Open Philanthropy has just launched a large new Request for Proposals for technical AI safety research. Here we're sharing a reference guide, created as part of that RFP, which describes what projects we'd like to see across 21 research directions in technical AI safety.

This guide provides an opinionated overview of recent work and open problems across areas like adversarial testing, model transparency, and theoretical approaches to AI alignment. We link to hundreds of papers and blog posts and approximately a hundred different example projects. We think this resource acts as a good starting point for technical people getting started in alignment research. We'd also welcome feedback from the LW community on our prioritization within or across research areas.

For each research area, we include:

• Discussion of key technical problems and why they matter
• Related work and important papers in the field
• Example projects we'd be excited [...]

---

Outline:

(01:34) Synopsis of Research Areas

(02:29) Adversarial machine learning

(04:57) Exploring sophisticated misbehavior in LLMs

(07:26) Model transparency

(10:48) Trust from first principles

(12:09) Alternative approaches to mitigating AI risks

(13:06) The Research Areas

(13:10) \*Jailbreaks and unintentional misalignment

(20:32) \*Control evaluations

(24:52) \*Backdoors and other alignment stress tests

(29:36) \*Alternatives to adversarial training

(35:33) Robust unlearning

(38:21) \*Experiments on alignment faking

(44:00) \*Encoded reasoning in CoT and inter-model communication

(49:06) Black-box LLM psychology

(54:40) Evaluating whether models can hide dangerous behaviors

(59:16) Reward hacking of human oversight

(01:02:56) \*Applications of white-box techniques

(01:08:12) Activation monitoring

(01:14:23) Finding feature representations

(01:21:27) Toy models for interpretability

(01:27:11) Externalizing reasoning

(01:31:39) Interpretability benchmarks

(01:34:21) †More transparent architectures

(01:40:05) White-box estimation of rare misbehavior

(01:43:15) Theoretical study of inductive biases

(01:49:45) †Conceptual clarity about risks from powerful AI

(01:55:02) †New moonshots for aligning superintelligence

---

First published:

Source:

Narrated by TYPE III AUDIO.