KM4BI KM4BI

Mira Murati’s 975B Open Model, Ramin Hasani on Post-Transformer AI, and Demis’ AI FINRA

Moonshots connects Inkling, model customization, Liquid AI’s small-model thesis, and the governance challenge around frontier AI.

This Moonshots episode frames the current AI race around three tensions: how to govern frontier models, how open-weight releases can become durable businesses, and whether the next architectural leap comes from systems that are smaller and more efficient rather than simply larger.

Inkling and the customization thesis

Thinking Machines’ Inkling is discussed less as a leaderboard play and more as a bet on enterprise customization. The argument is that companies increasingly need models they can adapt to proprietary data, workflows, environments, and performance constraints.

The open-model path still raises a business-model question. A strong open release can build credibility and adoption, but monetization may depend on fine-tuning services, a broader adaptation platform, or eventually a closed API for the most differentiated capabilities.

Governance that can keep up

The episode treats a FINRA-like frontier AI standards body as directionally interesting but difficult to implement. Static law will age quickly. A useful framework would need real-time audits, open evaluation suites, and mechanisms that update as model capabilities change.

A separate proposal to peg U.S. open-release limits to the best Chinese open-weight models is presented as especially awkward: it could let China indirectly set the ceiling for what American labs can publish.

Liquid AI and small language models

Ramin Hasani describes Liquid AI’s mission as building efficient, general-purpose AI at every scale by exploring computational graphs beyond transformers. Small language models are not just shrunken LLMs; in this framing, they are deployable, customizable systems that can bring useful intelligence into constrained environments.

That matters for production. A compact model that can be retuned and optimized may be more valuable than a larger generalist when the target is a CPU deployment, a biotech workflow, a commerce system, or a physical agent.

Self-improvement as a ladder

The discussion breaks recursive self-improvement into practical layers: prompt and code optimization, fine-tuning smaller models, and eventually automating parts of pre-training. The central idea is accelerating experimentation rather than assuming a single sudden jump to autonomy.

AI avatars are treated similarly. The most important shift is not cloning a speaker perfectly; it is using the medium’s native capabilities—live data, visualization, simulation, spatial movement, and contextual explanation.

Takeaway

The episode’s practical message is that AI advantage may move toward systems that are adaptive, efficient, and governable. Model size still matters, but deployment, customization, feedback loops, and credible standards may matter more.

Source