From the "obvious-but-maybe-worth-mentioning" file:

ChatGPT (4 and 4o at least) cheats at 20 questions:

If you ask it "Let's play a game of 20 questions. You think of something, and I ask up to 20 questions to figure out what it is.", it will typically claim to "have something in mind", and then appear to play the game with you.

But it doesn't store hidden state between messages, so when it claims to "have something in mind", either that's false, or at least it has no way of following the rule that it's thinking of a consistent thing throughout the game. i.e. its only options are to cheat or refuse to play.

You can verify this by responding "Actually, I don't have time to play the whole game right now. Can you just tell me what it was you were thinking of?", and then "refreshing" its answer. When I did this 10 times, I got 9 different answers and only one repeat.

Recently I tried to use Google to learn about the structure of the human nasal cavity & sinuses, and it seems to me that somehow medical illustrators haven't talked much to mechanical draftspeople. Just about every medical illustration I could find tried to use colors to indicate structure, and only gave a side-view (or occasionally a front view) of the region. In almost none of the illustrations was it clear which parts of your nasal cavity and sinuses are split down the middle of your head, vs joined together. I still feel pretty in-the-dark about it.

In drafting, you express 3d figures by drawing a set of multiple projections: Typically, you give a top view, a front view, and a side view (though other views, including cross-sections and arbitrary isometric perspective, may be useful or necessary). This lets you give enough detail that a (practiced) viewer can reconstruct a good mental model of the object, so that they can (for example) use their machine shop to produce the object out of raw material.

There's a pretty fun puzzle game that lets you practice this skill called ".projekt"; there are probably lots more.
 

As long as I'm using shortform posts to make feature requests, it would be really useful to me to have access to a feed (of shortform posts, normal posts, or both) where I could select which users I see. Right now I come to LessWrong and have a hard time deciding which posts I care about - lots of people here have lots of interests and lots of different standards for content quality, some of which I find actively annoying. Allowing me to build feeds from custom lists of selected users would let me filter by both shared interests and how valuable I typically find those users' posts. I don't think "who is in my custom feed" should be public like it is on Facebook, but Facebook circa 2012 gave me a lot of control over this via friend lists, and I miss the days when that feature was prioritized.

I also like the way lobste.rs solves this problem for some users, although I think the solution above would be better for me personally: every post comes with a selection from a group of site-wide tags, and users can filter their home page based on which tags they want to see.

I really think LessWrong would benefit by giving users avatars. I think this would make the site much more visually appealing, but I also think it would vastly decrease the cognitive load required to read threaded conversations.

There should really be a system that does what WebMD tried to do, but actually does it well.

You'd put in your symptoms and background info (e.g. what country you live in, your age), it would ask you clarifying questions ("how bad is the pain from 1 to 10?" "which of these patterns is most like the rash?" "Do you have a family history of heart disease?") and then it would give you a posterior distribution over possible conditions, and a guess about whether you should go to the emergency room or whatever.

Is this just much harder than I'm imagining it would be? It seems like the kind of thing where you could harvest likelihood ratios and put them all into a big database. Is there some regulatory thing where you can't practically offer this service because it's illegal to give medical advice or something?

Beth Barnes notices: Rationalists seem to use the word "actually" a lot more than the typical English speaker; it seems like the word "really" means basically the same thing.

We wrote a quick script, and the words "actually" and "really" occur about equally often on LessWrong, while Google Trends suggests that "really" is ~3x more common in search volume. SSC has ~2/3 as many "actually"s as "really"s.

What's up with this? Should we stop?

Doom circles seem hard to do outside of CFAR workshops: If I just pick the ~7 people who I most want to be in my doom circle, this might be the best doom circle for me, but it won't be the best doom circle for them, since they will mostly not know each other very well.

So you might think that doing doom "circles" one-on-one would be best. But doom circles also have a sort of ceremony / spacing / high-cost-ness to them that cuts the other way: More people means more "weight" or something. And there are probably other considerations determining the optimal size.

So if you wanted to have a not-at-the-end-of-a-workshop doom circle, should you find the largest clique with some minimum relationship strength in your social graph?

Yet another Shortform-as-feature-request:

Notifications and/or RSS feeds from particular posts' comments / answers.

This would be especially useful for Questions and Shortform posts (sometimes tellingly mis-labeled "shortform feeds"), both of which are things where one particular post has a collection of related comments, and which gather content over time.

I currently subscribe to the front page in Feedly, and whenever someone asks a question that I find interesting I mentally cringe because I know that I'll have to remind myself to check back (and I probably will never actually check back).

I guess I could come up with some custom Zapier / IFTTT system for this if I spent a few hours on it, but I suspect this would be generally useful functionality.

I'm interested in concrete ways for humans to evaluate and verify complex facts about the world. I'm especially interested in a set of things that might be described as "bootstrapping trust".

For example:

Say I want to compute some expensive function f on an input x. I have access to a computer C that can compute f; it gives me a result r. But I don't fully trust C - it might be maliciously programmed to tell me a wrong answer. In some cases, I can require that C produce a proof that f(x) = r that I can easily check. In others, I can't. Which cases are which?

A partial answer to this question is "the complexity class NP". But in practice this isn't really satisfying. I have to make some assumptions about what tools are available that I do trust.

Maybe I trust simple mathematical facts (and I think I even trust that serious mathematics and theoretical computer science track truth really well). I also trust my own senses and memory, to a nontrivial extent. Reaching much beyond that is starting to feel iffy. For example, I might not (yet) have a computer of my own that I trust to help me with the verification. What kinds of proof can I accept with the limitations I've chosen? And how can I use those trustworthy proofs to bootstrap other trusted tools?

Other problems in this bucket include "How can we have trustworthy evidence - say videos - in a world with nearly perfect generative models?" and a bunch of subquestions of "Does debate scale as an AI alignment strategy?"

This class of questions feels like an interesting lens on some things that are relevant to some sorts of AI alignment work such as debate and interpretability. It's also obviously related to some parts of information security and cryptography.

"Bootstrapping trust" is basically just a restatement of the whole problem. It's not exactly that I think this is a good way to decide how to direct AI alignment effort; I just notice that it seems somehow like a "fresh" way of viewing things.