Moonshots public presentation · Standard

# Moonshots EP271: Mira Murati's 975B Open Model, Ramin Hasani on Post-Transformer AI, and Demis' AI FINRA

By Brandon St-Martin

> For education, information, and entertainment only. Personal observations and opinions. Not financial advice. Verify independently and consult your own adviser. No outcome is promised. Full disclosures.
> The author holds long-term positions in Tesla, SpaceX and Bitcoin.

Moonshots public presentation
Public-source article based on Peter H. Diamandis / Moonshots.
Byline: BSM LLC Research Team Public YouTube citation: [https://www.youtube.com/watch?v=bAoXVyibE6Q](https://www.youtube.com/watch?v=bAoXVyibE6Q)
Disclosure: This material is for education and entertainment purposes only. It is not financial advice. Do your own research, draw your own conclusions, and consult a qualified financial professional before making financial decisions.
Word download: [Moonshots EP271 — 20 Takeaways (.docx)](/downloads/moonshots-ep271-20-takeaways.docx)

## Executive thesis
EP271 is about assurance under fast model change. Inkling is a major U.S. open-weight signal, Liquid is a serious edge-AI watch, and AI-FINRA is best read as a governance proposal rather than current law. The clean lesson is to separate openness, practicality, license terms, privacy, and authority.

## Exactly 20 human-reader takeaways

### 1. AI-FINRA is a proposal, not current law
Episode context: The episode says Demis Hassabis proposed a U.S.-led standards body modeled on FINRA.
Evidence/qualification: Hassabis’ essay and public reporting support the proposal; FINRA and SEC pages support the analogy, not an existing AI regulator.
Why it matters: Readers should not mistake a proposal for binding public law.
What would change the view: The view changes if a government or regulator establishes a formal body with clear scope.

### 2. Capture risk is a design concern
Episode context: A panelist worries that a FINRA-like structure could favor incumbent labs over open-weight, university, and startup work.
Evidence/qualification: The transcript supports the objection; it remains a plausible risk rather than proof of intent.
Why it matters: Good governance must minimize incumbent advantage and include transparent participation.
What would change the view: The view changes if future rules are low-burden, neutral, and genuinely inclusive.

### 3. Assurance should be task-specific
Episode context: Ramin argues that automotive, finance, biotech, e-commerce, and other verticals have different risk profiles.
Evidence/qualification: The transcript supports the design argument; it is not an empirical proof of one framework.
Why it matters: Model permission should vary by use case, data sensitivity, and possible harm.
What would change the view: The view changes if flat provider rules reliably outperform domain-specific checks.

### 4. Inkling’s 975B and 41B-active headline is real
Episode context: The episode reports Thinking Machines’ Inkling as a 975B-total, 41B-active MoE model.
Evidence/qualification: Thinking Machines’ announcement and model card support the core specification.
Why it matters: Inkling belongs on the open-weight frontier watchlist.
What would change the view: The view changes if official model card revisions materially alter the specs.

### 5. Inkling’s thesis is customization
Episode context: The episode says Murati is not claiming best-overall rank; the pitch is adaptable open weights.
Evidence/qualification: Thinking Machines’ materials position Inkling as a base for customization rather than a universal winner.
Why it matters: Task-specific adaptation may matter even when generic leaderboards favor another model.
What would change the view: The view changes if generic frontier models keep winning after cost and customization are considered.

### 6. Open weights are not practical self-hosting for everyone
Episode context: The episode says Inkling can be downloaded and run on premises.
Evidence/qualification: The source paper notes very high aggregate VRAM requirements, far beyond ordinary readers’ hardware.
Why it matters: Open availability and practical operation are separate questions.
What would change the view: The view changes if smaller official checkpoints or accessible hosted terms appear.

### 7. On-prem privacy claims need nuance
Episode context: The episode argues that local fine-tuning can protect proprietary data better than closed APIs.
Evidence/qualification: The general data-locality principle is sound, but blanket allegations about providers are not supported.
Why it matters: Privacy depends on contracts, retention terms, logging, access, and task design.
What would change the view: The view changes with workflow-specific no-training, no-retention, and audit terms.

### 8. Liquid’s scientific lineage is credible
Episode context: Ramin traces Liquid’s work to C. elegans, liquid neural networks, and continuous-time ideas.
Evidence/qualification: Public Liquid materials support Hasani’s research background and the liquid-neural-network lineage.
Why it matters: The science history is interesting without proving every product claim.
What would change the view: The view changes if current products fail to show useful efficiency or reliability gains.

### 9. Post-Transformer needs careful wording
Episode context: The episode discusses exploration beyond transformer architectures.
Evidence/qualification: The source paper warns that current Liquid language models can include attention components, so “transformer-free” is too strong.
Why it matters: Precise architecture language prevents hype from becoming technical confusion.
What would change the view: The view changes when future releases document a cleaner attention-free design.

### 10. Edge AI is about fit and usefulness
Episode context: Ramin describes constrained automotive chips and a sub-gigabyte multimodal foundation model.
Evidence/qualification: Public Liquid materials support a small-model edge direction, but product performance still needs reproducible checks.
Why it matters: Small models can matter when latency, privacy, or offline use is critical.
What would change the view: The view changes if small models cannot deliver useful quality under real constraints.

### 11. Automotive claims need company confirmation
Episode context: The episode mentions Mercedes as an early automotive entry point.
Evidence/qualification: The podcast is a source for what was said; specific rollout scope needs official company notice.
Why it matters: Customer and product details should come from named company sources.
What would change the view: The view changes when Mercedes or Liquid publishes scope, geography, dates, and product details.

### 12. Update loops are measurable
Episode context: Ramin describes data flywheels, drift handling, adapter updates, and personalization.
Evidence/qualification: The source paper treats these as process claims that need measurement.
Why it matters: Long-lived edge AI must handle drift, privacy, rollback, and maintenance.
What would change the view: The view changes if tests show fragile updates, privacy gaps, or high maintenance cost.

### 13. License terms are separate from model quality
Episode context: The source materials distinguish technical fit from legal permission.
Evidence/qualification: Open weights can carry different licenses and conditions; public terms must be reviewed before commercial use.
Why it matters: A model can be useful yet unsuitable for a given organization’s rules.
What would change the view: The view changes when license review clearly permits or limits the intended use.

### 14. Real-time audits and open evals are the useful seed
Episode context: The episode says AI moves too fast for traditional bureaucracy alone.
Evidence/qualification: The transcript supports open evaluation and audit ideas as design principles.
Why it matters: Public methods should test outputs, provenance, risk, and failures instead of relying on brand trust.
What would change the view: The view changes if evaluation methods become checklist theater without catching real problems.

### 15. Independent review matters for credibility
Episode context: The FINRA analogy turns on expert participation and credible review.
Evidence/qualification: The episode itself notes that AI talent constraints differ from securities oversight.
Why it matters: Higher-risk systems need a separation between the proposer and the checker.
What would change the view: The view changes if self-review proves equally reliable across hidden problem sets.

### 16. Liability should focus on harmful actions
Episode context: The panel contrasts regulating inputs or capabilities with responsibility for harmful outcomes.
Evidence/qualification: The source paper treats action-side remediation as the more concrete public lesson.
Why it matters: The practical question is what a system does, who is accountable, and how harm is corrected.
What would change the view: The view changes if capability-only rules prove clearly more effective.

### 17. Capability thresholds need permission thresholds
Episode context: The episode notes choke points, non-state actors, and unresolved enforcement paths.
Evidence/qualification: Public standards and timelines remain uncertain; capability alone does not determine risk.
Why it matters: A weaker system with tool access can be riskier than a stronger read-only system.
What would change the view: The view changes if formal public rules clearly settle thresholds and enforcement.

### 18. Mercedes rollout details remain unconfirmed
Episode context: Ramin gives specific details about model-year scope, updates, and size.
Evidence/qualification: The source paper says those specifics were not confirmed by an official Mercedes source in the checked materials.
Why it matters: Specific customer claims should not be generalized from a podcast statement alone.
What would change the view: The view changes when official notices verify product, date, scope, and feature set.

### 19. Inkling’s next trigger is a smaller or clearer path
Episode context: The current full-size model is important but demanding.
Evidence/qualification: The source paper highlights smaller checkpoints or hosted terms as the practical items to watch.
Why it matters: Readers should distinguish strategic signal from near-term usability.
What would change the view: The view changes with official smaller weights, accessible hosted terms, and strong independent results.

### 20. Fast model change requires fresh checks
Episode context: The episode spans model specs, architecture, licenses, privacy, and governance.
Evidence/qualification: The public takeaway is that assumptions decay quickly as releases and policies change.
Why it matters: Any meaningful change should restart comparison, privacy, and safety checks.
What would change the view: The view changes only if repeated public tests show stable behavior across updates.

## Public sources and limitations
The episode is the main public source. Additional public materials below are included for context and for future checks.
- [https://www.youtube.com/watch?v=bAoXVyibE6Q](https://www.youtube.com/watch?v=bAoXVyibE6Q)
- [https://demishassabis.substack.com/p/a-framework-for-frontier-ai-and-the-dawning-of-a-new-age](https://demishassabis.substack.com/p/a-framework-for-frontier-ai-and-the-dawning-of-a-new-age)
- [https://www.finra.org/about](https://www.finra.org/about)
- [https://www.sec.gov/rules-regulations/self-regulatory-organization-rulemaking/finra](https://www.sec.gov/rules-regulations/self-regulatory-organization-rulemaking/finra)
- [https://thinkingmachines.ai/news/introducing-inkling/](https://thinkingmachines.ai/news/introducing-inkling/)
- [https://thinkingmachines.ai/model-card/inkling/](https://thinkingmachines.ai/model-card/inkling/)
- [https://www.liquid.ai/team/ramin-hasani](https://www.liquid.ai/team/ramin-hasani)
- [https://www.liquid.ai/research/liquid-neural-networks-research](https://www.liquid.ai/research/liquid-neural-networks-research)

## Limitations
- The FINRA comparison is an analogy unless a formal public institution exists.
- Open weights do not by themselves settle cost, privacy, license, or useful performance.
- Liquid architecture claims need careful wording because current systems can be hybrid.
- Automotive rollout details need company confirmation before being treated as fact.

## Neutral monitoring ideas
- AI standards-body proposals and public rule text
- Inkling model card and smaller or hosted options
- Liquid architecture papers and license terms
- Edge-model memory use and latency
- Automotive customer notices and product scope

Also available: [Markdown version](/blog/moonshots-ep271-ai-systems-finra-decentralized-science.md) and [fact sheet (JSON)](/blog/moonshots-ep271-ai-systems-finra-decentralized-science.facts.json).