@prologic Here's mine.

@lyse I am a big fan of "obvious" math facts that turn out to be wrong. If you want to understand how reusing space actually works, you are mostly stuck reading complexity theory papers right now. Ian wrote a good survey: https://iuuk.mff.cuni.cz/~iwmertz/papers/m23.reusing_space.pdf . It's written for complexity theorists, but some of will make sense to programmers comfortable with math. Alternatively, I wrote an essay a few years ago explaining one technique, with (math-loving) programmers as the intended audience: https://www.falsifian.org/blog/2021/06/04/catalytic/ .

@lyse Still melting!

@sorenpeter Sorry, I realized that shortly after posting. Here's another attempt to post the images:

@prologic Have you tried Google's robots.txt report? https://support.google.com/webmasters/answer/6062598?hl=en . I would expect Google to be pretty good about this sort of thing. If you have the energy to dig into it and, for example, post on support.google.com, I'd be curious to hear what you find out.

@prologic I wrote 1/4 (one slash four) by which I meant "the first out of four". twtxt.net is showing it as ¼, a single character that IMO doesn't have that same meaning (it means 0.25). Similarly, 3/4 got replaced with ¾ in another twt. It's not a big deal. It just looks a little wrong, especially beside the 2/4 and 4/4 in my other two twts.

@prologic

See https://dev.twtxt.net

Yes, that is exactly what I meant. I like that collection and "twtxt v2" feels like a departure.

Maybe there's an advantage to grouping it into one spec, but IMO that shouldn't be done at the same time as introducing new untested ideas.

See https://yarn.social (especially this section: https://yarn.social/#self-host) -- It really doesn't get much simpler than this 🤣

Again, I like this existing simplicity. (I would even argue you don't need the metadata.)

That page says "For the best experience your client should also support some of the Twtxt Extensions..." but it is clear you don't need to. I would like it to stay that way, and publishing a big long spec and calling it "twtxt v2" feels like a departure from that. (I think the content of the document is valuable; I'm just carping about how it's being presented.)

@prologic Thanks for pointing out it lasts four hours. That's a big window! I wonder when most people will be on. I might aim for halfway through unless I hear otherwise. (12:00Z is a bit early for me.)

@movq Yes, the tools are surprisingly fast. Still, magrep takes about 20 seconds to search through my archive of 140K emails, so to speed things up I would probably combine it with an indexer like mu, mairix or notmuch.

@prologic Thanks for writing that up!

I hope it can remain a living document (or sequence of draft revisions) for a good long time while we figure out how this stuff works in practice.

I am not sure how I feel about all this being done at once, vs. letting conventions arise.

For example, even today I could reply to twt abc1234 with "(#abc1234) Edit: ..." and I think all you humans would understand it as an edit to (#abc1234). Maybe eventually it would become a common enough convention that clients would start to support it explicitly.

Similarly we could just start using 11-digit hashes. We should iron out whether it's sha256 or whatever but there's no need get all the other stuff right at the same time.

I have similar thoughts about how some users could try out location-based replies in a backward-compatible way (append the replyto: stuff after the legacy (#hash) style).

However I recognize that I'm not the one implementing this stuff, and it's less work to just have everything determined up front.

Misc comments (I haven't read the whole thing):

• Did you mean to make hashes hexadecimal? You lose 11 bits that way compared to base32. I'd suggest gaining 11 bits with base64 instead.

• "Clients MUST preserve the original hash" --- do you mean they MUST preserve the original twt?

• Thanks for phrasing the bit about deletions so neutrally.

• I don't like the MUST in "Clients MUST follow the chain of reply-to references...". If someone writes a client as a 40-line shell script that requires the user to piece together the threading themselves, IMO we shouldn't declare the client non-conforming just because they didn't get to all the bells and whistles.

• Similarly I don't like the MUST for user agents. For one thing, you might want to fetch a feed without revealing your identty. Also, it raises the bar for a minimal implementation (I'm again thinking again of the 40-line shell script).

• For "who follows" lists: why must the long, random tokens be only valid for a limited time? Do you have a scenario in mind where they could leak?

• Why can't feeds be served over HTTP/1.0? Again, thinking about simple software. I recently tried implementing HTTP/1.1 and it wasn't too bad, but 1.0 would have been slightly simpler.

• Why get into the nitty-gritty about caching headers? This seems like generic advice for HTTP servers and clients.

• I'm a little sad about other protocols being not recommended.

• I don't know how I feel about including markdown. I don't mind too much that yarn users emit twts full of markdown, but I'm more of a plain text kind of person. Also it adds to the length. I wonder if putting a separate document would make more sense; that would also help with the length.

@prologic Wikipedia claims sha1 is vulnerable to a "chosen-prefix attack", which I gather means I can write any two twts I like, and then cause them to have the exact same sha1 hash by appending something. I guess a twt ending in random junk might look suspcious, but perhaps the junk could be worked into an image URL like . If that's not possible now maybe it will be later.

git only uses sha1 because they're stuck with it: migrating is very hard. There was an effort to move git to sha256 but I don't know its status. I think there is progress being made with Game Of Trees, a git clone that uses the same on-disk format.

I can't imagine any benefit to using sha1, except that maybe some very old software might support sha1 but not sha256.

@prologic Why sha1 in particular? There are known attacks on it. sha256 seems pretty widely supported if you're worried about support.

@prologic

There's a simple reason all the current hashes end in a or q: the hash is 256 bits, the base32 encoding chops that into groups of 5 bits, and 256 isn't divisible by 5. The last character of the base32 encoding just has that left-over single bit (256 mod 5 = 1).

So I agree with #3 below, but do you have a source for #1, #2 or #4? I would expect any lack of variability in any part of a hash function's output would make it more vulnerable to attacks, so designers of hash functions would want to make the whole output vary as much as possible.

Other than the divisible-by-5 thing, my current intuition is it doesn't matter what part you take.

1. Hash Structure: Hashes are typically designed so that their outputs have specific statistical properties. The first few characters often have more entropy or variability, meaning they are less likely to have patterns. The last characters may not maintain this randomness, especially if the encoding method has a tendency to produce less varied endings.

2. Collision Resistance: When using hashes, the goal is to minimize the risk of collisions (different inputs producing the same output). By using the first few characters, you leverage the full distribution of the hash. The last characters may not distribute in the same way, potentially increasing the likelihood of collisions.

3. Encoding Characteristics: Base32 encoding has a specific structure and padding that might influence the last characters more than the first. If the data being hashed is similar, the last characters may be more similar across different hashes.

4. Use Cases: In many applications (like generating unique identifiers), the beginning of the hash is often the most informative and varied. Relying on the end might reduce the uniqueness of generated identifiers, especially if a prefix has a specific context or meaning.

@quark It looks like the part about traditional topics has been removed from that page. Here is an old version that mentions it: https://web.archive.org/web/20221211165458/https://dev.twtxt.net/doc/twtsubjectextension.html . Still, I don't see any description of what is actually allowed between the parentheses. May be worth noting that twtxt.net is displaying the twts with the subject stripped, so some piece of code is recognizing it as a subject (or, at least, something to be removed).

@movq

Maybe I’m being a bit too purist/minimalistic here. As I said before (in one of the 1372739 posts on this topic – or maybe I didn’t even send that twt, I don’t remember 😅), I never really liked hashes to begin with. They aren’t super hard to implement but they are kind of against the beauty of the original twtxt – because you need special client support for them. It’s not something that you could write manually in your twtxt.txt file. With @sorenpeter’s proposal, though, that would be possible.

Tangentially related, I was a bit disappointed to learn that the twt subject extension is now never used except with hashes. Manually-written subjects sounded so beautifully ad-hoc and organic as a way to disambiguate replies. Maybe I'll try it some time just for fun.

@xuu Thanks for the link. I found a pdf on one of the authors' home pages: https://ahmadhassandebugs.github.io/assets/pdf/quic_www24.pdf . I wonder how the protocol was evaluated closer to the time it became a standard, and whether anything has changed. I wonder if network speeds have grown faster than CPU speeds since then. The paper says the performance is around the same below around 600 Mbps.

To be fair, I don't think QUIC was ever expected to be faster for transferring a single stream of data. I think QUIC is supposed to reduce the impact of a dropped packet by making sure it only affects the stream it's part of. I imagine QUIC still has that advantage, and this paper is showing the other side of a tradeoff.

@movq @prologic Another option would be: when you edit a twt, prefix the new one with (#[old hash]) and some indication that it's an edited version of the original tweet with that hash. E.g. if the hash used to be abcd123, the new version should start "(#abcd123) (redit)".

What I like about this is that clients that don't know this convention will still stick it in the same thread. And I feel it's in the spirit of the old pre-hash (subject) convention, though that's before my time.

I guess it may not work when the edited twt itself is a reply, and there are replies to it. Maybe that could be solved by letting twts have more than one (subject) prefix.

But the great thing about the current system is that nobody can spoof message IDs.

I don't think twtxt hashes are long enough to prevent spoofing.

@prologic The headline is interesting and sent me down a rabbit hole understanding what the paper (https://aclanthology.org/2024.acl-long.279/) actually says.

The result is interesting, but the Neuroscience News headline greatly overstates it. If I've understood right, they are arguing (with strong evidence) that the simple technique of making neural nets bigger and bigger isn't quite as magically effective as people say --- if you use it on its own. In particular, they evaluate LLMs without two common enhancements, in-context learning and instruction tuning. Both of those involve using a small number of examples of the particular task to improve the model's performance, and they turn them off because they are not part of what is called "emergence": "an ability to solve a task which is absent in smaller models, but present in LLMs".

They show that these restricted LLMs only outperform smaller models (i.e demonstrate emergence) on certain tasks, and then (end of Section 4.1) discuss the nature of those few tasks that showed emergence.

I'd love to hear more from someone more familiar with this stuff. (I've done research that touches on ML, but neural nets and especially LLMs aren't my area at all.) In particular, how compelling is this finding that zero-shot learning (i.e. without in-context learning or instruction tuning) remains hard as model size grows.

@prologic Thanks. It's from a non-Euclidean geometry project: https://www.falsifian.org/blog/2022/01/17/s3d/