Leopold Aschenbrenner — 2027 AGI, China/US super-intelligence race, & the return of history

Leopold Aschenbrenner argues scaling plus unhobbling produces AGI by 2027 and that OpenAI fired him for raising CCP espionage concerns to the board.

• Aschenbrenner was fired from OpenAI primarily because he shared an internal security memo with the board after a major security incident; HR told him worrying about CCP espionage was racist.
• He predicts AGI (drop-in remote worker level) by 2027–2028, driven by 10GW training clusters costing hundreds of billions.
• Compute scaling trend: GPT-4 cluster ~$500M / 10MW; 2026 cluster ~$10B / 1GW; 2028 cluster ~$100B+ / 10GW; 2030 trillion-dollar cluster at 100GW, over 20% of US electricity.
• The data wall is underrated as a risk: Llama 3 already near the limit of all available internet data; repeating data yields diminishing returns past ~16x repetition.
• Test-time compute overhang is the key near-term unlock: GPT-4 doing millions of tokens of coherent reasoning (vs. hundreds today) could substitute for orders-of-magnitude model size gains.
• Building AGI clusters outside the US (e.g. UAE) is equivalent to doing the Manhattan Project in a foreign country — Aschenbrenner argues it is a serious national security error.
• Post-AGI intelligence explosion math: half an order of magnitude compute + half an order of magnitude algorithmic progress per year compounds to ~10x effective compute annually, dwarfing normal economic growth.
• OpenAI’s stated leak was Aschenbrenner sharing a preparedness brainstorming doc with three external researchers; the sensitive line cited was predicting AGI by 2027–2028, something Sam Altman says publicly.

2024-06-04 · Watch on YouTube