Situational Awareness
- What parts of this could be turned into predictions?
- "We are racing through the OOMs extremely rapidly, and the numbers indicate we should expect another ~100,000x effective compute scaleup—resulting in another GPT-2-to-GPT-4-sized qualitative jump—over four years"
- What does it mean to have "models as smart as PhDs or experts that can work beside us as coworkers"
- Notes
- Seems to underrate
- Multiple counting
- Separate labs slowing down
- Processese that can't happen in parallel (learning from big runs, improving AI agents)
- Regulation
- Improving humans
- Compute for running models, not agents
- The more quickly new tech appears, the less return on it
- AI being used for a wide range of things, as opposed to just more AI