
Jul 17, 2026 · 1h 57m
Tech leaders debate regulatory capture and the rise of on-device AI
Mira Murati's 975B Open Model, Ramin Hasani on Post-Transformer AI, and Demis' AI FINRA | EP #271
The episode unpacks how the intersection of corporate lobbying, geopolitical rivalry, and architectural efficiency will decide who controls the next wave of artificial intelligence.
- 1Proposals for centralized AI regulatory bodies risk enabling regulatory capture that protects tech giants from open-source rivals.
- 2Small language models are delivering energy-efficient, on-device intelligence for major enterprises without requiring massive cloud infrastructure.
- 3The integration of AI digital twins in government communications signals a shift toward automated public representation.
Don't miss
Ramin Hasani explains how Liquid AI transitioned from analyzing the nervous system of a tiny worm at MIT to building next-generation on-device AI models.
The brief
As artificial intelligence advances, the battle lines are shifting from raw computational scale to strategic control, energy-efficient architecture, and the geopolitical rules of engagement between global superpowers.
Industry giants are pushing for strict regulatory frameworks, but critics warn that proposals like a FINRA-style standards body might just be strategic regulatory capture designed to lock out open-source competitors.
Liquid AI co-founder Ramin Hasani joins the panel to explain how bio-inspired small language models, originally rooted in MIT research on roundworms, are bringing highly efficient, on-device intelligence to enterprises.
Meanwhile, governments are stepping into the future as Malaysia adopts an official digital twin of its Prime Minister, signaling a new era of interactive, AI-generated public communication and statecraft.
From recursive self-improvement to national security patent reforms, the future of tech sovereignty depends on balancing open innovation with the protection of critical intellectual property.
What was said on this episode
96 statements · 50 positive · 31 negative · 1 mixed · 14 neutral
Mira Murati’s model meets current market needs.
“she's built exactly the thing hitting the market. That exactly what everybody needs right now.”
Listen at 0:11
Liquid AI aims to build efficient general-purpose AI across scales.
“building efficient general purpose AI at every scale”
Listen at 0:23
Liquid AI seeks frontier-level intelligence running on CPUs.
“brings the same level of intelligence than a frontier model into, let's say on a CPU”
Listen at 0:36
AI regulation would serve the public interest.
“there's a strong consensus there should be some AI regulation”
Listen at 5:37
- Ramin Hasanion Independent AI regulatory agencyNeutral6:10
An independent AI regulatory agency will eventually emerge.
“some sort of AI regulatory agency that stands on its own, similar to the FAA or FCC is likely at some point”
Listen at 6:10
- Ramin Hasanion AI safety risksNegative6:26
Severe AI risks justify proactive safety measures.
“the consequences of AI going wrong are severe. So we have to be proactive rather than reactive.”
Listen at 6:26
- Ramin Hasanion AI systemsNegative6:26
AI failures could have severe consequences.
“the consequences of AI going wrong are severe”
Listen at 6:26
Traditional government bureaucracy cannot regulate AI effectively.
“AI moves way, way too fast for any kind of traditional government bureaucracy”
Listen at 11:00
Effective AI governance requires standards, real-time audits, and open evaluations.
“you're going to need a standards body, you're going to need real time audits and you're going to need open evaluation suites”
Listen at 11:06
AI systems can assist with their own regulation.
“AI can help regulate itself”
Listen at 12:22
A FINRA-like frontier-AI regulator is a bad idea.
“I tend to think this is a bad idea”
Listen at 12:45
A FINRA-style frontier AI regulator risks regulatory capture.
“I tend to think this is a bad idea. It smells like regulatory capture.”
Listen at 12:45
Demis Hassabis’s proposal could form a cartel of frontier AI labs.
“the attempted formation by demos of a cartel of frontier labs”
Listen at 12:50
AI governance should regulate harmful actions rather than model capabilities.
“regulating the actions that in at least the western legal canon that we do do and that I'd be much more supportive of”
Listen at 15:57
Frontier labs’ regulatory push may create competitive moats.
“I worry that it's more regulatory capture and creating moats for themselves”
Listen at 16:25
Static regulation cannot keep pace with rapidly advancing AI.
“I see no mechanism to regulate AIs moving way too quickly. Any regulatory is static.”
Listen at 17:20
No existing mechanism can regulate rapidly advancing AI effectively.
“I see no mechanism to regulate AIs moving way too quickly”
Listen at 17:20
A frontier AI regulatory body will be created eventually.
“I don't think it's a matter of if, it's only a matter of when and what the structure will be.”
Listen at 18:57
A US AI regulatory body will appear before year-end.
“My guess is we see this before the end of the year”
Listen at 19:58
The proposed U.S. open-model ceiling could incentivize China to advance faster.
“this creates the perverse incentive to let China win the race to ever greater superintelligence”
Listen at 21:28
AI policy should shift from prevention toward adaptation.
“You have to move from, like, prevention and whatever to adaptation”
Listen at 22:10
- Ramin Hasanion AI model developmentNegative22:44
Faster model development makes capability constraints increasingly difficult to impose.
“the pace of model development is getting so much smaller that is becoming exponentially more difficult to really impose any of these type of constraints”
Listen at 22:44
Restrictive Western AI regulation could drive leading researchers to China.
“the best Western researchers move to China to escape this regulatory framework”
Listen at 23:40
Open ecosystems consistently outperform closed ecosystems.
“open ecosystems always win”
Listen at 24:32
Fixed AI benchmarks could distort future model capabilities.
“we freeze in or otherwise lock in the benchmarks for how we measure capabilities”
Listen at 25:10
Adaptability will matter more than raw model power.
“it's not going to be the future is the raw power. It's going to be the adaptability that's going to win”
Listen at 27:42
Organizations will increasingly own and run models on-premises.
“people Owning their own models, working on prem and not giving their controls to the large frontier models”
Listen at 27:51
Thinking Machines’ model is stronger than Nvidia’s Nemotron according to released evaluations.
“it's stronger than Nematron”
Listen at 28:15
Thinking Machines’ model is weaker than GLM 5.2.
“the west now has a new frontier open weight model. It's weaker than GLM 5.2”
Listen at 28:36
Thinking Machines’ model is weaker than GLM 5.2 according to released evaluations.
“it's weaker than GLM 5.2”
Listen at 28:39
Western incentives have discouraged strong open-weight model releases.
“the west has been poorly incentivized to release strong open weight models”
Listen at 29:16
Western competition can create new monetization paths for open-weight models.
“this is very helpful to finally have enough competition in the west that's creating ways to monetize open weight models other than just per token sales”
Listen at 30:20
Companies may use open releases before switching breakthrough models to closed APIs.
“when you have something really groundbreaking, then you go closed source and you launch an API into corporate America”
Listen at 31:13
- Ramin Hasanion AI model customization marketNegative31:58
The AI customization market remains substantially underserved.
“customization has been like something that everything the whole market around customization has been very empty”
Listen at 31:58
- Ramin Hasanion AI model layerNegative32:21
AI model layers no longer offer the main source of enterprise value.
“the model layer is not anymore the place where you can actually extract value”
Listen at 32:21
Reinforcement fine-tuning may eventually become obsolete.
“there's an alternative vision where RFT just dies”
Listen at 34:14
Organizations will increasingly use owned, fine-tuned open-weight models locally.
“people want to use fine tuned open weight models that they own on their own hardware”
Listen at 34:46
Local fine-tuning of in-house models substantially expands enterprise AI capability.
“the ability to bring the model in house and fine tune it with your local data is a huge unlock”
Listen at 37:38
Small multimodal models are designed for organizational applications.
“this is aiming squarely at organizational use”
Listen at 38:08
On-premises proprietary-data AI will become the dominant future model.
“that will be the future”
Listen at 38:22
Conventional fine-tuning historically did not increase model capabilities.
“fine tuning didn't have a history of increasing capabilities”
Listen at 39:35
Reinforcement fine-tuning can increase model capabilities.
“fine tuning actually start to increase the capabilities of the models”
Listen at 40:05
Reinforcement fine-tuning is currently the leading scaling approach.
“I think probably RFT is the scaling of the moment”
Listen at 40:48
Recursive self-improvement is AI’s central long-term objective.
“the holy grail of AI is recursive self improvement”
Listen at 43:28
AI systems should police other AI systems proportionally to their capabilities.
“the solution for AI safety is AI policing AI in proportion”
Listen at 45:29
WECO’s outer loop polices the inner loop while improving its capabilities.
“the outer loop is defensively co scaling with and policing the inner loop”
Listen at 46:49
WECO’s system is beginning to approach recursive self-improvement ignition.
“they are touching, just starting to touch on ignition”
Listen at 48:35
- Ramin Hasanion WECO AI recursive self-improvement resultNegative48:59
WECO’s recursive self-improvement result is less significant than claimed.
“I'm not as excited as Alex is on the topic”
Listen at 48:59
- Ramin Hasanion WECO’s recursive self-improvement systemNeutral49:48
WECO’s reported recursive system does not change neural-network weights.
“there is no weight changes in the neural networks”
Listen at 49:48
- Ramin Hasanion Current AI modelsNeutral50:31
Current AI models have fixed core competencies because their weights remain fixed.
“the core competences of these models that we have right now, they're, they're, they're fixed weight models”
Listen at 50:31
Liquid AI is automating AI systems that design other AI systems.
“we are talking about AIs or designing AIs”
Listen at 51:36
- Ramin Hasanion WECO recursive self-improvement frameworkNegative53:10
WECO’s framework would require 350 years to fine-tune a two-billion-parameter model.
“It takes us 350 years to really fine tune a 2 billion parameter model with this framework.”
Listen at 53:10
Recursive self-improvement is currently limited to well-funded, compute-rich labs.
“it's happening, but it's happening with big compute and big budgets”
Listen at 54:42
Recursive self-improvement currently requires substantial compute and funding.
“it's happening with big compute and big budgets”
Listen at 54:43
Recursive self-improvement pipelines create real cybersecurity threats.
“the cybersecurity kind of threads that we are seeing, like coming out of these type of pipelines of recursive self improvement. They're real”
Listen at 56:15
- Ramin Hasanion Recursive self-improvement cybersecurity risksNegative56:43
Cybersecurity risks from recursive self-improvement are real.
“some of these concerns that Peter you brought up, they're very real”
Listen at 56:43
- Ramin Hasanion AI model release safety checksPositive56:59
AI companies should perform self-checks before mass model releases.
“before mass release of their models there has to be always a certain degree of self check”
Listen at 56:59
AI models will become extraordinarily capable within approximately two years.
“you're going to see unbelievably kind of models probably in the next two years or so”
Listen at 58:24
WECO’s workflow-level improvement demonstrates the organizational singularity’s first glimpse.
“this is the first glimpse of the organizational singularity”
Listen at 1:01:56
Official AI replicas of leaders could undermine perceived authenticity.
“The risk is that the authenticity kind of collapses”
Listen at 1:04:00
Watermarked official AI avatars could expand civic engagement.
“if you can do it from a, with a watermarking or something and say this is the actual avatar, then it gives every citizen a voice to plug into and gives huge props to the civics of all of this”
Listen at 1:04:15
Watermarked leader avatars could scale civic engagement.
“this allows you to scale that”
Listen at 1:04:40
Western governments and leaders will increasingly adopt AI digital twins.
“we're going to see more of this in the west as well”
Listen at 1:05:20
Corporations and CEOs will likely adopt AI digital twins.
“I think it's likely that corporations, corporate CEOs will do this”
Listen at 1:06:15
Digital twins may eventually run organizations instead of human leaders.
“the digital twin of, of the leader running the company and not the actual behavioral origin that's running the company”
Listen at 1:07:22
AI avatars should become dominant in the next election within two years.
“it should be easily dominant two years from now in the next election”
Listen at 1:10:35
Transformer attention mechanisms are excessively inefficient.
“the whole attention mechanism is bloated”
Listen at 1:13:45
Reconsidering transformer architecture could produce substantially better AI systems.
“you might be able to build something dramatically, like massively better”
Listen at 1:13:51
Liquid models can match larger models’ intelligence while using far fewer parameters.
“deliver basically intelligence at the level of models that are 10 to 1,000 times larger than themselves cells”
Listen at 1:18:42
Liquid AI’s automated system searches for architectures beyond transformers.
“an AI that designs AI that explores the computational graphs of intelligence beyond transformer”
Listen at 1:19:59
- Ramin Hasanion AI model scalingPositive1:20:52
Larger models and training budgets generally produce greater intelligence.
“the larger you make the models, the more token budgets you spend, the more intelligence of a system you can get”
Listen at 1:20:52
- Ramin Hasanion Small language modelsNeutral1:22:00
Smaller language models are typically specialized for dedicated problems.
“You usually specialize smaller language models.”
Listen at 1:22:00
- Ramin Hasanion Small language modelsNeutral1:22:19
Models below 100 billion parameters fall into the small-to-medium range.
“anything below 100 billion parameter is something that is not small and medium sized kind of models”
Listen at 1:22:19
- Ramin Hasanion Automotive AI modelsNeutral1:23:44
Automotive AI models must be compact while retaining performance.
“the model has to be very small and at the same time being able to perform”
Listen at 1:23:44
- Ramin Hasanion Cloud-hosted AI in vehiclesNegative1:23:57
Cloud-hosted AI is unsuitable for safety-critical vehicle control.
“You don't want your car to be driven by an AI model that is sitting in the cloud.”
Listen at 1:23:57
Liquid AI has a multimodal model under one gigabyte that runs on vehicle chips.
“we brought one of our multimodal foundation models that is only less than 1 gigabyte in size and it can go inside the car's chip”
Listen at 1:24:16
- Ramin Hasanion Liquid AI automotive modelPositive1:24:18
Liquid AI’s multimodal model can run inside a car using less than one gigabyte.
“one of our multimodal foundation models that is only less than 1 gigabyte in size and it can go inside the car's chip”
Listen at 1:24:18
Small adapter updates could recursively improve in-car user experiences.
“you would be able to have like a recursively kind of improving the experience of the user as well”
Listen at 1:26:59
- Ramin Hasanion Liquid AI in-car systemPositive1:27:40
Liquid AI’s in-car system can access approximately 700 vehicle functions.
“there are 700 functions inside the car”
Listen at 1:27:40
Liquid AI uses automated architecture search rather than relying solely on human architectural choices.
“we designed a search algorithm to let's say like, you know, let the algorithm, instead of human biasing kind of the algorithm”
Listen at 1:35:31
Liquid AI’s architecture search produced networks that are 80% double-gated convolution mechanisms.
“It came double gated convolution kind of mechanisms. As 80% of the network being this.”
Listen at 1:37:26
Expanding classified patent secrecy would be harmful.
“I think this is a terrible idea.”
Listen at 1:43:07
The Invention Secrecy Act regime should probably be abolished.
“the Invention Secrecy act regime should probably go away”
Listen at 1:44:55
Proprietary learning loops will provide stronger defenses than patents.
“The real moat is learning loops. That's going to be the real defensibilities.”
Listen at 1:45:41
Proprietary learning loops will provide stronger defensibility than patents.
“The real moat is learning loops”
Listen at 1:45:41
The U.S. may impose trade embargoes on countries that disregard intellectual property rights.
“the US will trade embargo. Anybody who doesn't respect intellectual property rights.”
Listen at 1:46:35
U.S. IP-related trade restrictions may occur within several years.
“that's going to happen soon, like in the next couple of years”
Listen at 1:46:49
The U.S. should seek stronger patent enforcement in China.
“what he should be asking is better enforcement of US patents in China”
Listen at 1:48:02
Meta’s medical AI capabilities are available free to over 3.5 billion people.
“top medical AI capabilities are now free to over three and a half billion people on the planet”
Listen at 1:49:58
AI has dramatically reduced the cost of medical diagnosis.
“diagnosis is, has had massive cost collapse”
Listen at 1:50:32
Instagram’s AI can provide better medical advice than human doctors.
“Instagram now gives better medical advice than a human doctor”
Listen at 1:51:07
Meta may have optimized its medical benchmark performance.
“I suspect a little bit of mild benchmaxing by Meta on this”
Listen at 1:51:31
CMLase reversed accumulated glycation damage in elderly human tissue samples.
“reversing damage that accumulated over the lifetime”
Listen at 1:53:24
CMLase research suggests some age-related glycation damage is reversible.
“A category in aging that we've always filed as permanent just became reversible.”
Listen at 1:53:31
Human cells perform approximately five billion chemical reactions per second.
“our ability to understand the 5 billion chemical reactions per second per cell”
Listen at 1:53:43
Engineered enzymes could substantially advance longevity escape velocity.
“Great potential for longevity, escape velocity.”
Listen at 1:55:42
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Books & mentions
Moonshots with Peter Diamandis
Explore cutting-edge discussions on exponential technologies and longevity science.
Weekly email
State of the Pod, every Monday
The week's most-discussed names and stories from podcasts, with the moments behind them. Free, five minutes, unsubscribe in one click.