Chapter 26
26Open Problems (Conclusion)
26.1Hallucination and calibration
Planned: Why models are confidently wrong; what partial fixes exist.
26.2Interpretability
Planned: Features, circuits, and the dream of reading a model's mind.
26.3Alignment at and beyond human level
Planned: Scalable oversight, deception, and evaluation we can trust.
26.4Data, efficiency, and the wall question
Planned: Running out of tokens; synthetic data; whether scaling continues.
26.5Where the field might go
Planned: A sober take on the next few years and what to keep learning.