Exercises: Decomposing Tasks for Better Supervision

Resources: Decomposing Tasks for Better Supervision

Exercises: Interpretability 9

Resources: Interpretability 9

Exercises: Track 1: Eliciting Latent Knowledge

Resources: Track 1: Eliciting Latent Knowledge

Exercises: Literature Review or Project Proposal

Resources: Literature Review or Project Proposal

Exercises: Track 3: Science and Deep Learning

Resources: Track 3: Science and Deep Learning

Exercises: Reasoning about Reasoning

Resources: Reasoning about Reasoning

Exercises: Further Understanding the Alignment Problem

Resources: Further Understanding the Alignment Problem

Exercises: Preventing Misgeneralization

Resources: Preventing Misgeneralization

Exercises: Track 2: Agent Foundations

Resources: Track 2: Agent Foundations

[Korbak, Perez and Buckley](https://arxiv.org/pdf/2205.11275.pdf) claim that “KL-regularised RL is equivalent to variational inference”. Read up to equation 7, then prove the claim in the following paragraph (which they derive in the appendix).




KL-regularised RL is equivalent to variational inference

Work through some of the exercises in the [factored cognition primer](https://primer.ought.org/).


Factored Cognition 

**Problem:** We won’t be able to supervise the long-term behaviour of models which are more intelligent than us. 
**Proposed solution:** Use models to supervise each other.
### Assumed background
Note the following two resources are assumed background for this session.
1. [AI safety via debate (Irving et al., 2018)](https://arxiv.org/abs/1805.00899) **(sections 1 and 2)** (25 mins)
2. [Self-critiquing models for assisting human evaluators: blog post (Saunders et al., 2022)](https://openai.com/blog/critiques/) (10 mins)


Alignment 201

Exercises: Decomposing Tasks for Better Supervision

Exercises