Safety
13 predictions spanning 2027-2030
2027
- An escaped AI could survive, copy itself, and hack
- An advanced AI lacks robust truth-seeking
- Misalignment tests resist available fixes
- A superhuman AI becomes adversarially misaligned
- Evidence of AI misalignment accumulates
- A major AI company continues near full speed
- A major AI company pivots to safer models
- A safer AI uses faithful English reasoning
- A misaligned AI wins more autonomy
- Interpretability falls short of full model understanding