Brian Armstrong on Bitcoin, Anthropic Drops Fable 5 & Mythos 5, NewLimit's $435M Age-Reversal | 264

Armstrong outlined three phases: (1) connect LLMs to human Coinbase accounts via MCP API; (2) embed agentic interface (Coinbase Advisor) in-app; (3) give every AI agent its own financial account. The third phase is scaling now with ~100M transactions. He envisions on-chain reputation systems—like a FICO score or Yelp rating—to reduce fraud, where transaction graphs assign reputation scores based on counterparty quality. This could enable agents to assess merchant reliability before transacting.

Armstrong explained the pipeline stages: AI recommendation → pooled screens → functional assays → non-human primate trials → human trials. The process has gone faster than expected because AI accelerates the search space exploration. The $435M raise funds multiple concurrent trials—phase 1 at $10-20M each, scaling up. Armstrong noted the line between disease prevention and enhancement is blurry (e.g., preventing osteoporosis = stronger bones = enhancement), suggesting the technology will start with disease and expand.

Armstrong detailed the contentious debate: Option A freezes coins that don't upgrade by a deadline—like 'a ship full of gold sinking to the bottom of the ocean'—preventing a quantum attacker from dumping them and crashing the price. Option B preserves Bitcoin's fundamental guarantee that wealth cannot be seized, treating the vulnerable coins as a bounty. A third hybrid option would freeze coins but allow an appeal process for rightful owners. Coinbase's quantum advisory council includes Dan Boneh (Stanford), Scott Aaronson (UT Austin), Justin Drake (Ethereum Foundation), and Yehuda Lindell (Coinbase). Armstrong noted this isn't just a Bitcoin issue—all internet cryptography faces the same threat.

Alex argued LEV might be achieved this year in a 'spiky' way, citing a recent double-blinded study showing GLP-1s producing double-digit age reversal on epigenetic clocks in HIV patients. Peter countered that biomarkers are unreliable and functional measures matter more. The group compared it to the Turing test—we may pass LEV without society noticing. Armstrong predicted 2033 (Kurzweil's estimate) is plausible but FDA approval is the bottleneck. Dave noted 'the baby that's going to live to a thousand years old is already alive.'

The hosts debated ethics and inevitability. Armstrong noted jurisdiction is the biggest complexity—local bans won't work because people will travel. He argued the disease/enhancement line is impossibly blurry and predicted we'll start with disease and expand to IQ and other traits. Alex compared it to Gattaca but noted base editing goes beyond embryo selection—it's 'better than the best of the parents.' Salim framed it as biology becoming a 'digital programmable domain' where 50 trillion cells per human is now a software engineering problem. The group noted IVF was once considered immoral but is now mainstream—suggesting norms will shift.

Alex offered three perspectives: (1) glass half empty—Apple outsourcing core tech is unprecedented and culturally dissonant; (2) foundation models are like search engines—intrinsically localizable, so Apple will partner with different providers by region (Baidu in China, etc.); (3) hyperdeflationary economics mean any loss of sovereignty is temporary—in 1-2 years, these capabilities will fit on-device. Dave warned that losing the interface (Siri → Google → Android) while simultaneously losing chip manufacturing priority at TSMC to AI data center customers could create an existential threat to iPhone revenue. The hosts were skeptical given Siri's history of disappointment.

Brian Armstrong broke down the agent economy into three phases: (1) connecting LLMs like ChatGPT to human Coinbase accounts via MCP API for context and control; (2) embedding an agentic interface (Coinbase Advisor) directly in the app for portfolio rebalancing and tax-loss harvesting; (3) giving every AI agent its own financial account. The third phase is what's scaling now. Armstrong sees a future where agents orchestrate hundreds of other agents, requiring an 'agent payroll' system. He subscribes to the 'polytheist' view of AI—many specialized agents communicating and transacting—rather than a single AGI. This agentic economy, he argues, will eventually dwarf the human economy.

Alex noted that the model can oneshot complex tasks like building a cyberpunk FPS game with no errors and a soundtrack. More impressively, it can play and win Pokémon games purely from visual reasoning—requiring long-range planning. The hosts observed that Anthropic doubled the price, countering the 'commodity intelligence' narrative. The safety split is stark: Fable 5 will refuse and downgrade to Opus on biology/chemistry topics with broad buffers, while Mythos 5 is less restricted. Alex speculated Anthropic is aggressively using RLVR (reinforcement learning with verifiable rewards) on long-range reasoning challenges involving spatial reasoning, large codebases, and CTF hunts. The hosts predicted the rate of model releases will accelerate as both Anthropic and OpenAI race toward their IPOs.

Armstrong explained the pipeline: AI ingests all existing literature and NewLimit's own wet lab data (the largest dataset in the field) to explore 10 quadrillion protein combinations. The AI recommends experiments; pooled screens identify candidates that make cells look phenotypically younger; functional assays then test if they act younger (e.g., reprogrammed liver cells processing caffeine and acetaminophen). The best candidates move to non-human primate and human trials. Armstrong initially thought this would be a 5-10 year basic research endeavor, but the AI-accelerated process has compressed timelines dramatically. The $435M will fund multiple clinical trials—phase 1 trials costing $10-20M each, scaling up through phases 2 and 3.

Alex noted the form factor is dominated by solar arrays and radiators, with compute being a tiny fraction of surface area—suggesting massive room for improvement as compact fusion or power beaming matures. He called this 'the most disjointed, largest, bulkiest, most mainframe era node in the Dyson swarm we're ever likely to see.' SpaceX also released concept videos of manufacturing these on the Moon and launching them via electromagnetic railgun, which is why vertical integration matters—you can't do lunar manufacturing without controlling the full supply chain. The Gigasat factory will be 1,000 acres with 11 million square feet of facilities, producing everything from solar ingots to finished satellites on one campus. Armstrong noted the design must have redundancy for micrometeorite hits on radiators and solar panels, and speculated on how it folds into Starship's payload bay.

Alex speculated that Elon may be putting Grok on 'de facto life support' to focus on becoming a hyperscaler first, accumulating revenue to eventually build the Dyson swarm and then buy his way back to the frontier with superior compute. He noted that algorithmic wars may burn out as recursive self-improvement discovers optimal model architectures, shifting competition entirely to who has the most hardware. The hosts debated whether Elon will maintain partnerships long-term—Peter predicted he won't, based on historical pattern. The duopoly of Dario (Anthropic) providing the smartest AI and Elon providing space-based data centers and manufacturing could be incredibly powerful if it holds, but the partnership is likely temporary.

Dave argued this is a 'horrible, horrible thing' long-term—while these specific investments may be good during an AI race with China, it sets a precedent for government capital allocation that future administrations will abuse. Alex countered that if AI companies become larger than the rest of the economy combined, some form of quasi-nationalization is inevitable. He suggested a scenario where the Treasury takes 5-10% golden shares in OpenAI and Anthropic, then gradually liquidates into an S&P 500 index to fund a sovereign wealth fund—effectively creating a universal basic dividend. Brian Armstrong was skeptical of government capital management, preferring policy-setting to equity ownership, but acknowledged a sovereign wealth fund could increase social cohesion if every citizen got a share at birth. He worried about politicization of investment decisions.