SITUATIONAL AWARENESS: The Decade Ahead
Source: https://situational-awareness.ai/ Author: Leopold Aschenbrenner Date: 2025-04-04 (saved), originally published June 2024
Summary
Leopold Aschenbrenner’s 165-page memo on AI progress, safety, and geopolitics. Former OpenAI safety researcher. Core thesis: AGI is coming this decade, national security implications are being underestimated, and the US needs to treat AI development as a matter of strategic importance comparable to nuclear weapons.
Key Claims
- By 2027: models will be capable of autonomous AI research, potentially running in data centers doing “a year of science in a day.”
- The “intelligence explosion” is plausible: once AI can do AI research, progress compounds rapidly.
- National security framing: AI weights are like nuclear secrets — theft by adversaries could be catastrophic. Current security practices are insufficient.
- Compute cluster scale required for AGI: estimated 10^27-10^29 FLOPs, achievable by ~2027 with current trajectory.
- Safety concern: the “race dynamic” between labs and between nations may force unsafe deployment decisions.
- The “schlep” problem: most of the work to AGI is engineering, not research — and it’s being done by a small number of people.
Entities
- Leopold Aschenbrenner — former OpenAI, wrote this after leaving
- OpenAI — subject of several claims
- Anthropic — mentioned as safety-focused alternative
Concepts
- Scaling & Compute — compute trajectory is central argument
- Autonomous Research — AI doing AI research is the crux of intelligence explosion
- AI Alignment — safety in race conditions