Overview
Leopold Aschenbrenner (June 2024, situational-awareness.ai) — Former OpenAI researcher argues that AGI by 2027 is “strikingly plausible” based on compute and algorithmic efficiency trendlines. Frames AGI as both a national security crisis and an intelligence explosion catalyst.
Core Argument
GPT-2 to GPT-4 took ~preschooler to ~smart high-schooler abilities in 4 years. Tracing trendlines:
- Compute: ~0.5 OOMs/year
- Algorithmic efficiency: ~0.5 OOMs/year
- “Unhobbling” gains: chatbot → agent capabilities
Combined, these predict another qualitative jump of equivalent magnitude by 2027.
Structure
I. From GPT-4 to AGI: Counting the OOMs
The math on compute scaling, algorithmic improvements, and capability unlocking points to AGI-level systems by 2025-2027.
II. From AGI to Superintelligence: Intelligence Explosion
Hundreds of millions of AGIs could automate AI research, compressing a decade of algorithmic progress (5+ OOMs) into ~1 year.
III. The Challenges
- IIIa. Racing to the Trillion-Dollar Cluster — Industrial mobilization including growing US electricity production by tens of percent
- IIIb. Lock Down the Labs — Current AI labs “basically handing AGI secrets to the CCP on a silver platter”
- IIIc. Superalignment — Controlling AI smarter than humans is unsolved; rapid intelligence explosion makes this extremely tense
- IIId. The Free World Must Prevail — Superintelligence as decisive economic/military advantage; China still competitive
IV. The Project
By 2027/28, the national security state gets involved. Some form of government AGI project emerges. “No startup can handle superintelligence.”
Key Quote
“Everyone is now talking about AI, but few have the faintest glimmer of what is about to hit them.”
Assessment
This is the strongest version of the AGI acceleration thesis. Whether correct or not, the compute buildout trajectory it describes is verifiably happening (trillion-dollar cluster plans, power contract scrambles). The geopolitical framing is the most novel contribution.
Related
- ai-guide-thinking-humans — Mitchell’s more skeptical view of AI capability claims
- alignment-problem — Christian’s treatment of the alignment challenges Aschenbrenner flags
- the-coming-wave — Suleyman’s complementary argument on AI containment