Bernie Demands the Labs Stop, Wall Street Turns GPUs Into Bonds, Grok 4.7 Takes #1 with Emad Mostaque | EP #279

Bernie Sanders sent a formal letter to the CEOs of Anthropic, Meta, and OpenAI demanding an immediate pause on AI development, citing each company's own prior safety commitments and arguing that threshold has been crossed. Sanders referenced AI being used to design a new virus, quoted the CIA director describing frontier models as akin to digital nuclear weapons, and cited Yoshua Bengio calling recent developments a wake-up call. Speakers pushed back that pausing is an absurdly coarse instrument because open models cannot be uninvented and the rest of the world will not comply, and argued the defensive side of AI must be co-scaled with offensive capabilities. Emad Mostaque offered a more targeted alternative, saying restricting reagents and input processes for dangerous biological agents is more tractable than a broad AI pause. Stanford researchers synthesized roughly 300 bacteriophage designs and produced 16 viable phages capable of infecting E. coli, and Mostaque noted that Evo 2, a 40-billion parameter open-source biological foundation model, can run on a MacBook. Multiple speakers identified DNA and RNA synthesizer equipment as the practical unregulated control point, and one speaker argued deploying ubiquitous cheap DNA sequencers everywhere is a more valuable defensive strategy than AI-level output restrictions.

Grok 4.6 launched the morning of the episode at two and six dollars per million input and output tokens, matching GPT-5.6 on the Artificial Analysis Intelligence Index at a score of 61. Mostaque provided specific parameter counts, saying Grok 4.5 had 1.5 trillion parameters which was post-trained to produce 4.6, that Grok 4.7 scales to 2 trillion parameters and will be trained on SpaceX physics and engineering knowledge, and that Grok 5 is planned at 6 and then 10 trillion parameters. Mostaque predicted Grok 4.7 will surpass Opus and take the number one position within two weeks. Alex claimed Grok 4.6 post-training leaned heavily on the Cursor acquisition, and that xAI licensed reasoning trace data from Cursor prior to the acquisition being finalized, framing this as a westernized version of what Chinese frontier labs do when distilling reasoning traces from competitors. Alex noted Elon has made public comments about wanting to start a new pre-training session approximately monthly, a cadence no other frontier lab is matching, with Google DeepMind operating on an annual basis.

Nvidia partnered with Apollo, BlackRock, Blackstone, Brookfield, and KKR to mobilize over 500 billion dollars in third-party capital for AI infrastructure, positioning itself as the architect of the financing layer rather than just a silicon supplier. Dave described Jensen Huang's approach as creating standalone financial structures tied to individual compute clusters that can be stamped out infinitely. Dave raised a core disagreement with the structure: securitizing GPU cash flows over ten years creates technological risk because a breakthrough in new architecture could strand compute assets suddenly, and exponential technologies do not provide the predictable depreciation that financial assets require. Dave noted that Larry Fink himself referenced mortgage-backed securities in describing the structure, and predicted that if ratings agencies correctly price technological risk, the build-out could see smooth growth for ten years or more.

Emad Mostaque reported that Higgs Field is generating approximately 700 million dollars in annual revenue run rate roughly one and a half years after founding. The company produced a 110-minute feature film with 28 people in four weeks at a total budget of approximately one million dollars using Sora Dance 2.5, compared to a traditional feature film costing 20 to 100 million dollars and taking one to 18 months. Mostaque estimated a similar film could be made for 100,000 dollars in compute today and potentially 10,000 dollars by year end, partly because the average Hollywood film uses three-second shots, enabling a tenfold reduction in compute cost. Nine of the top 10 text-to-video models on the AI analysis leaderboard are from China, and Alex argued these models are learning physics, motion, and causality relevant to robotics and autonomous driving, with enterprise code and reasoning work currently separate but unlikely to remain so.

Mark Zuckerberg published a 6,500-word essay outlining a vision of personal superintelligence distributed to every person on Earth via billions of personal agents. Meta is open-sourcing Muse Glimmer, a 30-billion parameter dense model running on a laptop, and plans to open the weights for Muse Spark 1.2. One speaker argued Meta's economic incentive in open-sourcing is to commoditize the model layer and shift value toward distribution and the social graph where Meta is dominant across its 3 billion users. Meta acquiring Scale AI was compared to Elon acquiring Cursor as a strategy to obtain reasoning traces and leapfrog back to the frontier.

The longevity XPRIZE carries a total prize pool of 101 million dollars with 80 million as the grand prize, and 10 million dollars has already been awarded across 10 finalist teams from an original field of 800. Finalist teams must conduct human trials with approximately 150 people in control groups, and the competition targets reversing functional age by 20 years in cognition, muscle, and immune system by 2030. Current average US healthy lifespan ends at 63, leaving roughly 19 years of poor health before death at the average age of 79, and the global cost of age-related diseases is estimated at 20 trillion dollars annually.