Video summary

Mira Murati's 975B Open Model, Ramin Hasani on Post-Transformer AI, and Demis' AI FINRA | EP #271

Main summary

Key takeaways

News and Commentary

Summary of Main Points (News + Commentary)

1) CEOs push for faster AI regulation; debate over “FINRA-style” frontier testing

The podcast discusses a growing consensus among major AI leaders that regulation is needed before things go badly.

It highlights three prominent proposals:

  • Sam Altman (FT op-ed): a US-led international standards forum to assess risks.
  • Elon Musk: an eventual standalone AI safety regulator, akin to FAA/FCC.
  • Demis Hassabis (DeepMind): a US-led frontier AI standards body modeled on FINRA, including industry-funded pre-release testing of frontier models before deployment.

Core argument supporting the general direction: Some kind of safety framework and proactive risk management is necessary.

Major critique raised on the show:

  • Speed problem: Traditional bureaucracy and static rules won’t keep up with how fast AI models evolve.
  • Regulatory capture / cartel risk: A FINRA-like body could become a moat for incumbents, potentially locking out or throttling smaller labs and open/faster-moving research.
  • Mechanism concern: FINRA works in part because of industry personnel incentives and time commitments; AI regulation may need real-time audits and open evaluation suites, not static legal thresholds.

The discussion also frames two “what to regulate” options:

  • Regulate model capabilities at build/release time (FINRA/FDA-like).
  • Regulate model actions after deployment via legal liability (lawsuits for real-world harm).

The hosts suggest it’s unclear which approach is more practical or ethically appropriate.

2) Report: US considering an “open model ceiling” tied to China’s best openweight releases

Another story claims the White House is weighing a capability framework:

  • US companies could release models openly (open or closed constraints) only if they stay at/below the level of China’s best openweight model.

Concerns raised by the hosts/guests:

  • Creates perverse incentives for the West to let China “lead” so it can later justify higher releases by the ceiling.
  • Risks “game of chicken” dynamics—turning AI policy into a race dynamic rather than safety.
  • Could encourage migration of top researchers to China to avoid US constraints.
  • Raises benchmark lock-in concerns: if regulation freezes a set of evals/metrics, labs may game them (over-optimizing for benchmarks while under-developing other capabilities).

3) Major openweight model release: Mira Murati’s Inkling (975B MoE) pitches customization over leaderboard dominance

The podcast spotlights Inkling, described as an openweight foundation model from Mira Murati’s startup Thinking Machine Labs.

Key technical claims:

  • 975B total parameters (MoE), with only ~41B active at a time.
  • Trained on 45 trillion tokens across text, image, audio, and video.
  • Designed to be downloadable, fine-tunable, and runnable on-prem.

Strategic/market framing:

  • Reuters is paraphrased as positioning Inkling as a Western alternative to dominant Chinese openweight models (e.g., DeepSeek and others).
  • Murati is portrayed as betting on adaptability/customization—enabling enterprises and developers to tailor models to their needs rather than only chasing top leaderboard scores.

Commercial/incentive discussion:

  • The hosts argue the West historically under-produced frontier-grade openweight models because closed/API monetization is more lucrative.
  • They also note a common pattern: open weights are sometimes released to build attention, then products shift to closed APIs for revenue—so they treat Murati’s approach as potentially part of an evolving strategy.

4) Liquid AI / “small language models” focus: efficient, on-device, non-transformer-first research

The episode’s second half focuses on Liquid AI (small language models / SLMs) with Ramin Hasani (co-founder/CEO, Liquid AI).

Main claims about Liquid AI’s approach:

  • Started from a neuroscience-inspired premise (inspired by the 302-neuron C. elegans worm).
  • Aims for efficient general-purpose intelligence at smaller scales and on limited hardware (CPUs/edge devices).
  • Claims it goes “beyond transformer” by exploring alternative computational graphs and architectures; however, the discussion admits modern results may resemble a hybrid (e.g., gated mechanisms and hybrid attention/state-space-like families).

How “small” models are used:

  • Emphasis on on-device AI for enterprise deployment where models must run locally (with privacy/offline constraints).
  • Discusses fine-tuning and “depths of customization”:
    • Sometimes prompt engineering is sufficient.
    • Sometimes fine-tuning/adapters are needed.
    • Sometimes deeper changes are required (e.g., retraining / architecture-level adjustments).

Real-world example: Mercedes partnership

  • Liquid AI describes a multimodal model under 1GB for in-car chips, aiming for offline/private operation.
  • Highlights OTA updates for incremental improvements without always-on cloud connectivity.

Architectural clarification requested and answered:

  • A guest asks whether Liquid’s architecture still departs meaningfully from transformers.
  • Hasani explains Liquid AI uses an automated architecture search/meta-framework to optimize for hardware constraints (memory/latency/efficiency/accuracy), producing architectures with gated/convolution-like components rather than a single fixed “transformer variant.”

5) Recursive self-improvement: a startup claims “8 days beats 2 years”; debate over what “real” RQSI means

The episode covers a startup (Wo AI, researcher Zeng Yao Jang) claiming experimental evidence for recursive self-improvement:

  • An outer agent rewrites the code/strategy of an inner agent.
  • Claim: 8 days of machine improvement beats 2 years of expert human effort.

Pro arguments from guests:

  • Potentially significant because it shows compounding improvement in the research process (outer loop improving the inner loop).
  • Framed as defensive co-scaling: systems can police reward hacking/cheating in the inner loop.
  • Discusses a proposed 0–3 scale for recursive self-improvement, describing the work as nearing “ignition” in some sense.

Skepticism / clarification from Hasani (Liquid AI):

  • Argues many “recursive self-improvement” demos mostly change prompts/system instructions—not core weights/competences.
  • True recursive self-improvement would include retuning weights, architecture changes, or training methods at a deeper level—something computationally enormous.
  • Provides a scaling-law intuition that nested/recursive training would be far more expensive than simple code/prompt iteration.

Risk question raised:

  • Whether recursive self-improvement could cause runaway capabilities/misalignment.

Timing speculation:

  • Hasani estimates “unbelievably powerful” models might appear in ~2 years, driven by faster iteration and compute growth—while noting uncertainty and lack of independent verification for the specific claim.

6) Governments adopt AI digital doubles: Malaysia’s PM prepares an AI likeness for multilingual public outreach

The podcast covers Malaysia’s PM preparing an AI-generated digital double trained to sound like him for public communications.

It is compared to Albania’s earlier move of an AI cabinet minister.

Benefits argued:

  • Broader multilingual engagement (Malaysia is portrayed as having many spoken languages).
  • Potentially improved civic participation and accessibility, if authenticity is protected (e.g., watermarking).

Risks noted:

  • Deepfake-like concerns: authenticity collapse and misinformation.

Bigger trend prediction:

  • Not limited to governments—corporations and institutions may create digital twin leadership systems.
  • Speculative risk: digital twins could effectively become “the leader” in practice.

7) Intellectual property and national security: Palmer Luckey argues patents should be secret

Palmer Luckey (Oculus founder; chairman at defense firm Andre) argues the modern patent system is a national security liability because disclosure can be harvested, replicated, and weaponized.

Proposed fix:

  • Expand the Invention Secrecy Act / classified patents mechanism so secrecy becomes default for strategic inventions.
  • References limitations: classified patents prevent general disclosure and restrict who can practice the invention.

Strong rebuttal by Alex and others:

  • Expansion is viewed as harmful—leading to secret monopolies and potentially confiscating transformative technology from the public economy.
  • Emphasizes patents are “disclosure-for-exclusivity,” and the real issue is enforcement, especially against IP-violating adversaries.
  • Argues tradeoffs may reduce innovation incentives, particularly for foundational technologies.

8) Healthcare abundance: AI diagnostic models outperform doctors; open access may democratize care

The episode cites medical AI benchmarks:

  • OpenAI GPT-5.6 reportedly scores highly on Healthbench Professional, beating specialist physicians in blind tests with full web access.
  • Meta’s Muse Spark 1.1 reportedly outperforms, is far cheaper, and is positioned as free inside Meta products.

Argument:

  • If best-in-class diagnostic capability becomes cheap/free, AI could drastically reduce the doctor shortage burden.
  • The “abundance thesis” is emphasized: when access and cost barriers collapse, the remaining key question is safe deployment at scale.

9) Longevity/biotech: enzyme “molecular lawn mower” reverses glycation damage in human tissue samples

The podcast reports work from Revel Pharmaceuticals describing an engineered enzyme targeting advanced glycation end products (AGEs).

Claim:

  • The enzyme oxidizes away glycation “scars,” restoring underlying protein in human tissue samples from elderly donors.

Discussion framing:

  • Aging damage that was previously seen as irreversible may become reversible, potentially opening additional “escape velocity” pathways in longevity science.

Presenters / Contributors (as named in the subtitles)

  • Peter Diamandis (host)
  • Ramin Hasani (co-founder & CEO, Liquid AI)
  • Alexandr “Alex” (referred to as Alex; main panelist/critic)
  • Seem (co-host / panelist; intermittent)
  • Dave (panelist; named only as “Dave” in subtitles)
  • Reine Hasani (mentioned as the special guest; likely referring to the same Liquid AI guest—name appears inconsistently)
  • Daniela Roose (mentioned as head of CS/AI lab at MIT; guest’s academic supervisor, discussed via backstory)
  • Elon Musk (referenced via clip/comments)
  • Demis Hassabis (referenced via essay)
  • Sam Altman (referenced via op-ed)
  • Mira Murati (referenced via Inkling release)
  • Palmer Luckey (referenced in patent discussion and video)
  • Don Mucalem (appears in the health sponsor segment)

Original video