Alignment and safety
What it means for a system to do what we intend rather than what we wrote down — specification gaming, reward hacking and Goodhart's law (simulated), the outer/inner alignment framing, proposals for supervising systems we can't fully check (debate, weak-to-strong generalisation), dangerous-capability evaluations, honesty, and the real disagreement among experts about how much any of this matters.
50 min
After this lesson you can
- Define specification gaming and reward hacking, with documented examples
- Explain regressional Goodhart and why heavy-tailed proxy error makes optimisation pressure dangerous
- State the outer/inner alignment distinction as Hubinger et al. frame it
- Describe debate and weak-to-strong generalisation, and what each has and hasn't shown
- Summarise dangerous-capability evaluations and the main positions in the expert debate, with attribution
In this lesson
- Specification gaming and reward hacking
- Goodhart’s law, simulated
- Outer and inner alignment
- Supervising what you can’t check: scalable oversight
- Evaluations for dangerous capabilities
- Honesty and hallucination
- How worried should we be? The range of views
- Open problems
This lesson is part of Descent Pro
Descent Pro unlocks every lesson on the route, from base camp to the summit. Every plan starts with a free trial.