTO Mira Murati's startup, Thinking Machine Labs, released Inkling, a 975-billion-parameter mixture-of-experts open-weight model activating 41 billion parameters across text, image, audio, and video. Dr. Alexander Wissner-Gross notes Inkling beats Nvidia's Nemotron 3 but trails China's GLM 5.2, while Ramin Hasani praises Murati's strategy of leaving room for enterprise fine-tuning.
Significance (High): Marks a major Western push into frontier open-weight models, shifting the monetization strategy from raw API token sales to enterprise customization.
Sources in support: Peter H. Diamandis (Host; Founder of XPRIZE, Singularity University, ZeroG, and A360), Salim Ismail (Founder of Open ExO; GP at Exponential Venture Capital), Ramin Hasani (Co-founder and CEO of Liquid AI)
Neutral sources: Dr. Alexander Wissner-Gross (Computer scientist and founder of Reified), Dave Blundin (Founder & GP of Link Ventures)
5. Dave Blundin and Panelists Debate Fine-Tuning
Timestamp: 00:33:30 to 00:42:00
Dave Blundin explains that enterprises need local model fine-tuning to keep proprietary secrets off frontier lab servers, noting that Fable 5 now makes fine-tuning as simple as a prompt. Dr. Alexander Wissner-Gross counters that while reinforcement fine-tuning currently boosts reasoning capabilities, future generalist models may render fine-tuning obsolete, before Peter H. Diamandis calls for more women in AI leadership.
Significance (Medium): Clarifies the tension between immediate enterprise demand for private, fine-tuned models and the long-term trajectory toward generalist base models.
Sources in support: Dave Blundin (Founder & GP of Link Ventures), Peter H. Diamandis (Host; Founder of XPRIZE, Singularity University, ZeroG, and A360), Salim Ismail (Founder of Open ExO; GP at Exponential Venture Capital)
Sources against: Dr. Alexander Wissner-Gross (Computer scientist and founder of Reified)
6. Weco AI's Recursive Self-Improvement Claim
Timestamp: 00:43:05 to 00:48:55
Peter H. Diamandis introduces London startup Weco AI's AIDE2 system, which claims eight days of dual-loop machine self-improvement beat two years of human effort. Dr. Alexander Wissner-Gross praises the experiment as proof of defensive co-scaling where an outer AI loop polices an inner loop against reward hacking, and outlines Weco AI's four-level recursive self-improvement scale.
Significance (High): Frames outer-loop AI oversight as an organic alignment mechanism that prevents reward hacking while accelerating software research.
Sources in support: Dr. Alexander Wissner-Gross (Computer scientist and founder of Reified), Peter H. Diamandis (Host; Founder of XPRIZE, Singularity University, ZeroG, and A360)
7. Ramin Hasani Critiques Prompt-Level Loops
Timestamp: 00:48:56 to 00:54:52
Ramin Hasani challenges the excitement around Weco AI's AIDE2, explaining that editing system prompts and code patches with frozen model weights is not true recursive self-improvement. Citing the Chinchilla scaling law's 20-to-1 token-to-parameter ratio, he calculates that retuning a two-billion-parameter model's actual weights using Weco AI's framework would take 350 years.
Significance (High): Grounds recursive self-improvement hype in computational reality by separating superficial prompt optimization from actual neural weight evolution.
Sources in support: Ramin Hasani (Co-founder and CEO of Liquid AI), Dave Blundin (Founder & GP of Link Ventures)
Sources against: Dr. Alexander Wissner-Gross (Computer scientist and founder of Reified)
Neutral sources: Peter H. Diamandis (Host; Founder of XPRIZE, Singularity University, ZeroG, and A360)
8. Ramin Hasani's Three Depths of Customization
Timestamp: 00:54:52 to 01:02:35
Ramin Hasani outlines three depths of AI self-improvement: kernel and Python code optimization, autonomous fine-tuning of smaller models, and full pre-training automation, noting Anthropic hired Andrej Karpathy to tackle the third. While Hasani warns that multi-agent self-improvement creates real cybersecurity threats requiring enterprise self-checks within two years, Salim Ismail celebrates workflow-level self-improvement.
Significance (High): Maps the concrete engineering milestones between current code-patching agents and autonomous pre-training of next-generation foundation models.
Sources in support: Ramin Hasani (Co-founder and CEO of Liquid AI), Salim Ismail (Founder of Open ExO; GP at Exponential Venture Capital)
Neutral sources: Peter H. Diamandis (Host; Founder of XPRIZE, Singularity University, ZeroG, and A360)
9. Malaysia's Prime Minister Debuts AI Clone
Timestamp: 01:02:35 to 01:12:03
Peter H. Diamandis reports that Malaysian Prime Minister Anwar Ibrahim is preparing an official AI digital double to communicate across Malaysia's 135 spoken languages, following Albania's 2025 appointment of an AI cabinet minister. Salim Ismail, Dr. Alexander Wissner-Gross, and Dave Blundin predict that bidirectional digital twins with real-time visual capabilities will transform civic engagement, corporate leadership, and elections within two years.
Significance (Medium): Illustrates how state-sanctioned AI avatars shift political communication from one-to-many broadcasting to interactive, multilingual dialogue.
Sources in support: Peter H. Diamandis (Host; Founder of XPRIZE, Singularity University, ZeroG, and A360), Salim Ismail (Founder of Open ExO; GP at Exponential Venture Capital), Dr. Alexander Wissner-Gross (Computer scientist and founder of Reified), Dave Blundin (Founder & GP of Link Ventures)
10. From Nematode Neurons to Liquid AI's Founding
Timestamp: 01:12:03 to 01:20:17
After Dave Blundin and Peter H. Diamandis disclose their early investment in Liquid AI via MIT CSAIL Director Daniela Rus, Ramin Hasani explains how studying the 302 analog neurons of the C. elegans worm inspired Liquid Neural Networks. By modeling continuous-time differential equations rather than bloated transformers, Hasani and his co-founders built compact models capable of running physical robotics and edge devices on standard CPUs.
Significance (High): Shows how biological priors from a 302-neuron organism yielded a continuous-time mathematical alternative to standard transformer attention.
Sources in support: Ramin Hasani (Co-founder and CEO of Liquid AI), Dave Blundin (Founder & GP of Link Ventures), Peter H. Diamandis (Host; Founder of XPRIZE, Singularity University, ZeroG, and A360)
11. Deploying Sub-1GB Liquid Models in Mercedes
Timestamp: 01:20:17 to 01:28:32
Ramin Hasani defines small language models as efficient systems under 100 billion parameters specialized for on-device deployment. He details Liquid AI's partnership with Mercedes-Benz to deploy a multimodal model under 1 gigabyte onto a $60 automotive chip with 2 to 8 gigabytes of RAM, rolling out via a 600-megabyte over-the-air update to North American vehicles from 2022 onward to control 700 to 1,200 car functions offline.
Significance (High): Demonstrates commercial viability for offline, on-device multimodal AI running on constrained $60 automotive hardware without cloud latency or privacy risks.
Sources in support: Ramin Hasani (Co-founder and CEO of Liquid AI), Peter H. Diamandis (Host; Founder of XPRIZE, Singularity University, ZeroG, and A360)
12. Automated Architecture Search at Liquid AI
Timestamp: 01:28:32 to 01:37:46
Ramin Hasani outlines Liquid AI's enterprise deployments with Insilico Medicine, Shopify, and AMD before Dr. Alexander Wissner-Gross questions whether Liquid AI has drifted toward conventional transformers. Hasani responds that Liquid AI uses its STAR and Automated Foundation Model Design algorithms to search across dynamical systems for memory, speed, latency, and accuracy, discovering an architecture composed 80% of double-gated convolutions.
Significance (High): Directly answers technical skepticism by detailing how automated architecture search validates input-dependent gating mechanisms over pure attention.
Sources in support: Ramin Hasani (Co-founder and CEO of Liquid AI)
Neutral sources: Dr. Alexander Wissner-Gross (Computer scientist and founder of Reified), Peter H. Diamandis (Host; Founder of XPRIZE, Singularity University, ZeroG, and A360)
13. Rejecting Palmer Luckey's Secret Patent Push
Timestamp: 01:39:36 to 01:48:07
Peter H. Diamandis plays a clip of Anduril Chairman Palmer Luckey arguing that public US patents act as an instruction manual for China and calling to massively expand classified patents under the 1951 Invention Secrecy Act. Dr. Alexander Wissner-Gross strongly opposes the idea, explaining that secrecy orders restrict inventions to military use and create secret monopolies, while Salim Ismail and Dave Blundin debate proprietary learning loops versus IP trade embargoes.
Significance (Medium): Exposes the legal reality of the Invention Secrecy Act, showing that classifying patents confiscates civilian commercial rights rather than protecting free-market innovators.
Sources in support: Dr. Alexander Wissner-Gross (Computer scientist and founder of Reified), Salim Ismail (Founder of Open ExO; GP at Exponential Venture Capital)
Sources against: Palmer Luckey (Founder of Oculus and Chairman of Anduril (featured in clip))
Neutral sources: Peter H. Diamandis (Host; Founder of XPRIZE, Singularity University, ZeroG, and A360), Dave Blundin (Founder & GP of Link Ventures)
14. AI Medical Benchmarks and Reversing Glycation
Timestamp: 01:48:08 to 01:55:58
Peter H. Diamandis highlights that OpenAI's GPT-5.6-Sol beat specialty-matched doctors across 20,000 judgments on HealthBench Professional, only for Meta's free Muse Spark 1.1 to beat GPT-5.6-Sol at one-seventh the cost for 3.56 billion users. He and Dr. Alexander Wissner-Gross then celebrate a Nature Communications study by Revel Pharmaceuticals and Calico showing an engineered bacterial enzyme, CmlA, reverses advanced glycation end-products in elderly human tissue.
Significance (High): Pairs the collapse of global medical diagnostic costs with a biochemical breakthrough that reverses previously permanent age-related protein cross-linking.
Sources in support: Peter H. Diamandis (Host; Founder of XPRIZE, Singularity University, ZeroG, and A360), Salim Ismail (Founder of Open ExO; GP at Exponential Venture Capital), Dr. Alexander Wissner-Gross (Computer scientist and founder of Reified)
This analysis was generated by skim (skim.plus), an AI-powered content analysis platform by Credible AI. Scores and classifications represent the platform's AI-generated assessment and should be considered alongside other sources.
