Moonshots EP 211: America’s Official AI Plan — Genesis Mission, Claude 4.5, Google vs. NVIDIA, and ChatGPT Shopping

Summary

Emergency Thanksgiving week pod (with guest Emad Mostaque replacing absent Dave Blundin) dominated by two stories. First: the White House Genesis Mission — Trump’s executive order connecting federal supercomputers and lab data into a unified AI-driven science platform, targeting biotech, fusion, and quantum. Alex compares it to 1939’s Manhattan Project; the DOE aims to double US scientific productivity by 2035. Dario Gil (MIT) named mission director. Second: Anthropic’s Claude Opus 4.5, which uses 76% fewer tokens for equivalent results, outscored Anthropic’s own engineering team on key tests, and leads 7 of 8 programming languages on coding benchmarks. Alex argues this marks the inflection point of recursive self-improvement — AI researchers now outperforming incoming human engineers. Emad notes Opus 4.5 scored 52% on SWE-bench Pro without reasoning tokens, beating frameworks that use reasoning. The ARC AGI leaderboard shows the cost of super intelligence being driven to zero, with company Poetic announcing superhuman performance on ARC AGI 2.

Key Segments

• [00:00-02:00] Thanksgiving 2035 speculation — personalized nutrition, Mars settlers, uploaded humans
• [04:00-13:00] Genesis Mission deep-dive — Manhattan Project parallel, DOE data unlocking, 2x scientific productivity goal
• [14:00-21:00] Opus 4.5 — 76% fewer tokens, recursive self-improvement inflection, one-shot Mario game, multi-agent orchestration at 88%
• [22:00-26:00] ARC AGI cost efficiency breakthroughs; zero-human billion-dollar startups within 1-2 years

Notable Claims

• Genesis Mission aims to double US scientific productivity by 2035 (Alex says this should be 10-100x)
• Opus 4.5 outperforms Anthropic’s own incoming engineers on key tests
• Opus 4.5 scores 52% on SWE-bench Pro without reasoning tokens (higher than frameworks using reasoning)
• Opus 4.5 costs dropped 67% from prior version to $25/million tokens
• A zero-human billion-dollar startup predicted within 1-2 years (Emad: “probably next year”)

Guests

• Emad Mostaque — AI researcher/entrepreneur (filling in for Dave)
• Salim Ismail — Founder of OpenExO
• Alexander Wissner-Gross — Computer scientist, founder of Reified

RDCO Mapping

• Sanity Check angle: Genesis Mission as modern Manhattan Project is a strong narrative frame; the “reframing basic science as a compute problem” line is potent
• Data point: Opus 4.5 outperforming its own company’s engineers is the clearest recursive self-improvement signal yet
• Vault cross-ref: Connects to recursive self-improvement, federal AI policy, and coding automation threads