Murati's Open Model and AI's Regulation Dilemma

Three things happened in the same week that, taken together, tell you something important about where AI power is actually concentrating — and who's trying to control it.

Mira Murati shipped her first model. Demis Hassabis published an essay calling for a FINRA-style regulatory body for frontier AI. And a London-based startup claimed experimental evidence for recursive self-improvement. On the Moonshots podcast, Peter Diamandis and his panel — investor Dave Blundin, computer scientist Alexander Wissner-Gross, Salim Ismail, and Liquid AI CEO Ramin Hasani — dug into all three. The conversation was more candid than most, and the tensions it surfaced are worth sitting with.

The Regulation Question Nobody Can Answer Cleanly

Start with the governance story, because it frames everything else. Hassabis, in an essay titled A Framework for Frontier AI and the Dawning of a New Age, called for a US-led standards body modeled on FINRA — the industry-funded, self-regulatory organization that polices Wall Street under SEC oversight. He reportedly wants something operational before the end of the year. Elon Musk has been predicting a standalone AI regulator for years. Sam Altman published a framework proposal in the Financial Times the week prior.

When three of the largest frontier labs are all, simultaneously, asking to be regulated, it's reasonable to wonder what they're actually asking for.

Wissner-Gross was direct: "I worry that it's more regulatory capture and creating moats for themselves in a hypercompetitive landscape." His concern isn't just that incumbents write the rules in their favor — it's a more specific fear that any ratified benchmark system for measuring "frontier" capabilities would distort the entire field. Labs would optimize for whatever gets measured, ignore what doesn't, and the result would be a stunted, benchmark-gamed version of AI development frozen around whoever happened to be at the table when the standards were set.

Ismail called it simply "not workable," pointing to the deeper problem that non-state actors won't comply, and any static regulation is outdated the moment it passes.

Hasani offered the most structurally interesting take. He framed regulation as a game-theoretic problem — a Stackelberg game, specifically, where policymakers move first and agents react. The challenge is that the frequency gap between regulatory action and AI development is enormous. By the time a rule is written, the thing it's regulating has already changed. His preference: capability-based thresholds that differ by vertical, since the governance needs for an on-device model in a car are genuinely different from those for a model processing financial data.

Blundin made a practical point about the FINRA analogy's limits: at a frontier AI lab, you can't ask your best researchers to go work at a regulatory body for two years the way a bank executive might do a stint at FINRA. The talent economics don't work. And unlike FINRA, where regulation is fundamentally a human conversation, AI regulation could theoretically involve AI — which changes the whole structure of the problem.

The regulatory challenges here aren't unique to AI, but the velocity is. The honest summary from the panel: everyone agrees something needs to exist, nobody agrees on what, and the window before it becomes moot may be shorter than the legislative timeline.

The US-China Ceiling That Nobody Likes

Woven into the regulation discussion is a reported White House proposal that the panel found genuinely strange: pegging the permissible ceiling for US open-weight model releases to the capability level of China's best open-weight model. The logic is that DeepSeek and similar models have already been downloaded millions of times globally — you can't un-ship them. So rather than a ban, define a permissible range.

Wissner-Gross called this "the moral equivalent of throwing the steering wheel out the window in a game of chicken." The perverse incentive is obvious: it gives Western labs a reason to want China to release more capable open-weight models faster, so Western releases can follow. That's not a policy. That's an inversion.

The broader point — made by multiple panelists — is that open ecosystems have historically won, and the US's competitive advantage has always been permissionless innovation. Capping that advantage to China's pace of release is, as one panelist put it, "the weirdest strategically bizarre thing we've ever seen."

What Murati's Inkling Actually Is

The model release that generated the most energy in the conversation was Inkling, from Mira Murati's startup Thinking Machine Labs. It's a mixture-of-experts architecture with 975 billion total parameters, activating 41 billion at inference time — which keeps it fast and runnable on reasonable hardware. It was trained on text, image, audio, and video, and reasons natively across all four modalities. It can be downloaded, fine-tuned, and run on-premises.

Murati's stated bet is deliberately contrarian: she's not claiming Inkling is the best model on any leaderboard. Her own blog says as much. The pitch is customizability — that enterprises should own their models, train them on proprietary data, and not route their most sensitive information through a foundation lab's API.

The panel evaluated this clearly. According to evals that Thinking Machine Labs released, Inkling outperforms Nvidia's Nemotron on their benchmarks, which makes it one of the stronger Western open-weight models available. It's weaker than the leading Chinese open-weight models and obviously behind closed Western frontier models. It's not the best model in the world. That's not what it's trying to be.

Hasani read the business logic astutely: "They're deliberately leaving some room for fine-tuning so that people can come in and using their business API... that would be even printing money at a larger speed on the customization side." The model layer, especially below the closed frontier, is no longer where you extract value. The value is in the customization pipeline. Thinking Machine Labs is betting that reinforcement fine-tuning as a service — letting enterprises tune Inkling on their own data, in their own environment — is where the money is.

Wissner-Gross noted the countervailing risk: that's a bet on the permanence of the reinforcement fine-tuning paradigm, which may not hold. If base models become generalist enough that fine-tuning yields no capability gains, the whole business model deflates. "I think probably RFT is the scaling of the moment, but in the future I can totally imagine a generalist-based model that is just so generally capable that it doesn't actually benefit from any further reinforcement fine-tuning."

The open-weight angle matters beyond Thinking Machine Labs. The question of AI optimism and pessimism running through the broader policy debate is really a question about who gets to develop AI at all — and open-weight models are the mechanism by which that access gets distributed, or doesn't.

The Architecture Question, and the Worm

Hasani's own company, Liquid AI, represents the other thread worth tracking: what comes after the transformer.

The origin story is unusual. Hasani's PhD research, conducted at MIT's CSAIL under Daniela Rus, focused on the nervous system of C. elegans — a 302-neuron worm whose neural architecture, due to its transparency and genetic similarity to more complex organisms, has been an important model organism in biology. The insight was that the mathematical structure governing how those neurons process information over time could be applied to build neural networks with fundamentally different properties than transformer-based attention mechanisms: more efficient, more adaptable, capable of running on far smaller hardware.

Liquid AI's mission, as Hasani described it, is building "efficient general-purpose AI at every scale that explores the computational graphs of intelligence beyond transformer." The practical expression of that is models running on CPUs, on-device, embedded in automotive systems — a Mercedes partnership being one example — without the compute and memory footprint that transformer-based models require.

This matters for the governance and openness conversation. A model architecture that can run inference on a CPU changes who can deploy AI, where it can run, and how much it costs. The whole discussion about on-premises deployment, data sovereignty, and enterprise customization becomes more tractable if the model doesn't require a server rack.

The RSI Question

The panel also took up a claim from London-based startup Evo AI (referred to as Wo AI in the discussion) about "experimental evidence for the first recursive self-improvement." Their system — an outer AI loop tasked with rewriting the research strategy and code of an inner AI loop — reportedly showed that 8 days of machine self-improvement beat two years of expert human effort on a benchmark.

Hasani was measured in his response. The work is impressive engineering, he said, but it doesn't meet the definition of recursive self-improvement as the major labs understand it — because the model weights aren't changing. The inner loop is being optimized via code edits and system prompt changes, not by actually retraining the neural network. The models are fixed-weight. "For us, recursive self-improvement means that you have an AI system or an army of AI systems that they can also retune themselves — they can adapt very similar to how humans do it."

He ran the math: using Chinchilla scaling laws as a rough guide, the computational cost of actually fine-tuning a two-billion parameter model via the framework Evo AI described would take approximately 350 years. The nested optimization problem is computationally intractable at meaningful scale with current approaches.

Wissner-Gross was more optimistic about the symbolic significance — he found the emergent self-policing behavior of the outer loop genuinely interesting as a case study in what he calls "defensive co-scaling": the idea that aligned AI behavior emerges from AI policing AI, rather than from any single alignment algorithm.

Hasani's two-year timeline for models "going above our understanding" wasn't a dismissal of the field — it was a calibration. The work is happening at the frontier labs, at scale, with compute that makes the Evo AI experiment look like a classroom demo. It just isn't being announced.

The more interesting question may not be when we'll have recursive self-improvement, but what governance structure — if any — will exist when we do. Right now, those two timelines appear to be running at very different speeds.

---

Dev Kapoor covers open source software, developer communities, and the politics of code for Buzzrag.