I'm an editor who does feedback requests for LessWrong. Many of these feedback requests are for proofreading/clarity of language, so I have lots of experience reading community draft posts and helping improve them. I want to share some high level takeaways - low effort self-editing techniques that, based on my experience, I think many to most LW posters could benefit from.
I wrote all the example text in this post myself. I make no claims that any of it is reasonable on the object level - it's just meant to resemble lots of the stuff I edit.
Bolded text in examples is for emphasis of the points under discussion, not a sign that it would make sense to actually bold those words or phrases.
As usual, reverse advice as necessary.
Here's my main advice, in decreasing order of confidence - I'll elaborate on each:
The problem: many sentences on LessWrong begin with "this". At worst, the thing "this" is pointing to is fully ambiguous between multiple live options, making it impossible to parse the author's intended meaning. But even in milder cases, it can waste a few seconds for the reader to trace out what a post is talking about.
The solution: scan through your draft for words like "this" or "that" (but for whatever reason overwhelmingly "this" specifically), and when in doubt about clarity, just replace them with whatever their intended antecedents are.
Many key figures in the community have been shortening their timelines recently, since major advances just keep coming one after the other. This is especially worrying, especially compared to discussions of "AI winter" just a few years ago.
What, precisely, does "this" point to in the sentence above? The fact of key figures shortening their timelines? The major advances? When we get later in the sentence and see that we're comparing whatever "this" is to "discussions", we can infer that we're talking about the shortening of timelines (vs. the actual technical advances), but it takes up precious reader bandwidth to figure this out, or, worse, the reader becomes overloaded by ambiguity and unconsciously switches to "skimming mode".
Individually, overuse of "this" isn't a big deal. But it adds up. It's the single greatest contribution to a hand-wavy feeling in the posts I review, like maybe the author isn't totally sure what they're saying. And it's easy to avoid! Just ctrl-f "this" and go to town!
The problem: LessWrongers care a lot about making factually accurate claims. This (gasp!) is a virtue. But like any virtue, it's vulnerable to Goodharting: when people learn that hedging is a sign of epistemic responsibility, you end up with a whole lot of hedging. Occasionally, you actually end up with hedging that is logically incoherent or actively redundant.
The solution: Make sure you hedge precisely as much as necessary, and especially make sure you don't hedge the same claim multiple times in the same way. If "you suspect" something "might" happen, it doesn't also have to be "possible" that it "could" have "potentially" significant effects.
The other salient quality of insect suffering is its sheer scale. Insects are the most populous animals on the planet, so it may be that if their suffering is morally relevant at all, I believe it could swamp almost any other considerations about animal welfare.
Some hedging is virtuous here. We're dealing with uncertainty, so it'd be misleading to follow typical writing guides and say something bold like: "Because there are so many insects, their suffering adds up to more than that of larger animals." But in this example, the writer has gone too far. Taken literally, they are speculating on their own belief, and not any object level claim. But they have direct access to their belief! They can just say what they believe full stop, and have easier-to-read and more literally accurate text. Something like:
Insects are the most populous animals on the planet, so I believe that if they can suffer in a morally relevant way, their suffering in aggregate might exceed that of larger animals.
We're importantly still hedging here! We're just not doing it in so many layers of the hypothetical that it becomes hard to read and incoherent when fully parsed.
The problem: LessWrongers tend to write really long sentences. It's an understandable temptation: we're often talking about genuinely complex ideas, and the sentence is traditionally the vehicle for transmitting a single idea. But I think there's some Goodharting here, too. Many of LessWrong's top authors have a breathless, low-punctuation style, which others try to emulate. That style is fine if you're good at it, but it's dangerous for beginners or people less confident in their writing.
People read differently. Some people, like me, tend to actively hear text using their inner voice. It's harder to parse super duper long sentences for people like us. And even for people who have a different cognitive experience of reading, I suspect very long sentences cause some overload. Plus, it's easier to get lost as a writer in long sentences, too, and fall into sloppier thinking/excess hedging/hand-waving.
The solution: Keep it tight! Sometimes this requires replacing entire sentences rather than tuning them internally. But it's worth it, and an external editor may not be bold enough (or have enough time) to suggest this kind of change, so it's extra good to do it yourself.
We have to worry about mesa optimizers even if we can guarantee that the outer level optimizer is myopic and we don't currently have any promising solutions to this issue, though robust interpretability tools could help us figure out and take preventative measures against mesa optimizers as they arise.
Sentences like this one are hard to read, and there are too many opportunities for sneaky ambiguity or hand-waving. Here's a broken up version:
We might hope that myopia could protect us from unexpected goals or behaviors. But even if an outer level optimizer is provably myopic, it might chance upon a mesa optimizer that isn't. We don't have any current techniques to get around this issue, though interpretability tools seem promising here: if we can see mesa optimizers as they arise, we may be able to quash dangerous ones early.
This rework is a little longer, but I think it's a lot clearer. Readers should immediately know what the writer is talking about, and can take issue with the individual points atomically, rather than having a vague feeling that something might be off or confusing.
The problem: A lot of writing in LessWrong is high context. And worse, sometimes different people have different interpretations of that context. I've noticed in particular people taking "Goodhart" to mean anything from "A measure made a goal ceases to be a good measure" to "Literally stated objectives and fuzzily meant objectives diverge dangerously at the limit". These interpretations rhyme, but if you just say "Goodhart" and leave it at that, then, well, your readers' interpretations may also diverge at the limit. Plus you may have a new reader to the site who just gives up from too many terms and acronyms they don't recognize.
The solution: Aim your writing to people who know significantly less than your actual target audience. You may be overestimating your target audience's understanding, or their median attention span, or the degree to which their precise interpretations of terms align with your own. So use lots of links, diagrams and examples for bonus opportunities to let something "click", and commas to help readers parse conceptual boundaries within sentences.
It's clearly past time for someone to pull the fire alarm but it isn't clear either what the fire alarm mechanism is or who could actually pull it.
We're assuming readers are familiar with this "fire alarm" term. Linking the original post, plus adding a comma, makes for a much better reader experience:
It's clearly past time for someone to pull the fire alarm, but it isn't clear either what the fire alarm mechanism is or who could actually pull it.
One other benefit of going to find the link is you might realize that your interpretation of a term is a little off. I chose the "fire alarm" example because the entire point of the introducing post is that there isn't a fire alarm. Discovering this could be a good chance to drop the term and say something more aligned (but less jargon-y) that better captures your meaning, like:
It's clearly past time for it to be common knowledge that AGI is dangerously close, but it isn't clear how best to communicate this fact to key players such as top labs, governments, and arguably the public at large.
Sometimes when trying to remove a term that means something a little different than you thought, you'll discover that you didn't actually have a crisp meaning behind the term, and that rephrasing something in your own words is hard. This confusion (and noticing it) is a feature, not a bug.
Here's a final example of some text that could be significantly improved, just by keeping the points in this post in mind:
Top researchers are exploring many directions like corrigibility, interpretability, safety through debate and myopia but I think the general consensus is that none seems particularly likely to be promising yet, and worse it's not clear more money or even more marginal researchers might help very much with this.
Bolded words aren't necessarily problematic, just things that an editing process of the type I'm advocating here would flag. Here's one possible rewrite:
Top AGI researchers are exploring a variety of research directions, including corrigibility, interpretability, safety through debate, and myopia. But there's no consensus view that any of these, or others I've left out, are especially promising. Worse, it's not clear how to either make current avenues of research more effective, or to find new avenues that might work better: neither more money nor more researchers would necessarily do the trick.
It's easier to disagree with this reworked version - the choice of links may not be optimal, the claims are more direct, and it's simply easier to understand what the author is trying to say. It can be uncomfortable to write in a clearer, more beginner-friendly style, but in my view, the rewards are worth it.
Editing is hard, and requires lots of practice. The LessWrong feedback feature continues to be open for business, and in my (biased) view, most people who use it seem to be satisfied with the outcome (and request more later). So if you'd like to improve your writing along these or other axes, don't be shy about asking for help!
Before I went through and edited the post, I had the word "this" here. Truly no one is safe from "this"!
Love this! I used to manage teams of writers/editors and here are some ideas we found useful for increasing readability:
To remove fluff, imagine someone is paying you $1,000 for every word you remove. Our writers typically could cut 20-50% with minimal loss of information.
Long sentences are hard to read, so try to change your commas into periods.
Long paragraphs are hard to read, so try to break each paragraph into 2-3 sentences.
Most people just skim, and some of your ideas are much more important than others, so bold/italicize your important points.
Shoulder!Justis telling me to replace "it", "this", etc with real nouns is maybe the most legible improvement in my writing over the last few years.
Good advice all around, thanks for writing this.
Someone told me in grade school: “Never use ‘this’ as a noun, only an adjective”. I think it's a good rule and I believe that I've followed it ever since then.
(I think that your complaint about ‘this’ pretty much exclusively applies to this-as-a-noun, not this-as-an-adjective, although you're welcome to disagree.)
Yeah, that's a good pithy summary! I often suggest replacing "this" with "this [x]".
What do you think about encouraging writers to add TLDRs on top of their posts? TLDRs make the purpose and content immediately clear so readers can decide whether to read on, and it plausibly also helps the writers to be more focused on their key points. (Advice that’s emphasized a lot at Rethink Priorities.)
If your article could be substituted with a TLDR, then just write that in the first place.
IMO, articles should include TLDRs, but shouldn't just be TLDRs. You have a short, high-context, high-trust summary. Then you write a longer article for people who don't have all the necessary background to understand your summary, or don't immediately trust that your summary is correct.
As a silly example, if you did an experiment to determine the acceleration due to gravity, your TLDR could simply be, "The acceleration due to gravity is 9.8 m/s^2." And for many readers, that's all they need to know. But you should definitely also explain your methodology and present the data from your experiment.
use lots of links
I agree with most of the post; I don't really agree with your use of links.
I find defining terms - either inline or a footnote - better than using links. (Partially because links - especially external links - can change out from under you! Partially because the article is no longer standalone, and can't be e.g. saved to read offline. And partially because linking to the 'original' definition of something can result in a mismatch where the original definition is X and you're using it as X', but your link to the definition still says it's X.)
(Now, this isn't to say "don't use links". Linking elsewhere is useful. Just not as a substitute for definitions please.)
Yeah, that critique is part of why "use more links" is among my least confident advice of the stuff in this post. I like links mostly as an alternative to nothing - if there's a term of background that ideally your readers should already know, a link is an economical way to give readers below your target audience in background knowledge a leg up. But for really central terms, yeah, better to summarize in your own words.
I find footnotes superior for that sort of thing.
I actively dislike the use of inline links. When I am reading an article and there is a link I now have a choice, continue reading or follow that link. I immediately feel myself leaving the thread of the article itself to make a quick assessment about whether the linked material likely to be more or less interesting than the article I am currently reading. Even in the best case this is a small speedbump that can break flow.
I think that maybe the links can all be at the end in a reference section like place. It might feel quite formal, but it doesn't break the flow.
This fits into my bigger impression that the second-best writing is easily navigable, with headers and signposts that can let me find the bits I want. The best writing just sucks me in from the first sentence and I don't at any point even think to assess which bits might be relevant because I am so immersed.
Thank you for the post, and thank you for all the editing you've done!
Interesting. Grammarly always points at the "This" sentences as problems in my drafts. For a while I looked at them; in approximately 100% of cases I've decided that the sentence is clear and made no change. Nowadays, I dismiss all such warnings without looking. This :) has sped up the process. I guess I'll give them more consideration from now on.
Grammarly also points to some of the other things you mention in the post, which I tend to take seriously.