"Mira Murati's 975B Open Model, Ramin Hasani on Post-Transformer AI, and Demis' AI FINRA | EP #271" — Peter Diamandis (Moonshots)

Why this is in the vault

This episode sits at the intersection of three active RDCO tracking threads: the open-weights competitive landscape (Inkling/Thinking Machines vs DeepSeek), the post-transformer architecture bet (Liquid AI on-device SLMs with real enterprise deployments), and the AI governance conversation sparked by Demis Hassabis's FINRA proposal. Ramin Hasani's in-depth technical pushback on the WoCo recursive self-improvement claim - and his naming of Anthropic as the RSI leader - is the most substantive primary-source technical content in any Moonshots episode this quarter.

Episode summary

EP #271 is a five-segment news roundup anchored by a rare extended technical interview with Ramin Hasani (co-founder and CEO of Liquid AI). The panel debates the Hassabis FINRA proposal, a White House open-model policy that pegs US release ceilings to China's best open-weight model, and the significance of Mira Murati's Inkling (975B MOE, open-weight). Hasani provides a detailed origin story for Liquid AI (C. elegans neuroscience → CSAIL → STAR automated architecture search) and current enterprise deployments with Mercedes-Benz and Shopify. Late segments cover the Malaysian PM's AI digital double, Palmer Luckey's patent secrecy argument, AI healthcare benchmarks, and a longevity breakthrough in glycation reversal.

Key arguments / segments

[00:04:02] Demis Hassabis FINRA proposal and AI regulation

• Sam Altman (FT op-ed), Elon Musk, and Demis Hassabis all calling for US-led AI standards bodies this week
• Demis proposes a FINRA-equivalent — industry-funded, under SEC-like oversight — to test frontier models before release; wants it operational by end of 2026
• Alex Wisner-Gross: "It smells like regulatory capture... an attempted formation of a cartel of frontier labs"
• Ramin: regulations need to be capability-indexed per vertical; notes Liquid AI already submitted jointly with AMD to DoD on design of such frameworks
• Dave Blundin: FINRA analogy breaks because top AI talent won't leave for 2-year regulator stints the way finance people do

[00:20:00] White House open-model ceiling pegged to China

• Washington Post report: White House weighing a policy that caps permissible US open-weight releases at or below China's best open-weight model
• Panel diagnosis: creates perverse incentive to let China win the open-weight race so US labs can escape regulation; "throwing the steering wheel out the window in a game of chicken" (Alex)
• Salim: "trying to uninvent the printing press"

[00:26:00] Mira Murati's Inkling — 975B open-weight model

• Thinking Machine Labs (Murati's startup): 975B total parameters, MoE firing 41B at a time; trained on 45T tokens across text, image, audio, video; reasons natively across all four modalities
• Deliberately not claiming SOTA: "customization over leaderboard dominance is what's going to win her the day"
• Benchmarks: stronger than NVIDIA Neatron; weaker than GLM 5.2 (strongest open-weight overall)
• Business model: fine-tuning as a service; leaves room in the architecture for enterprise fine-tuning
• Ramin on the business thesis: "fine-tuning generates one to two orders of magnitude more tokens than inference... even printing money at a larger speed"
• Alex: OpenAI launched reinforcement fine-tuning (RFT) as a service first — "no one used it" — and has since wound down their fine-tuning API

[00:43:00] WoCo AI recursive self-improvement claim

• Startup WoCo AI (London, UC London grads) published "AID²" — outer AI rewrites inner AI's code and research strategy
• Claim: 8 days of machine RSI beat 2 years of expert human effort
• WoCo's RSI scale: Level 0 (delegation) → Level 1 (net positive) → Level 2 (ignition) → Level 3 (inflection); they claim Level 1 sparks
• Alex enthusiastic: frames outer loop policing inner loop as "defensive co-scaling" — alignment through capability co-scaling, not through pause
• Ramin skeptical: "This gives you the assumption that every single AI model in your pipeline is already fixed with certain capabilities — there are no weight changes." True RSI requires retuning weights; their framework would take 350 years to meaningfully fine-tune a 2B parameter model given Chinchilla ratios
• Ramin: Anthropic "has been having a lead on all of these things because they thought about this before everybody else"; Andrej Karpathy joined Anthropic to work on pre-training automation ("automation of automation")
• Dave: distinction between learning (20-year timescale) and recursive evolution (10M-year timescale)

[01:02:00] Malaysian PM AI digital double

• PM Anwar Ibrahim deploying an official AI-generated digital double for public communications (Malaysia has 135 spoken languages)
• Salim: excited about scaling civic engagement; risk is authenticity collapse without watermarking
• Alex: generalizes social media broadcast from one-to-many to bidirectional; predicts digital twins of CEOs and religious leaders; flags inflection point where the digital twin becomes the actual decision-maker

[01:12:00] Liquid AI deep dive with Ramin Hasani

• Origin: Ramin's PhD in Vienna under Radu Grosu (2015), inspired by C. elegans worm (302 neurons, 78% genome similarity to humans); joined MIT CSAIL (2017) under Daniela Rus
• Core idea: pack more information density into individual compute units via continuous-time neural networks; liquid neural networks use nested non-linearities and physics-inspired dynamics
• STAR framework (STAr): automated architecture search across 100+ operation variants; four objective functions (memory, compute efficiency, latency, accuracy); found double-gated convolution mechanisms dominate when running without human bias
• Current architectures: hybrid models, gated linear attention variants — post-transformer but not pure-recurrent in production
• Products: SLMs (below 100B parameters) optimized for on-device deployment

[01:22:00] Mercedes-Benz and Shopify deployments

• Mercedes: multimodal foundation model <1GB running on the car's chip (~$60 Qualcomm/Samsung chip); 600MB OTA update for all Mercedes-Benz North America 2022+ cars; gives access to 700-1,200 car functions; private (offline-capable); memory-persistent per user
• Shopify: 1B+ requests served; hundreds of millions of users; 10B products; in production 6 months
• AMD partnership: AI-PC experiences; Liquid submitting joint DoD proposals with AMD
• Model-plus-X platform: sells not just model weights but ongoing customization capability to enterprises

[01:40:00] Palmer Luckey — patents as national security liability

• Palmer: "the entire patent office could be downloaded every morning and used to fight a war against you"; wants expansion of 1951 Invention Secrecy Act (currently ~6,000 secrecy orders) into default mechanism
• Alex: "terrible idea" — secrecy orders grant exclusive rights to the military only, not the inventor; would create secret monopolies; argues Invention Secrecy Act should probably go away
• Salim: real defensibility is learning loops and trade secrets, not ownership; "the only people that win in this model are the lawyers"

[01:48:00] AI healthcare benchmarks

• GPT-5.6 "SAW" sets record on HealthBench Professional (OpenAI's hardest medical benchmark)
• In blind test of ~20,000 physician judgments: GPT-5.6 beat specialty-matched physicians given unlimited web access and unlimited time
• Meta's Llama Spark 1.1 ("Muse Spark") beat GPT-5.6 across HealthBench at 7x lower cost; free in WhatsApp/Facebook serving 3.56B daily active users
• Alex notes mild benchmarking optimism warranted but "it's a great day when Instagram gives better medical advice than human doctors"

[01:52:00] Glycation reversal — Revel Pharmaceuticals (Nature Communications)

• Engineered enzyme CMLA acts as "molecular lawn mower" — breaks advanced glycation end-product (AGE) crosslinks in proteins
• Demonstrated in human tissue from elderly donors, reversing lifetime accumulation
• Developed via directed evolution from a bacterial enzyme; involves Maillard-class reactions
• Calico (Alphabet) reportedly deeply involved; Alex sees large longevity potential alongside GLP-1s

Notable claims

• [00:27:00] Inkling specs: 975B total parameters, 41B active (MoE), trained on 45T tokens across 4 modalities
• [00:28:00] Inkling stronger than NVIDIA Neatron on evals released by Thinking Machines; weaker than GLM 5.2 (strongest open-weight globally)
• [00:29:00] Alex: Anthropic approaching IPO at ~$1T valuation; OpenAI planning to IPO at ~$1T — cited as why western labs have been poorly incentivized to release open-weight models
• [00:52:00] Ramin: using WoCo's framework, meaningfully fine-tuning a 2B parameter model would take ~350 years given Chinchilla compute ratios
• [00:53:00] Ramin: Andrej Karpathy joined Anthropic specifically to work on pre-training automation
• [01:14:00] Liquid AI reached unicorn status from MIT CSAIL faster than any prior company in MIT history
• [01:24:00] Mercedes deployment: model under 1GB, chip as cheap as $60, OTA update is 600MB, covers all 2022+ Mercedes-Benz North America vehicles
• [01:31:00] Liquid AI at Shopify: 1B+ requests, 100M+ users, 10B products; 6 months in production
• [01:49:00] Meta Muse Spark 1.1: beats GPT-5.6 on HealthBench at 7x lower cost; free to 3.56B daily Meta users

Guests

Ramin Hasani — Co-founder and CEO of Liquid AI. PhD from Vienna University of Technology and MIT CSAIL under Daniela Rus. Research origin: C. elegans-inspired continuous-time neural networks (liquid neural networks). Co-founded Liquid AI in early 2023 with Matias Lechner. Persian-Austrian, based in San Francisco. Board interests: competitive sports, game theory.

Peter H. Diamandis (host) — Founder of XPRIZE, Singularity University, Zero-G, and A360. Primary host/organizer.

Alex Wisner-Gross (AWG) — Scientist, investor; writes the Accelerando newsletter. Generally skeptical of regulatory capture; bullish on open ecosystems and defensive co-scaling as alignment strategy.

Dave Blundin (DB2) — Founder of Link Studios and Link Ventures; early investor in Liquid AI (introduced by Daniela Rus). Tends to the practical enterprise and business-model take.

Salim Ismail (Seem) — Founder of Open ExO; GP at Exponential Venture Capital. Organizational singularity thesis; frequently frames exponential institution dynamics.

Sponsorship

Blitzy (blitzy.com) — Autonomous software development platform using thousands of specialized AI agents with infinite code context. Claims enterprises achieve 5x engineering velocity. Mid-roll sponsor read at approximately 00:42-00:43.

Fountain Life (fountainlife.com/per) — Longevity health screening and diagnostics. Features interview with chief medical officer Don Mucalem on cognitive health and dementia prevention. Mid-roll sponsor at approximately 01:38-01:39.

Mapping against Ray Data Co

Strong signal — open-weight model landscape: Inkling's release and the White House open-model ceiling debate are directly relevant to RDCO's AI vendor tracking. The panel's analysis of why western labs under-invest in open-weight models (per-token API revenue too good) maps to the strategic question of whether RDCO should hedge Claude dependence with open-weight fine-tuned alternatives. Ramin's framing of fine-tuning-as-a-service and Liquid AI's Model+X platform is the commercial shape RDCO would need to evaluate if pursuing on-prem deployment for any phData client work requiring data sovereignty.

Strong signal — Anthropic's RSI lead and Karpathy hire: Ramin citing Anthropic as ahead on recursive self-improvement and Andrej Karpathy's hire for pre-training automation is a primary-source signal (from a founder peer) about Anthropic's internal capability trajectory. This is directly load-bearing for RDCO's primary bet on Claude as the production model.

Medium signal — Demis FINRA proposal: Demis's FINRA framework proposal is moving from think-piece to serious advocacy (Bloomberg, FT, industry op-eds all this week). RDCO should track whether phData clients in regulated verticals (banking, defense) will see AI procurement gated by a FINRA-equivalent; this would accelerate demand for sovereign/on-prem SLMs of exactly the type Liquid AI offers.

Medium signal — healthcare AI benchmarks: GPT-5.6 / Muse Spark beating specialty-matched physicians on HealthBench is directly relevant to RDCO's Sanity Check positioning: AI as diagnostic accelerator is no longer futurist framing, it is current benchmark reality. 3.56B free-tier medical AI users is a macro context point.

Weak signal — longevity glycation: Revel CMLA enzyme is early-stage (Nature Communications → clinical trials ahead), but RDCO's health/longevity tracking thread should note it. Calico involvement suggests Alphabet is still a serious longevity player.

Related

• [[2026-07-14-hassabis-frontier-ai-finra-framework]] — Demis Hassabis's original essay on frontier AI governance; the primary source for the FINRA proposal this episode debates
• [[2026-07-13-moonshots-ep270-grok-vs-gpt]] — Prior Moonshots episode (EP #270); same panel minus Ramin
• [[2026-07-16-innermost-loop-open-weights-chip-capex-singularity]] — Open-weights context and chip capex analysis from same week
• [[2026-07-17-innermost-loop-kimi-k3-frontier-moat]] — Same-day analysis of Kimi K3 and frontier model moat dynamics
• [[2026-06-01-moonshots-opus48-hassabis-agi]] — Earlier Moonshots episode with Hassabis on AGI timelines