As I dug through my drafts for today’s InkHaven’s post, I found an old draft of a blog post written in late 2022, arguing that LessWrong/Alignment Forum/etc should have a stronger norm for including related work sections.[1] When I first started posting actively on LessWrong and meddling in papers back in 2022, one of the things I pushed for was substantial related work sections. Nowadays, I'm a bit more mixed on the value of related work sections on the margin. So I’m writing a post about why I was a big proponent of more effort on related work sections in 2022, and why I’m not as large of a proponent nowadays.
There were several reasons why I pushed for related work sections. The first is because I thought a lot of people on LessWrong were doing work that either already existed in academia, or that would’ve benefited substantially from knowing the literature. I thought that making people write related work sections was a good way to make them read the existing literature (at least enough that they can explain how their work differs from existing work). The second is because I thought it was important to integrate with academics in general, both in an effort to recruit them for particular research topics in AI, but also because I thought there was a lot of great work in the independent AI Safety community that the academics weren’t engaging with. I saw following academic norms such as related work sections as a way to improve communication with academics.
I also thought the reasons that people cited at the time for why they didn’t think related work sections were valuable were kind of silly. For example, I remember that some people were very put off by having to write these sections because they thought of related work sections as being forced to share credit with academics. (In my opinion, a good related work section is often an explanation for why the existing academic work is bad.) Other people thought that there was nothing of value in the academic literature, perhaps because their topics were so unique, or perhaps because academics were so unreasonable, that there was not much of a point in engaging. (While I do think it’s true that the academic literature rarely contains exactly the answer you’re looking for – the people who dismissed AI safety work as “already done in 1985” were generally just incorrect – it often has useful ideas!)
In 2026, the situation is quite different. Firstly, as AI safety topics have become more mainstream and more academic, there’s less of a need to push people to cater more to academic audiences on the margin.
More importantly, I’ve become more disillusioned with the quality of many related work sections. It seems that while people are forced to at least know the names of some of the authors, it rarely causes them to do a more systematic search of what work already exists. And based on my experience reviewing for ML conferences since 2022, many related work sections seem to flagrantly misrepresent the work that they’re citing, or at least suggest the author has done nothing more than read the abstract of the paper without engaging with the core ideas. Insofar as “write related work” is a way to make people engage with the literature, I think it doesn’t really accomplish this task. (Relatedly, due to the prevalence of LLMs, writing a related work section now requires even less engagement with the literature.)
A related reason is that I’ve also become more convinced that “knowing the literature” is, for most people, much harder of a task than I thought it would be. There are many, many papers, and most researchers correctly spend most of their time working on their own research (rather than reading other people’s research). On top of that, I’ve come to realize that the reasons more junior people don’t know the literature is not because they don’t think it’s important, but because they often lack some knowledge in how to actually understand the literature (e.g. how to quickly read papers, how to check that you actually understood an idea, etc). Pushing more formalities that are supposed to cause people to “know the literature” generally doesn’t actually help much, at least compared to passing on research skills or writing literature reviews myself.
That is to say: I still think it’s quite valuable, as a researcher, to have a big-picture sense of what work is happening regarding the topics that would go into your related work section. But I no longer think that pushing for related work sections are an effective way to make this happen on the margin.
Here's an excerpt from the draft, to give a sense of what is in it:
Abstract: Related work sections are required for academic publications but are generally absent from LessWrong and Alignment Forum posts. I outline the good and bad reasons to include related work sections in technical work. I then argue that the lack of related work sections (and in general, a lack of citations) on LW/AF both leads to reinventing the wheel—that is, multiple people spending time rederiving the same results—and also causes posts to be less clear than they can be about their core contributions or key claims. I conclude by speculating why this happens, and what can be done to encourage more citations of related work.
Written very quickly for the Inkhaven Residency.
As I dug through my drafts for today’s InkHaven’s post, I found an old draft of a blog post written in late 2022, arguing that LessWrong/Alignment Forum/etc should have a stronger norm for including related work sections.[1] When I first started posting actively on LessWrong and meddling in papers back in 2022, one of the things I pushed for was substantial related work sections. Nowadays, I'm a bit more mixed on the value of related work sections on the margin. So I’m writing a post about why I was a big proponent of more effort on related work sections in 2022, and why I’m not as large of a proponent nowadays.
Let’s start by explaining the situation as I remember it in 2022. At the time, research posts on LessWrong had quite limited related work sections, if they had it at all. For example, Neel Nanda’s Alignment Forum post on his work on reverse engineering a modular arithmetic model had no related work section. John Wentworth’s work on natural abstractions did reference some related work in fields such as physics or statistics, but did not reference similar work in the field of machine learning (e.g. representation learning or the universality hypothesis).
There were several reasons why I pushed for related work sections. The first is because I thought a lot of people on LessWrong were doing work that either already existed in academia, or that would’ve benefited substantially from knowing the literature. I thought that making people write related work sections was a good way to make them read the existing literature (at least enough that they can explain how their work differs from existing work). The second is because I thought it was important to integrate with academics in general, both in an effort to recruit them for particular research topics in AI, but also because I thought there was a lot of great work in the independent AI Safety community that the academics weren’t engaging with. I saw following academic norms such as related work sections as a way to improve communication with academics.
I also thought the reasons that people cited at the time for why they didn’t think related work sections were valuable were kind of silly. For example, I remember that some people were very put off by having to write these sections because they thought of related work sections as being forced to share credit with academics. (In my opinion, a good related work section is often an explanation for why the existing academic work is bad.) Other people thought that there was nothing of value in the academic literature, perhaps because their topics were so unique, or perhaps because academics were so unreasonable, that there was not much of a point in engaging. (While I do think it’s true that the academic literature rarely contains exactly the answer you’re looking for – the people who dismissed AI safety work as “already done in 1985” were generally just incorrect – it often has useful ideas!)
In 2026, the situation is quite different. Firstly, as AI safety topics have become more mainstream and more academic, there’s less of a need to push people to cater more to academic audiences on the margin.
More importantly, I’ve become more disillusioned with the quality of many related work sections. It seems that while people are forced to at least know the names of some of the authors, it rarely causes them to do a more systematic search of what work already exists. And based on my experience reviewing for ML conferences since 2022, many related work sections seem to flagrantly misrepresent the work that they’re citing, or at least suggest the author has done nothing more than read the abstract of the paper without engaging with the core ideas. Insofar as “write related work” is a way to make people engage with the literature, I think it doesn’t really accomplish this task. (Relatedly, due to the prevalence of LLMs, writing a related work section now requires even less engagement with the literature.)
A related reason is that I’ve also become more convinced that “knowing the literature” is, for most people, much harder of a task than I thought it would be. There are many, many papers, and most researchers correctly spend most of their time working on their own research (rather than reading other people’s research). On top of that, I’ve come to realize that the reasons more junior people don’t know the literature is not because they don’t think it’s important, but because they often lack some knowledge in how to actually understand the literature (e.g. how to quickly read papers, how to check that you actually understood an idea, etc). Pushing more formalities that are supposed to cause people to “know the literature” generally doesn’t actually help much, at least compared to passing on research skills or writing literature reviews myself.
That is to say: I still think it’s quite valuable, as a researcher, to have a big-picture sense of what work is happening regarding the topics that would go into your related work section. But I no longer think that pushing for related work sections are an effective way to make this happen on the margin.
Here's an excerpt from the draft, to give a sense of what is in it: