This graph represents a conceptual scenario called the "Intelligence Explosion," which outlines the potential trajectory of artificial intelligence (AI) progress. It uses "effective compute" (measured logarithmically and normalized to GPT-4) as a proxy for AI capability over time.
Key Elements:
-
Axes:
- Y-axis: The "Effective Compute" scale is logarithmic, ranging from to , normalized to GPT-4's compute level.
- X-axis: The timeline spans from 2018 to 2030, showing the progression of AI development over time.
-
Markers of AI Progress:
- GPT-2: Equivalent to a preschooler's cognitive capability.
- GPT-3: Compared to an elementary schooler's cognitive capability.
- GPT-4: Analogous to a "smart high schooler."
- "Automated Alec Radford?": Hypothetical level of compute where AI might fully automate its own development processes (possibly named as a nod to an AI researcher or key figure).
- Superintelligence: Theoretical level where AI surpasses human intelligence by orders of magnitude.
-
The Intelligence Explosion:
- The graph suggests a rapid acceleration in AI capability due to "Automated AI Research." This concept implies that once AI reaches a certain threshold of intelligence, it could take over its own development, exponentially improving itself.
-
Shaded Region:
- Represents uncertainty or variability in the timeline and speed of this hypothetical intelligence explosion. This shaded area spans the late 2020s, showing the range of possibilities for achieving superintelligence.
Interpretation:
- The graph conveys a hypothetical trajectory where advancements in AI (driven by self-optimization) lead to an "explosion" in capability within a short timeframe.
- It underscores the potential risks and opportunities of reaching a point where AI systems are capable of autonomous research and improvement.
