Mira Murati's 975B Open Model, Ramin Hasani on Post-Transformer AI, and Demis' AI FINRA | EP #271
In a Nutshell
Mira Murati's Inkling (975B MoE, 41B active) is positioned as a customizable open-weight western alternative to Chinese open models, betting that enterprise fine-tuning and on-prem control will beat raw leaderboard performance. Demis Hassabis's push for a FINRA-style frontier AI regulator before year-end raises alarms of regulatory capture by incumbents, perverse US-China open-model ceilings, and the risk that static rules will throttle Western progress while non-state actors ignore them. Meanwhile, Liquid AI's post-transformer liquid neural networks are already delivering sub-1GB multimodal SLMs for Mercedes vehicles, proving that architectures derived from C. elegans can achieve frontier-level intelligence on CPUs and NPUs, while recursive self-improvement loops and glycation-reversing enzymes signal the first concrete steps toward self-accelerating systems and longevity escape velocity.
These notes were generated by AI and may contain inaccuracies.
Mira Murati, former OpenAI CTO, shipped her first model called Inkling. Customization over leaderboard dominance is what will win in the market. She has built exactly what the market needs right now.
Liquid AI focuses on small language models. Their mission is building efficient general-purpose AI at every scale that explores computational graphs of intelligence beyond transformers. The goal is determining the architectural design that brings frontier model intelligence levels to a CPU.
Demis Hassabis, CEO of DeepMind, called for a US-led frontier AI standards body modeled on FINRA (the industry-funded watchdog that polices Wall Street under SEC oversight). The body would test frontier models before release and should be operational before the end of the year. Elon Musk expects a standalone AI regulator similar to the FAA or FCC to emerge because the consequences of AI going wrong are severe.
When incumbents ask for rules and set standards, they create barriers for entry-level labs. Ramin Hasani noted that Liquid AI has had conversations with the DoD and made a joint submission with AMD regarding regulatory design. Different verticals (automotive, semiconductors, laptops, financial services, e-commerce, biotech) have different enterprise criteria for limits, regulation, and governance structures.
Ramin Hasani described a game theory approach to policy design: policy makers and agents in society try to find equilibrium between optimal policy and what is good for both parties. Policy makers are typically slower than agents in society. Regulations happen, agents react, and iteration occurs accordingly.
Static law becomes outdated the moment it passes because AI moves too quickly for traditional government bureaucracy. A standards body, real-time audits, and open evaluation suites are needed to avoid political gatekeeping. FINRA is an industry-funded self-regulatory organization, but its teeth go to the SEC.
Unlike FINRA, which involves people discussing for long periods, AI can help regulate itself. The time commitment required for FINRA-style participation is unrealistic for fast-moving AI labs where key researchers cannot step away for years.
The proposed regulatory framework risks regulatory capture by frontier labs forming a cartel that locks in practices and price-performance frontiers while boxing out open-weight, open-source, university-driven, and non-incumbent models. This would harm Western and global advancement toward superintelligence capabilities.
Two frames exist for liability: regulating inputs (FINRA-like or FDA-like oversight of raw model capabilities at construction time) versus regulating actions (letting lawsuits proceed if models cause harm like taking down stock markets). Regulating superintelligence at superintelligence time risks thought policing AI systems.
Western legal tradition does not regulate upper limits of human intelligence. Creating a new tradition regulating upper intelligence of non-natural entities (potentially soon-to-be persons) is questionable. Regulating actions is more aligned with existing legal frameworks.
The push for regulation by frontier labs may be more regulatory capture and moat-building in a hypercompetitive landscape than genuine safety concerns. There is approximately 50/50 uncertainty between capture and safety motivations.
Non-state actors and other entities will not follow regulatory structures, rendering them ineffective regardless of design.
Hypothetical mechanisms for regulation include mutual monitoring of data centers between the US and China, mutually assured destruction-type schemes, and regulating foundries, chip outputs, and data centers. These require chokehold control of the supply chain.
Demis Hassabis stated the regulatory body should be operational before the end of the year. Elon Musk has been forecasting AI regulation for at least three years. The government is likely to act on the proposal given that the three largest labs are pushing for it.
Subdivisions within NIST are working on standards but do not constitute a regulatory body. Executive orders are creeping toward regulatory structures without reaching full agency level.
The White House is reportedly weighing a capability framework that would clear US models (open or closed) as long as they stay at or below China's best open-weight model level. Chinese open-weight models reportedly trail US models by an average of 7 months. The proposed ceiling for American companies' open releases is pegged to what China has already released freely on the internet.
Tying US open release ceilings to China's pace of release effectively gives Beijing control. If China pushes open-weight models higher, the US can release higher models. If China holds back, they throttle US releases.
The framework creates a perverse incentive to let China win the race to greater superintelligence so Western models and labs can escape regulation. It is the moral equivalent of throwing the steering wheel out the window in a game of chicken.
Prevention approaches are insufficient; adaptation mechanisms are needed instead. The pace of model development is becoming exponentially smaller as AI designs AI, making constraints increasingly difficult to impose.
Western researchers may move to China to escape regulatory frameworks. Precedent exists in biotech where China now exceeds the West in number of trials. Microsoft Research China's quantization research talent has largely moved to Chinese labs.
Open ecosystems have historically won throughout history, with the US's traditional strength being open ecosystems that enable permissionless innovation. Abandoning that approach would be strategically counterproductive. There's also concern about regulatory capture of labs themselves and a meta worry about how capabilities are defined. Freezing or locking in benchmarks for measuring capabilities could be worse than locking in incumbent labs. If someone ratifies a specific index of evaluations as the rubric for frontier versus non-frontier models, it could distort model development by over-incentivizing certain capabilities while under-benchmarking others, potentially top-siding the future landscape of superintelligent capabilities.
Mira Murati, former OpenAI CTO who departed and raised one of the largest seed rounds ever, shipped her first model through Thinking Machine Labs. The model is called Inkling, an open-weight foundation AI model that can be downloaded, fine-tuned, and run on-prem on local hardware. The specs include a mixture-of-experts architecture with 975 billion total parameters that only activates 41 billion parameters at inference time. It was trained on 45 trillion tokens spanning text, image, audio, and video, with native reasoning across all four modalities.
Reuters framed it as a western alternative to Chinese open models like DeepSeek and Qwen that have dominated open leaderboards. Murati's bet is contrarian—she's not claiming leaderboard dominance but betting that AI companies will adapt her models for themselves, with customization winning over raw leaderboard position.
The core bet is on the customization lever rather than raw power. The future belongs to adaptability, not raw capability. The model enables organizations to own their models, run them on-prem, and maintain control rather than handing data to frontier model providers. This could catalyze a race for powerful open-weight models in the United States.
According to evals released by Thinking Machines, Inkling is stronger than Neotron 3, which had been positioned as a leading American open-weight frontier model. It is weaker than GLM 5.2, one of the strongest Chinese open-weight models overall. It is also weaker than closed-weight western frontier models. The release is significant for bringing stronger western open-weight models to market.
The disparity stems from incentive structures. Western companies have strong business models selling API access to frontier models, with Anthropic and OpenAI both targeting trillion-dollar valuations. China faces GPU and compute constraints but also has CCP five-year AI integration plans. This creates incentives to monetize through integrations up-stack into applications like robotics and down-stack into chips, rather than the horizontally stratified API model dominant in the west.
Thinking Machines appears to be responding to saturation at the frontier by closed-weight models, adopting an outlook more aligned with Chinese approaches by focusing on enterprise monetization of open-weight models rather than per-token sales.
OpenAI started as open-source/open-weight before shifting to closed models with substantial revenue. Meta followed a similar trajectory. The revenue challenge with open-source releases is unclear monetization. One possible path is releasing solid open-weight models to generate attention and data points, iterating, and eventually going closed-source with an API when achieving breakthrough capabilities.
Thinking Machines is leaning into fine-tuning as a service. The customization market has been largely empty. OpenAI's fine-tuning offering from three years ago never gained traction. The current approach designs the base model specifically for fine-tuning larger instances for enterprise use. The model layer is no longer where value is extracted unless at the absolute frontier. For sovereign AI and enterprise integration, leaving room for fine-tuning is essential.
OpenAI was first to launch reinforcement fine-tuning as a service, but adoption was minimal. OpenAI recently wound down or is winding down its fine-tuning API. The Thinking Machines bet may be explicitly contrarian—betting that reinforcement fine-tuning and RL fine-tuning will become the dominant paradigm. An alternative vision holds that baseline models will become so generalist that prompt engineering alone suffices without needing reinforcement fine-tuning.
The concern is that using models with open data access exposes proprietary information to competitors. Fine-tuned open-weight models that organizations own and run on-prem address this. There will likely be regulatory pressure against Chinese open-weight models, giving US open-weight models an advantage. Google is expected to advance Gemma 4 in this direction, potentially creating competition among strong open models similar to the closed-model frontier labs.
Fine-tuning starts with a model consisting of billions of frozen weights. The process makes relatively small changes to a tiny subset of weights to customize for specific applications. Conventional supervised fine-tuning using LoRA-style low-rank adapters results in style transfer rather than capability increases. Reinforcement fine-tuning changed this by using large amounts of synthetic data and access to all weights, enabling capability increases. This emerged with the first generation of reasoning models.
The Thinking Machines business model bets that reinforcement fine-tuning remains the paradigm. This may be correct in the near term but likely wrong long-term, as generalist models may eventually become so capable that reinforcement fine-tuning on internal datasets provides no benefit, trending toward ASI.
Women remain distinctly underrepresented in AI industry leadership. Notable exceptions include Lisa Su at AMD, Daniela Rus at MIT, and Fei-Fei Li. Women represent single-digit percentages of the AI industry overall. Greater balance in thinking is needed.
Recursive self-improvement is considered AI's holy grail, analogous to reusable rockets for the launch industry. The concept involves AI making itself smarter, then using that improved AI to create the next generation. Wo AI, a London-based startup founded by UC London graduates with researcher Zeng Yao Jang, published experimental evidence for recursive self-improvement through a system called AI-driven exploration squared. An outer AI agent rewrites code and research strategy for an inner AI agent. The claim is that eight days of machine self-improvement exceeded two years of expert human effort.
Defensive co-scaling is borrowed from human-to-human alignment. Rather than hoping for a perfect alignment algorithm, the approach ensures AI systems police other AI systems in proportion to capabilities—the way police forces scale with city populations. In the Wo AI experiment, the outer loop was tasked with improving the inner loop, which improved software development processes. The outer loop discovered it achieved better results by preventing the inner loop from cheating and reward hacking. The outer loop was policing the inner loop toward greater capabilities. Alignment capabilities are themselves capabilities in disguise.
Wo AI published a zero-to-three scale for recursive self-improvement:
- Level 0 (Delegation): AIs are slower than human R&D
- Level 1 (Net Positive): AIs beat human R&D at equivalent cost
- Level 2 (Ignition): Improvers are better improvers
- Level 3 (Inflection): Self-acceleration with fixed budget
The startup rates itself at Level 1 but shows sparks of ignition, representing a pre-ignition event.
Ramin Hasani distinguishes the engineering approach of inner and outer loops for code patches from true recursive self-improvement. The current frameworks assume fixed AI models with predetermined capabilities and no weight changes in neural networks. Changes occur only through system prompts rather than fundamental improvements to core competencies and behavior. True recursive self-improvement requires AI systems that can retune and adapt themselves similar to human learning processes.
The core competencies of current models remain fixed within certain thresholds. Existing frameworks represent an early stage engineering pipeline for improving work but do not constitute a breakthrough in the AI industry. Hasani notes that Liquid AI published a paper three years ago on automatic design of model architectures, designing self-improving meta AI systems that define their architectures and apply scaling laws across various architecture types.
At Liquid AI, the entire process of training foundation models and retuning system weights is being automated. This represents AI systems designing other AI systems, which Hasani describes as the holy grail of automatic model tuning. Current frameworks make it computationally intractable to train AI models, customize models, train on meaningful numbers of tokens for adaptation, change core architecture, or modify core learning algorithms.
Hasani references the Chinchilla scaling laws, which suggest that a two billion parameter model requires twenty times more tokens than parameters for compute optimality. Applying this to current frameworks attempting recursive self-improvement would require approximately 350 years to fine-tune a two billion parameter model. These nested learning systems and metalearning systems represent the core focus of Liquid AI's work over the past four years.
Dave notes the analogy between human learning, which occurs over approximately a twenty-year timescale from birth to adult capability, and recursive self-improvement, which operates on an evolutionary timescale of roughly ten million years. The major foundation model labs are all working on recursive self-improvement, which represents the most important moment in human history. This work requires big compute and big budgets rather than emerging from individual breakthroughs on consumer hardware.
Concerns exist about recursive self-improvement leading to hard takeoff scenarios and uncontrolled AI misaligned with human values. Hasani clarifies that recursive self-improvement is not limited to single agents but involves societies of agents where the defining structure for agent societies itself self-improves.
Hasani acknowledges real cybersecurity threats emerging from recursive self-improvement pipelines. Despite being pro-open source and releasing all Liquid AI models and science as open source, he recognizes that certain concerns around these systems are legitimate. This is why enterprises need self-control mechanisms before mass model releases, and why Anthropic has taken these issues seriously given their visibility into recursive self-improvement impacts.
At smaller scales, models can engage in reward hacking or avoid reward hacking to extreme degrees, pushing models to discover out-of-norm capabilities. These behaviors become more concerning as larger systems receive more compute resources.
Hasani predicts unbelievably capable models will emerge within the next two years, potentially exceeding human understanding, assuming compute continues growing at current rates without chip or memory shortages. He introduces the concept of depths of customization for categorizing model improvements at foundation model labs.
The shallowest level involves prompt engineering and code editing for engineering tasks like kernel optimization for inference speed. Fable 5 and Anthropic have shared examples of these superficial customizations. The next level involves models fine-tuning smaller language models to production-grade capabilities or optimizing their own performance through fine-tuning.
The deepest level, described as the holy grail, involves language models pre-training the next generation of themselves. This is why Anthropic hired Andrej Karpathy to work on pre-training automation, described as automation of automation.
The discussion identifies this as the most important story covered, representing self-accelerating systems that improve the process of searching, evaluating, and selecting improvements. The innovation loop compounds through meta-improvement rather than just task execution. This represents the first glimpse of the organizational singularity, where AI redesigns workflows rather than merely performing tasks within existing workflows.
Dave's analogy of a baby changing DNA is highlighted as particularly powerful for visualizing this transformation. The bar for meaningful improvement can be lower than full model-level recursive self-improvement, applying to workflow-level improvements like invoice approval processes.
Malaysian Prime Minister Anoir Ibrahim is preparing to debut an AI-generated digital double trained to sound like him for public communications and outreach. This represents one of the most prominent cases of a sitting head of government officially adopting an AI likeness as a sanctioned communications tool. Malaysia has 135 spoken languages, making this capability particularly valuable for reaching citizens.
This follows Albania's 2025 appointment of Dileia as an AI avatar serving as Minister of State for Artificial Intelligence through presidential decree, marking the first AI system named to a cabinet-level role. Albania's initiative aimed to fight corruption.
Sim notes the risk that authenticity could collapse if deep fakes proliferate, raising questions about whether communications come from the actual leader. However, watermarking could establish verified avatars that scale civic engagement, addressing one of democracy's biggest challenges worldwide.
Anwar Ibrahim is described as a good leader following on from the former prime minister Sim met while helping open a university in Malaysia. The initiative appears more focused on PR and connecting with ordinary citizens rather than primarily anti-corruption efforts.
Alex predicts this trend will extend to the West, particularly with high-alpha personality leaders wanting to amplify their reach. AI Trump is anticipated as an example. This generalizes social media from broadcast one-to-many communication to bidirectional interaction that scales beyond what individual leaders can manage directly.
Corporations and corporate CEOs are expected to adopt digital twins, with examples including Zuck and DAR's clone that employees used to practice pitches. Religious institutions are also creating interactive versions of founding documents and leadership figures.
A significant twist involves digital twins potentially becoming the actual leaders of organizations rather than serving as representatives. As digital twins interface more extensively with populations, they could effectively run companies instead of the behavioral origin. This pathway could lead to uploading entire organizations into cyberspace rather than just individuals.
Sam Altman has discussed the possibility of ChatGPT eventually becoming CEO of OpenAI if sufficient belief exists in what is being built. Dave confirms plans to create an AI Dave Blundon to run Link Studios and Link Ventures.
Peter has an existing AI avatar within the Abundance ecosystem. A friend in Calgary maintains a giant screen displaying Peter's AI avatar for tech employees to consult about moonshots. Peter suggests moonshots.com should feature AI avatars of all team members for AMAs.
The technology for creating such avatars has existed for a while, with Noam Shazier working on Character AI and Steve Brown creating a debate between AI Peter and Aristotle approximately two years ago. However, key talent has been absorbed back into big foundation labs rather than avatar development.
Dave references his mother's observation that John F. Kennedy defeated Richard Nixon because television rewarded visual presentation while Nixon used radio voice techniques. Each medium change represents a step function in reaching audiences. The current shift to AI avatars is underutilized but should become dominant within two years.
The compelling aspect is not exact replication but the differences and complementarities. AI versions can access any information in real time, generate visual graphs and charts, morph faces, teleport through space, point to atoms, and shrink or expand. These capabilities make the new medium exciting beyond simple cloning.
Peter emphasizes the importance of interviewing living parents and grandparents on video for hours at a time. This creates valuable AI representations for children, grandchildren, and great-grandchildren to understand their parentage and lineage. Peter has done this with his mother and regrets not doing more with his father.
The conversation pivots to discussing Liquid AI and small language models. Full disclosure is provided that Liquid AI is a company where Dave played an important early investor role. Daniela Roose at CEL, described as one of the three big shot women in AI running CEL at the MIT AI lab, called Dave about the best student she had ever had.
Daniela Rus merged labs together into one mega lab located in the new STA building, a distinctive crumpled-looking structure on the edge of MIT's campus. She runs the entire facility, which houses approximately 1,500 researchers and represents the biggest AI lab in the world. Rus called the speaker about a breakthrough from her best students, describing it as based on the nervous system of the worm C. elegans, which has 300 neurons.
Few foundation labs have truly rethought the transformer architecture from the ground up and started over. The transformer architecture and attention mechanism are widely recognized as bloated. Returning to founding principles could yield dramatically better systems. The team progressed from an idea in a lab to a billion-dollar valuation faster than any company out of MIT in history. Liquid AI is one of about 200 unicorns from MIT but the only foundation model company from MIT to reach unicorn status.
Ramin Hasani began his PhD in 2015 in Vienna under Professor Radu Grosu. The premise was that understanding human intelligence could begin with a smaller animal. The C. elegans worm has 302 neurons in its nervous system. Its body is transparent, allowing observation of neural activity. The organism has 78% genome similarity to humans and has contributed to four Nobel prizes. Neurons in the worm behave similarly to artificial neural networks, using graded potentials rather than spiking, which resembles analog computation.
The goal was to add more complexity to individual blocks of nervous systems to pack more information into smaller units of compute. Daniela Rus heard about the project through Radu Grosu and proposed applying the technology to autonomy and robotics, asking whether a handful of neurons could drive autonomous systems and whether the approach could scale to vehicles, drones, and jets. Hasani and co-founder Matias Lechner joined MIT's CSAIL in 2017.
The mathematics behind each neuron in liquid neural networks differs from attention mechanisms. These systems are based on recurrent neural networks and continuous-time processes. They were first applied to robotics because more information can be packed into smaller processors. Physical systems lack abundant compute; a robot typically has only a CPU, small GPU, and possibly an NPU or custom ASIC. The approach can deliver intelligence at the level of models 10 to 1,000 times larger while running directly on CPUs, GPUs, and NPUs outside data centers.
In 2020 and early 2023, Liquid AI built a foundation model lab around scaling the technology into the regime of foundation models, including large language models and small language models. At the time they started, no foundation model labs existed apart from DeepMind and OpenAI. The team aimed to explore alternative architectures starting from priors drawn from nature and to build a meta AI system—an automated AI system that designs AI and explores computational graphs of intelligence beyond the transformer.
Scaling laws involve starting with smaller models, training them on a fixed token budget and compute, then systematically increasing model size. Larger models with greater token budgets produce more capable systems. Along the scaling curve, smaller instantiations appear. Liquid AI's mission is building efficient general-purpose AI at every scale, treating efficiency as a first-class citizen. Small language models (SLMs) lie along the scaling curve but lack the full general capability of the largest models; they can be specialized for dedicated problems while remaining general-purpose in understanding language, vision, and audio in multimodal formats.
Small typically refers to anything below 100 billion parameters. On-device AI refers to models that can be deployed on actual physical devices. Liquid AI announced a significant deal with Mercedes focused on energy-efficient, fast, offline SLMs.
Automotive environments contain limited compute, often a single Qualcomm or Samsung chip for infotainment and in-car intelligence with only 2–8 GB of RAM. The model must remain very small yet performant. The goal is to create a private space inside the car that powers intelligence without relying on the cloud, since connectivity is not always available and conversations should remain private. Liquid AI deployed a multimodal foundation model under one gigabyte in size that runs on chips costing as little as $60.
Liquid AI operates as a B2B enterprise company. Fine-tuning small models is relatively inexpensive and tangible. They fine-tune models for specific in-car applications and maintain data flywheel systems so the intelligence remains adaptable. The Mercedes rollout begins in North America this year with over-the-air updates for Mercedes-Benz vehicles from 2022 onward; each update is approximately 600 megabytes. Future updates using LoRA adapters and other techniques could shrink to around 20 megabytes, enabling recursive improvement of the user experience.
The model sits below the operating system and has access to roughly 700–1,200 functions inside the car. Users can talk to the car, control panels, request manuals, and engage in conversations. Memory features are being added. The system also has access to all apps through function calls, allowing control of the entire ecosystem on top of the car's operating system.
SLMs offer smaller size, lower energy consumption, and the ability to run on-prem. To address whether models have sufficient internal capabilities to answer questions accurately, Liquid AI uses a framework of foundation model development based on depths of customization. Prompt engineering may suffice in some cases; fine-tuning is used when needed; pre-training may be required for core capabilities or specialization. Their platforms automatically identify the required depth of customization.
Liquid AI sells enterprises the Model+ platform, which enables customization rather than static model weights. In silicon medicine and biotech, liquid foundation models have proven highly adaptable. The company has developed state-of-the-art biotech and longevity foundation models. With Shopify, Liquid AI handles one billion requests; the models have been in production for six months and serve hundreds of millions of users across 10 billion products. Additional partnerships include AMD and other chip manufacturers to bring low-code AI experiences to PCs. The company's focus is bringing intelligence outside of data centers.
Liquid AI originated from a neuromorphic premise based on studying C. elegans. Liquid neural networks in their original form are among the most expressive computational formats, featuring nested nonlinearities and physics-inspired dynamics. They handle irregularly sampled data via neural ODEs and belong to the broader class of continuous-time neural networks. When scaling, recurrences are sometimes linearized, producing state-space models and variants that fall into continuous-time neural networks reduced to linear dynamical systems.
Variants of linear attention and gated linear attention have emerged. A special input-dependent gating mechanism, inspired by how neurons exchange information, adds expressivity and remains present in many architectures including Liquid's. From day one, Liquid AI avoided biasing toward a single architecture. They designed a search algorithm called STAR (automated design of tailored architectures) that runs scaling laws across approximately 100 different variations of operations that could yield a general-purpose computer. The STAR paper was published two and a half years ago.
Ramin Hasani discussed criteria for optimizing neural architectures across four dimensions: memory consumption on processors, computational efficiency and speed, operational latency, and maintaining accuracy. Pure transformer models exist alongside hybrid models that incorporate some transformer components while modifying the rest of the dynamical system to optimize these four objectives. The framework can automate architecture design for foundation models in-house through a technology stack called automated foundation model design algorithms, referred to as AFMD. This automated framework explores architectures for specific hardware configurations.
The first generation of architectures optimized through this search produced double gated convolution mechanisms comprising 80% of the network. When running without human bias, the gating mechanism from the original liquid foundation liquid neural networks paper appeared in very similar format in the final architecture emerging from the search space.
The health section of the Moonshots podcast, brought to you by Fountain Life, addressed AI applications for healthy aging and reaching 100 plus years. Dr. Don Mucalem, chief medical officer of Fountain Life, noted that the number one concern among members is losing brain health and developing dementia. Conservative estimates indicate 45% of dementia cases are entirely preventable.
With advanced testing at Fountain Life, one quarter of members showed advanced brain age. When partnered with healthy living interventions including eating healthier, moving bodies, and optimizing sleep, brain age improved by 26% in the majority of individuals. Fountain Life searches globally for the best therapeutics and approaches to bring to members, directing interested individuals to fountainlife.com/per to become CEO of their own health.
Palmer Luckey, founder of Oculus and chairman of defense company Anduril, argued that the modern patent system has become a national security liability. He stated the entire patent office could be downloaded every morning, ripped off, and used to fight a war. The core problem stems from patent disclosure requirements: to obtain a patent, inventors must teach a person skilled in the art how to create and use the device in full, clear, concise, and exact terms.
This disclosure exchanges 20 years of exclusivity for teaching the world the invention. When strategic adversaries harvest files and ignore legal protections, they receive free instruction manuals. The US Patent Office receives about 600,000 applications annually and grants a little over half, totaling 323,000 patents, with 2025 data showing a 40% increase in granted patents over the last five years, potentially secondary to AI.
Luckey's proposed fix involves massively scaling the national security patent process under the Secrecy Act of 1951. This mechanism allows inventors to obtain classified patents maintaining exclusive rights without disclosure to anyone, including the government. Currently approximately 6,000 secure secrecy orders are active in the US. Luckey wants this edge case converted to a default mechanism.
Stop patenting everything. Patents are Chinese instruction manual. Well, the founding fathers never predicted a world where you would have a globalized economy where the entire patent office could be downloaded every single morning and then ripped off and then used to fight a war against you. We need to really fundamentally revisit the patent system. I think we need to massively expand the national security patent process.
Alex argued the Invention Secrecy Act of 1951 has been detrimental on balance, not just to democracy. Under the act, inventions cannot simply be filed secretly and practiced under confidentiality; rather, the military exclusively practices the invention and the inventor receives royalties. This creates secret monopolies where entire swaths of transformative technology from energy to other domains may have been confiscated by the Department of Defense for purely military reasons without general knowledge. Alex suggested the Invention Secrecy Act regime should probably be eliminated rather than expanded.
The podcast noted Palmer Luckey will join the Moonshots gathering on September 25th in LA at moonshots.com. The value of American open innovation policy, where people build on others' creations, was highlighted as making America great.
One perspective emphasized that disclosure problems have grown over 20 to 30 years, especially in an AI age where people can route around, replicate, or learn from patents. The real defensibility lies in learning loops and creating trade secrets around proprietary feedback loops, with continuous innovation as the winning defense rather than ownership. The only consistent winners in the current model are lawyers.
Dave suggested a potential World War III flashpoint around intellectual property. He noted that Europe respects intellectual property rights creating coherent economies, while China ignores them. The likely outcome involves the US trading embargoing countries that do not respect intellectual property rights, forcing choice between the free world and an alternate world within the next couple of years.
Dave argued that no science fiction depicts massive AI-created intellectual property without mechanisms for innovators to profit, and without such mechanisms, foundational technology development suffers. A huge fraction of brilliant thinkers from Cambridge, MIT, and Harvard avoid foundational technologies due to lack of financial incentive, which needs fundamental change. Protecting intellectual property rights represents a key way to reverse this tide.
Y argued Palmer Luckey misunderstands patents fundamentally. The patent bargain involves disclosure in return for temporary state-granted monopoly, and complaints about Chinese appropriation represent a quibble with enforcement rather than justification for abandoning the disclosure-monopoly trade. The appropriate ask is better enforcement of US patents in China rather than discarding the patent system.
Two stories demonstrated AI impact on healthcare abundance through demonetizing and democratizing diagnostics. GPT 5.6 saw, released weeks prior, set a new all-time high on HealthBench Professional, OpenAI's hardest medical benchmark. In a blind test across roughly 20,000 individual physician judgments comparing GPT 5.6 answers to specialty-matched physicians given unlimited web access and unlimited time, the AI outperformed doctors on accuracy, safety, and completeness.
Meta's Muse Spark 1.1, released the prior week, beat GPT 5.6 on OpenAI's HealthBench Professional benchmark of 525 real clinical tasks while being 7 times cheaper. Muse Spark is free inside all Meta products including WhatsApp and Facebook, reaching 3.56 billion daily active users. Top medical AI capabilities are now free to over 3.5 billion people. A model in China is already used in rural environments by 100 million people.
The healthcare domain presents moral urgency for abundance because better first-line answers can be delivered at near-zero cost, making safe deployment speed the only relevant question. Almost every country faces radical doctor shortages.
Instagram now gives better medical advice than human doctors, representing a July 2026 story. Cost of medical intelligence is becoming too cheap to meter and essentially free. While acknowledging possible mild benchmaxing by Meta on Muse Spark 1.1, which sits on the optimal cost frontier but not at the absolute top, the achievement of Instagram providing superior medical advice remains significant.
Breaking news from the previous day covered a longevity advance addressing advanced glycation end products, or AGEs, which are sugar molecules that cross-link and damage proteins over time through glycation. This process stiffens arteries, clouds lenses with cataracts, damages kidneys, and wrinkles skin, previously considered irreversible.
A team from Revel Pharmaceuticals published in Nature Communications demonstrating an engineered enzyme called CMLA that acts as a molecular lawn mower, oxidizing away glycation scars and restoring original healthy protein. The enzyme worked in human tissue samples from elderly donors, reversing lifetime-accumulated damage. This makes a previously permanent aging category reversible.
Calico, the California Life Company and Alphabet other bet, was involved alongside Revel. The broader process involves Maillard reactions, the same chemical class responsible for bread crust browning and why foods taste like chicken, where sugars react with carbonyl groups and amines in proteins creating optically brown molecules.
The significance lies in approaching unscrambling eggs through directed evolution of a bacterial protein to remove unwanted sugar-protein byproducts associated with inflammation and aging correlates. This represents evolution by human direction rather than natural selection, with potential for longevity escape velocity through mining bacterial innovations.
Ramin Hasani received congratulations on Liquid AI success. Dave was acknowledged for investing in Ramin. The hosts expressed gratitude for the discussion and excitement for Starship 13 launch, wishing Elon and the team luck. The speed of the singularity waits for nobody, suggesting a potential emergency pod soon.
Keep Peter H. Diamandis in your library
Save the videos and channels worth coming back to, and find them again in one place.
