LW - Thoughts on the impact of RLHF research by paulfchristiano

<a href="https://www.lesswrong.com/posts/vwu4kegAEZTBtpT6p/thoughts-on-the-impact-of-rlhf-research">Link to original article</a><br/><br/>Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Thoughts on the impact of RLHF research, published by paulfchristiano on January 25, 2023 on LessWrong. In this post I’m going to describe my basic justification for working on RLHF in 2017-2020, which I still stand behind. I’ll discuss various arguments that RLHF research had an overall negative impact and explain why I don’t find them persuasive. I'll also clarify that I don't think research on RLHF is automatically net positive; alignment research should address real alignment problems, and we should reject a vague association between "RLHF progress" and "alignment progress." Background on my involvement in RLHF work Here are some background views about alignment I held in 2015 and still hold today. I expect disagreements about RLHF will come down to disagreements about this background: The simplest plausible strategies for alignment involve humans (maybe with the assistance of AI systems) evaluating a model’s actions based on how much we expect to like their consequences, and then training the models to produce highly-evaluated actions. (This is in contrast with, for example, trying to formally specify the human utility function, or notions of corrigibility / low-impact / etc, in some way.) Simple versions of this approach are expected to run into difficulties, and potentially to be totally unworkable, because: Evaluating consequences is hard. A treacherous turn can cause trouble too quickly to detect or correct even if you are able to do so, and it’s challenging to evaluate treacherous turn probability at training time. It’s very unclear if those issues are fatal before or after AI systems are powerful enough to completely transform human society (and in particular the state of AI alignment). Even if they are fatal, many of the approaches to resolving them still have the same basic structure of learning from expensive evaluations of actions. In order to overcome the fundamental difficulties with RLHF, I have long been interested in techniques like iterated amplification and adversarial training. However, prior to 2017 most researchers I talked to in ML (and many researchers in alignment) thought that the basic strategy of training AI with expensive human evaluations was impractical for more boring reasons and so weren't interested in these difficulties. On top of that, we obviously weren’t able to actually implement anything more fancy than RLHF since all of these methods involve learning from expensive feedback. I worked on RLHF work to try to facilitate and motivate work on fixes. The history of my involvement: My first post on this topic was in 2015. When I started full-time at OpenAI in 2017 it seemed to me like it would be an impactful project; I considered doing a version with synthetic human feedback (showing that we could learn from a practical amount of algorithmically-defined feedback) but my manager Dario Amodei convinced me it would be more compelling to immediately go for human feedback. The initial project was surprisingly successful and published here. I then intended to implement a version with language models aiming to be complete in the first half of 2018 (aiming to build an initial amplification prototype with LMs around end of 2018; both of these timelines were about 2.5x too optimistic). This seemed like the most important domain to study RLHF and alignment more broadly. In mid-2017 Alec Radford helped me do a prototype with LSTM language models (prior to the release of transformers); the prototype didn’t look promising enough to scale up. In mid-2017 Geoffrey Irving joined OpenAI and was excited about starting with RLHF and then going beyond it using debate; he also thought language models were the most important domain to study and had more conviction about that. In 2018 he started a larger team working on fine-tuning on language models, w...

First published

01/25/2023

Genres:

education

Listen to this episode

0:00 / 0:00

Summary

Link to original articleWelcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Thoughts on the impact of RLHF research, published by paulfchristiano on January 25, 2023 on LessWrong. In this post I’m going to describe my basic justification for working on RLHF in 2017-2020, which I still stand behind. I’ll discuss various arguments that RLHF research had an overall negative impact and explain why I don’t find them persuasive. I'll also clarify that I don't think research on RLHF is automatically net positive; alignment research should address real alignment problems, and we should reject a vague association between "RLHF progress" and "alignment progress." Background on my involvement in RLHF work Here are some background views about alignment I held in 2015 and still hold today. I expect disagreements about RLHF will come down to disagreements about this background: The simplest plausible strategies for alignment involve humans (maybe with the assistance of AI systems) evaluating a model’s actions based on how much we expect to like their consequences, and then training the models to produce highly-evaluated actions. (This is in contrast with, for example, trying to formally specify the human utility function, or notions of corrigibility / low-impact / etc, in some way.) Simple versions of this approach are expected to run into difficulties, and potentially to be totally unworkable, because: Evaluating consequences is hard. A treacherous turn can cause trouble too quickly to detect or correct even if you are able to do so, and it’s challenging to evaluate treacherous turn probability at training time. It’s very unclear if those issues are fatal before or after AI systems are powerful enough to completely transform human society (and in particular the state of AI alignment). Even if they are fatal, many of the approaches to resolving them still have the same basic structure of learning from expensive evaluations of actions. In order to overcome the fundamental difficulties with RLHF, I have long been interested in techniques like iterated amplification and adversarial training. However, prior to 2017 most researchers I talked to in ML (and many researchers in alignment) thought that the basic strategy of training AI with expensive human evaluations was impractical for more boring reasons and so weren't interested in these difficulties. On top of that, we obviously weren’t able to actually implement anything more fancy than RLHF since all of these methods involve learning from expensive feedback. I worked on RLHF work to try to facilitate and motivate work on fixes. The history of my involvement: My first post on this topic was in 2015. When I started full-time at OpenAI in 2017 it seemed to me like it would be an impactful project; I considered doing a version with synthetic human feedback (showing that we could learn from a practical amount of algorithmically-defined feedback) but my manager Dario Amodei convinced me it would be more compelling to immediately go for human feedback. The initial project was surprisingly successful and published here. I then intended to implement a version with language models aiming to be complete in the first half of 2018 (aiming to build an initial amplification prototype with LMs around end of 2018; both of these timelines were about 2.5x too optimistic). This seemed like the most important domain to study RLHF and alignment more broadly. In mid-2017 Alec Radford helped me do a prototype with LSTM language models (prior to the release of transformers); the prototype didn’t look promising enough to scale up. In mid-2017 Geoffrey Irving joined OpenAI and was excited about starting with RLHF and then going beyond it using debate; he also thought language models were the most important domain to study and had more conviction about that. In 2018 he started a larger team working on fine-tuning on language models, w...

Duration

14 minutes

Parent Podcast

The Nonlinear Library: LessWrong Weekly

View Podcast

Share this episode

Similar Episodes

Announcing AlignmentForum.org Beta by Raymond Arnold.

Release Date: 12/03/2021

Description: Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Announcing AlignmentForum.org Beta, published by Raymond Arnold on the AI Alignment Forum. We've just launched the beta for AlignmentForum.org. Much of the value of LessWrong has come from the development of technical research on AI Alignment. In particular, having those discussions be in an accessible place has allowed newcomers to get up to speed and involved. But the alignment research community has at least some needs that are best met with a semi-private forum. For the past few years, agentfoundations.org has served as a space for highly technical discussion of AI safety. But some aspects of the site design have made it a bit difficult to maintain, and harder to onboard new researchers. Meanwhile, as the AI landscape has shifted, it seemed valuable to expand the scope of the site. Agent Foundations is one particular paradigm with respect to AGI alignment, and it seemed important for researchers in other paradigms to be in communication with each other. So for several months, the LessWrong and AgentFoundations teams have been discussing the possibility of using the LW codebase as the basis for a new alignment forum. Over the past couple weeks we've gotten ready for a closed beta test, both to iron out bugs and (more importantly) get feedback from researchers on whether the overall approach makes sense. The current features of the Alignment Forum (subject to change) are: A small number of admins can invite new members, granting them posting and commenting permissions. This will be the case during the beta - the exact mechanism of curation after launch is still under discussion. When a researcher posts on AlignmentForum, the post is shared with LessWrong. On LessWrong, anyone can comment. On AlignmentForum, only AF members can comment. (AF comments are also crossposted to LW). The intent is for AF members to have a focused, technical discussion, while still allowing newcomers to LessWrong to see and discuss what's going on. AlignmentForum posts and comments on LW will be marked as such. AF members will have a separate karma total for AlignmentForum (so AF karma will more closely represent what technical researchers think about a given topic). On AlignmentForum, only AF Karma is visible. (note: not currently implemented but will be by end of day) On LessWrong, AF Karma will be displayed (smaller) alongside regular karma. If a commenter on LessWrong is making particularly good contributions to an AF discussion, an AF Admin can tag the comment as an AF comment, which will be visible on the AlignmentForum. The LessWrong user will then have voting privileges (but not necessarily posting privileges), allowing them to start to accrue AF karma, and to vote on AF comments and threads. We’ve currently copied over some LessWrong posts that seemed like a good fit, and invited a few people to write posts today. (These don’t necessarily represent the longterm vision of the site, but seemed like a good way to begin the beta test) This is a fairly major experiment, and we’re interested in feedback both from AI alignment researchers (who we’ll be reaching out to more individually in the next two weeks) and LessWrong users, about the overall approach and the integration with LessWrong. Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

Explicit: No

Details

AMA on EA Forum: Ajeya Cotra, researcher at Open Phil by Ajeya Cotra

Release Date: 11/17/2021

Description: Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: AMA on EA Forum: Ajeya Cotra, researcher at Open Phil, published by Ajeya Cotra on the AI Alignment Forum. This is a linkpost for Hi all, I'm Ajeya, and I'll be doing an AMA on the EA Forum (this is a linkpost for my announcement there). I would love to get questions from LessWrong and Alignment Forum users as well -- please head on over if you have any questions for me! I’ll plan to start answering questions Monday Feb 1 at 10 AM Pacific. I will be blocking off much of Monday and Tuesday for question-answering, and may continue to answer a few more questions through the week if there are ones left, though I might not get to everything. About me: I’m a Senior Research Analyst at Open Philanthropy, where I focus on cause prioritization and AI. 80,000 Hours released a podcast episode with me last week discussing some of my work, and last September I put out a draft report on AI timelines which is discussed in the podcast. Currently, I’m trying to think about AI threat models and how much x-risk reduction we could expect the “last long-termist dollar” to buy. I joined Open Phil in the summer of 2016, and before that I was a student at UC Berkeley, where I studied computer science, co-ran the Effective Altruists of Berkeley student group, and taught a student-run course on EA. I’m most excited about answering questions related to AI timelines, AI risk more broadly, and cause prioritization, but feel free to ask me anything! Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

Explicit: No

Details

AMA: Paul Christiano, alignment researcher by Paul Christiano

Release Date: 12/06/2021

Description: Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: AMA: Paul Christiano, alignment researcher, published by Paul Christiano on the AI Alignment Forum. I'll be running an Ask Me Anything on this post from Friday (April 30) to Saturday (May 1). If you want to ask something just post a top-level comment; I'll spend at least a day answering questions. You can find some background about me here. Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

Explicit: No

Details

AI alignment landscape by Paul Christiano

Release Date: 11/19/2021

Description: Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: AI alignment landscape, published byPaul Christiano on the AI Alignment Forum. Here (link) is a talk I gave at EA Global 2019, where I describe how intent alignment fits into the broader landscape of “making AI go well,” and how my work fits into intent alignment. This is particularly helpful if you want to understand what I’m doing, but may also be useful more broadly. I often find myself wishing people were clearer about some of these distinctions. Here is the main overview slide from the talk: The highlighted boxes are where I spend most of my time. Here are the full slides from the talk. Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

Explicit: No

Details