I am slightly worried about the rate at which LW is shipping new features. I’m not convinced they are net positive. I see lesswrong as a clear success, but unclear user of the marginal dollar; I see lighthaven as a moderate success and very likely positive to expand at the margin.
The interface has been getting busier[1] whereas I think the modal reader would benefit from having as few distractions as possible while reading. I don’t think an LLM-enhanced editor would be useful, nor am I excited about additional tutoring functionality.
I am glad to see that people are donating, but I would have preferred this post to carefully signpost the difference between status-quo value of LW (immense) from the marginal value of paying for more features for LW (possibly negative), and from your other enterprises. Probably not worth the trouble, but is it possible to unbundle these for the purposes of donations?
Separately, thank you to the team! My research experience over the past years has benefitted from LW on a daily basis.
EDIT: thanks to Habryka for more details. After comparing to previous site versions I’m more optimistic about the prospects for active work on LW.
Yeah, I think this concern makes a bunch of sense.
My current model is that LW would probably die a slow death within 3-4 years if we started developing at a much slower pace than the one which we have historically been developing. One reason for that is that that is exactly what happened with LW 1.0. There were mods, and the servers were kept on and bugs were being fixed, but without substantial technical development the site fell into disrepair and abandonment surprisingly quickly.
The feature development here is important in the eternal race against spammers and trolls, but the internet is also constantly shifting, and with new modalities of how to read and interact with ideas, it does matter to have an active development team, even just for basic traffic and readability reasons. LW 1.0 missed a bunch of the transition to mobile and this was a large component of its decline. I think AI chat systems are likely a coming transition where you really want a team to actively iterate on how to best handle that shift (my current guess is 15-20% of new users are already referred to the site because ChatGPT or Claude told them to read things here), but it might also end up something VR or AR shaped.
I also want to push back a bit on this sentence:
The interface has been getting busier whereas I think the modal reader would benefit from having as few distractions as possible while reading.
I actually think LessWrong is very unique in how it has not been getting busier! I really try extremely hard to keep UI complexity under control. As an illustration, here is a screenshot of a post page from a year ago:
Here is a screenshot of that page today:
I think the second one is substantially cleaner and less distracting. UI on LessWrong gets less busy as frequently as it gets busy!
Overall I am very proud of the design of the site, which successfully combines a really quite large (and IMO valuable) feature set with a very clean reading experience. Reducing clutter is the kind of iteration we’ve been doing a lot off just this year, meaning that is were a substantial chunk of marginal development resources are going into.
Of course, you might still overall disagree with the kind of feature directions we are exploring. I do think AI-driven features are really important for us to work on, and if you disagree with that, it makes sense to be less excited about donating to us. For another example of the kind of thing that I think could be quite big and useful, see Gwern’s recent comment: https://www.lesswrong.com/posts/PQaZiATafCh7n5Luf/gwern-s-shortform?commentId=KGBqiXrKq8x8qnsyH
But overall, we did sure grow LessWrong a lot, and I expect future development will do that as well. It’s from my perspective often extremely hard to tell which things cause the site to grow and get better, but as one example of a very recent change, our rework of shortform into Quick Takes and Popular Comments on the frontpage have I think enabled a way for new content to get written on the site that now hosts close to 40% of my favorite content, and I think that’s huge. And that very much is the kind of thing that marginal feature development efforts go into.
Due to various complications in how our finances are structured in the upcoming year, our ability to marginally scale up or down is also very limited, making a discussion of the value of marginal contributions a bit hard. As I outlined in my comment to David, we actually have very little staff specialized just for software engineering, and as I said in the post, we already are an organization that grows extraordinarily slowly. And in the coming year, $2M of our $3M in expenses are mortgage payments where if we fail to meet them, we would end up with something close to bankruptcy, so that’s not really a place where we can choose to spend less.
This means, that at most we could reduce our burn rate by ~20%, even if we got rid of all of our specialized software engineering staff and let go of a third of our core staff.
And getting rid of our core staff seems like a bad choice, even if I agreed with your point about the value of marginal feature development. I’ve invested enormous amounts of resources into developing a good culture among my staff, there are large fixed costs associated with hiring, and the morale effects of firing people are huge. And they are already selected for being the kind of people that will just work with me on scaling up and improving Lighthaven, or going into policy, or pivoting into B2B Saas, exactly because I recognize that indeed different projects will hit diminishing returns. As such, I think the much more likely thing to do here would be to keep costs the same, be convinced by good arguments that we should do something else, and then invest more into things other than LW.
I think this overall means that from the perspective of a donor, there isn’t really a way to unbundle investment into LessWrong or Lighthaven or other projects. Of course, we will take into account arguments and preferences from people who keep the show running, so making those arguments and sharing your preferences about where we should marginally allocate resources is valuable, but I don’t think this could up with a way to allow donors to directly choose into which project we will invest marginal resources.
Thanks for these details. These have updated me to be significantly more optimistic about the value of spending on LW infra.
The LW1.0 dying to no mobile support is an analogous datapoint in favor of having a team ready for 0-5 year future AI integration.
The head-to-head on the site updated me towards thinking things that I’m not sure are positive (visible footnotes in sidebar, AI glossary, to a lesser extent emoji-reacts) are not a general trend. I will correct my original comment on this.
While I think the current plans for AI integration (and existing glossary thingy) are not great, I do think there will be predictably much better things to do in 1-2 years and I would want there to be a team with practice ready to go for those. Raemon’s reply below also speaks to this. Actively iterating on integrations while keeping them opt-in (until very clearly net positive) seems like the best course of action to me.
I wrote some notes on how we’ve been working to keep UI simpler, but habryka beat me to it. Meanwhile:
Some thoughts Re: LLM integration
I don’t think we’ll get to agreement within this comment margin. I think there’s a lot of ways LLM integration can go wrong. I think the first-pass at the JargonBot Beta Test isn’t quite right yet and I hope to fix some of that soon to make it a bit more clear what it looks like when it’s working well, as proof-of-concept.
But, I think LLM integration is going to be extremely important, and I want to say a bit about it.
Most of what LLMs enable is entirely different paradigms of cognition, that weren’t possible before. This is sort of a “inventing cars while everyone is still asking for slightly better-horses, or being annoyed by the car-centric infrastructure that’s starting to roll out in fits and starts. Horses worked fine, what’s going on?”
I think good LLM integrations make the difference between “it’s exhausting and effortful to read a technical post in a domain you aren’t familiar with” (and therefore, you don’t bother) to “actually it’s not that much harder than reading a regular post.” (I think several UI challenges need to get worked out for this to work, but they are not particularly impossible UI challenges). This radically changes the game on what sort of stuff you can learn, and how quickly somewhat who is somewhat interested in a field can get familiar with it. You can just jump into the post that feels relevant, and have the gaps between your understanding and the cutting edge filled in automatically (instead of having to painstakingly figure out the basics of a field before you can start participating).
Once this is working reliably and you actually deeply believe in it, it opens up new atomic actions that you brain can automatically consider that would previously have been too expensive to be worth it.
I don’t think we even need advances on current LLM-skill for this to work pretty well – LLMs aren’t very good at figuring stuff out at the cutting edge, but they are pretty good at filling it details that get you up to speed on the basics, and I think it’s pretty obvious how to improve them along the edges here.
This is in addition to the very straightforward LLM-integrations into an editor that save obvious boring bits of work (identifying all typos and slight wording confusions and predictably hard-to-understand sections) and freeing up that attention for more complicated problem solving.
I think it’s important for LessWrong in particular to be at the forefront here, because there are gnarly important bottlenecking-for-humanity’s-future problems, that require people to skill up rapidly to have a hope of contributing in time. (My inspiration was a colleague kind of casually deciding “I think I’m going to learn about the technical problems underlying compute governance”, and spinning up into the field so they could figure out how to contribute)
I am slightly worried about the rate at which LW is shipping new features. I’m not convinced they are net positive. I see lesswrong as a clear success, but unclear user of the marginal dollar; I see lighthaven as a moderate success and very likely positive to expand at the margin.
The interface has been getting busier[1] whereas I think the modal reader would benefit from having as few distractions as possible while reading. I don’t think an LLM-enhanced editor would be useful, nor am I excited about additional tutoring functionality.
I am glad to see that people are donating, but I would have preferred this post to carefully signpost the difference between status-quo value of LW (immense) from the marginal value of paying for more features for LW (possibly negative), and from your other enterprises. Probably not worth the trouble, but is it possible to unbundle these for the purposes of donations?
Separately, thank you to the team! My research experience over the past years has benefitted from LW on a daily basis.
EDIT: thanks to Habryka for more details. After comparing to previous site versions I’m more optimistic about the prospects for active work on LW.
(edit) in some places, less busy in others
Yeah, I think this concern makes a bunch of sense.
My current model is that LW would probably die a slow death within 3-4 years if we started developing at a much slower pace than the one which we have historically been developing. One reason for that is that that is exactly what happened with LW 1.0. There were mods, and the servers were kept on and bugs were being fixed, but without substantial technical development the site fell into disrepair and abandonment surprisingly quickly.
The feature development here is important in the eternal race against spammers and trolls, but the internet is also constantly shifting, and with new modalities of how to read and interact with ideas, it does matter to have an active development team, even just for basic traffic and readability reasons. LW 1.0 missed a bunch of the transition to mobile and this was a large component of its decline. I think AI chat systems are likely a coming transition where you really want a team to actively iterate on how to best handle that shift (my current guess is 15-20% of new users are already referred to the site because ChatGPT or Claude told them to read things here), but it might also end up something VR or AR shaped.
I also want to push back a bit on this sentence:
I actually think LessWrong is very unique in how it has not been getting busier! I really try extremely hard to keep UI complexity under control. As an illustration, here is a screenshot of a post page from a year ago:
Here is a screenshot of that page today:
I think the second one is substantially cleaner and less distracting. UI on LessWrong gets less busy as frequently as it gets busy!
Overall I am very proud of the design of the site, which successfully combines a really quite large (and IMO valuable) feature set with a very clean reading experience. Reducing clutter is the kind of iteration we’ve been doing a lot off just this year, meaning that is were a substantial chunk of marginal development resources are going into.
Of course, you might still overall disagree with the kind of feature directions we are exploring. I do think AI-driven features are really important for us to work on, and if you disagree with that, it makes sense to be less excited about donating to us. For another example of the kind of thing that I think could be quite big and useful, see Gwern’s recent comment: https://www.lesswrong.com/posts/PQaZiATafCh7n5Luf/gwern-s-shortform?commentId=KGBqiXrKq8x8qnsyH
But overall, we did sure grow LessWrong a lot, and I expect future development will do that as well. It’s from my perspective often extremely hard to tell which things cause the site to grow and get better, but as one example of a very recent change, our rework of shortform into Quick Takes and Popular Comments on the frontpage have I think enabled a way for new content to get written on the site that now hosts close to 40% of my favorite content, and I think that’s huge. And that very much is the kind of thing that marginal feature development efforts go into.
Due to various complications in how our finances are structured in the upcoming year, our ability to marginally scale up or down is also very limited, making a discussion of the value of marginal contributions a bit hard. As I outlined in my comment to David, we actually have very little staff specialized just for software engineering, and as I said in the post, we already are an organization that grows extraordinarily slowly. And in the coming year, $2M of our $3M in expenses are mortgage payments where if we fail to meet them, we would end up with something close to bankruptcy, so that’s not really a place where we can choose to spend less.
This means, that at most we could reduce our burn rate by ~20%, even if we got rid of all of our specialized software engineering staff and let go of a third of our core staff.
And getting rid of our core staff seems like a bad choice, even if I agreed with your point about the value of marginal feature development. I’ve invested enormous amounts of resources into developing a good culture among my staff, there are large fixed costs associated with hiring, and the morale effects of firing people are huge. And they are already selected for being the kind of people that will just work with me on scaling up and improving Lighthaven, or going into policy, or pivoting into B2B Saas, exactly because I recognize that indeed different projects will hit diminishing returns. As such, I think the much more likely thing to do here would be to keep costs the same, be convinced by good arguments that we should do something else, and then invest more into things other than LW.
I think this overall means that from the perspective of a donor, there isn’t really a way to unbundle investment into LessWrong or Lighthaven or other projects. Of course, we will take into account arguments and preferences from people who keep the show running, so making those arguments and sharing your preferences about where we should marginally allocate resources is valuable, but I don’t think this could up with a way to allow donors to directly choose into which project we will invest marginal resources.
Thanks for these details. These have updated me to be significantly more optimistic about the value of spending on LW infra.
The LW1.0 dying to no mobile support is an analogous datapoint in favor of having a team ready for 0-5 year future AI integration.
The head-to-head on the site updated me towards thinking things that I’m not sure are positive (visible footnotes in sidebar, AI glossary, to a lesser extent emoji-reacts) are not a general trend. I will correct my original comment on this.
While I think the current plans for AI integration (and existing glossary thingy) are not great, I do think there will be predictably much better things to do in 1-2 years and I would want there to be a team with practice ready to go for those. Raemon’s reply below also speaks to this. Actively iterating on integrations while keeping them opt-in (until very clearly net positive) seems like the best course of action to me.
I wrote some notes on how we’ve been working to keep UI simpler, but habryka beat me to it. Meanwhile:
Some thoughts Re: LLM integration
I don’t think we’ll get to agreement within this comment margin. I think there’s a lot of ways LLM integration can go wrong. I think the first-pass at the JargonBot Beta Test isn’t quite right yet and I hope to fix some of that soon to make it a bit more clear what it looks like when it’s working well, as proof-of-concept.
But, I think LLM integration is going to be extremely important, and I want to say a bit about it.
Most of what LLMs enable is entirely different paradigms of cognition, that weren’t possible before. This is sort of a “inventing cars while everyone is still asking for slightly better-horses, or being annoyed by the car-centric infrastructure that’s starting to roll out in fits and starts. Horses worked fine, what’s going on?”
I think good LLM integrations make the difference between “it’s exhausting and effortful to read a technical post in a domain you aren’t familiar with” (and therefore, you don’t bother) to “actually it’s not that much harder than reading a regular post.” (I think several UI challenges need to get worked out for this to work, but they are not particularly impossible UI challenges). This radically changes the game on what sort of stuff you can learn, and how quickly somewhat who is somewhat interested in a field can get familiar with it. You can just jump into the post that feels relevant, and have the gaps between your understanding and the cutting edge filled in automatically (instead of having to painstakingly figure out the basics of a field before you can start participating).
Once this is working reliably and you actually deeply believe in it, it opens up new atomic actions that you brain can automatically consider that would previously have been too expensive to be worth it.
I don’t think we even need advances on current LLM-skill for this to work pretty well – LLMs aren’t very good at figuring stuff out at the cutting edge, but they are pretty good at filling it details that get you up to speed on the basics, and I think it’s pretty obvious how to improve them along the edges here.
This is in addition to the very straightforward LLM-integrations into an editor that save obvious boring bits of work (identifying all typos and slight wording confusions and predictably hard-to-understand sections) and freeing up that attention for more complicated problem solving.
I think it’s important for LessWrong in particular to be at the forefront here, because there are gnarly important bottlenecking-for-humanity’s-future problems, that require people to skill up rapidly to have a hope of contributing in time. (My inspiration was a colleague kind of casually deciding “I think I’m going to learn about the technical problems underlying compute governance”, and spinning up into the field so they could figure out how to contribute)