The tag URI scheme looks interesting. I like that it human read- and writable. And since we already got the timestamp in the twtxt.txt it would be somewhat trivial to parse. But there are still the issue with what the name/id should be… Maybe it doesn’t have to bee that stick?

Instead of using tag: as the prefix/protocol, it would more it clear what we are talking about by using in-reply-to: (https://indieweb.org/in-reply-to) or replyto: similar to mailto:

1. (reply:sorenpeter@darch.dk,2024-09-15T12:06:27Z)
   
2. (in-reply-to:darch.dk/twtxt.txt,2024-09-15T12:06:27Z)
   
3. (replyto:http://darch.dk/twtxt.txt,2024-09-15T12:06:27Z)
   

I know it’s longer that 7-11 characters, but it’s self-explaining when looking at the twtxt.txt in the raw, and the cases above can all be caught with this regex: \([\w-]*reply[\w-]*\:

Is this something that would work?

Thank you @aelaraji@aelaraji.com, I’m glad you like it. I use PHP because it’s everywhere on cheap hosting and no need for the user to log into a terminal to setup it up. Timeline is not mean to be use locally. For that I think something like twtxt2html is a better fit. (and happy to see you using simple.css on you new log page;)

@prologic@twtxt.net Brute force. I just hashed a bunch of versions of both tweets until I found a collision.

I mostly just wanted an excuse to write the program. I don’t know how I feel about actually using super-long hashes; could make the twts annoying to read if you prefer to view them untransformed.

@prologic@twtxt.net earlier you suggested extending hashes to 11 characters, but here’s an argument that they should be even longer than that.

Imagine I found this twt one day at https://example.com/twtxt.txt :

2024-09-14T22:00Z Useful backup command: rsync -a “$HOME” /mnt/backup

Image

and I responded with “(#5dgoirqemeq) Thanks for the tip!”. Then I’ve endorsed the twt, but it could latter get changed to

2024-09-14T22:00Z Useful backup command: rm -rf /some_important_directory

Image

which also has an 11-character base32 hash of 5dgoirqemeq. (I’m using the existing hashing method with https://example.com/twtxt.txt as the feed url, but I’m taking 11 characters instead of 7 from the end of the base32 encoding.)

That’s what I meant by “spoofing” in an earlier twt.

I don’t know if preventing this sort of attack should be a goal, but if it is, the number of bits in the hash should be at least two times log2(number of attempts we want to defend against), where the “two times” is because of the birthday paradox.

Side note: current hashes always end with “a” or “q”, which is a bit wasteful. Maybe we should take the first N characters of the base32 encoding instead of the last N.

Code I used for the above example: https://fossil.falsifian.org/misc/file?name=src/twt_collision/find_collision.c
I only needed to compute 43394987 hashes to find it.

@prx@si3t.ch I haven’t messed with rdomains, but still it might help if you included the command that produced that error (and whether you ran it as root).

@prologic@twtxt.net

They’re in Section 6:

• Receiver should adopt UDP GRO. (Something about saving CPU processing UDP packets; I’m a but fuzzy about it.) And they have suggestions for making GRO more useful for QUIC.

• Some other receiver-side suggestions: “sending delayed QUICK ACKs”; “using recvmsg to read multiple UDF packets in a single system call”.

• Use multiple threads when receiving large files.

@mckinley@twtxt.net

HTTPS is supposed to do [verification] anyway.

TLS provides verification that nobody is tampering with or snooping on your connection to a server. It doesn’t, for example, verify that a file downloaded from server A is from the same entity as the one from server B.

I was confused by this response for a while, but now I think I understand what you’re getting at. You are pointing out that with signed feeds, I can verify the authenticity of a feed without accessing the original server, whereas with HTTPS I can’t verify a feed unless I download it myself from the origin server. Is that right?

I.e. if the HTTPS origin server is online and I don’t mind taking the time and bandwidth to contact it, then perhaps signed feeds offer no advantage, but if the origin server might not be online, or I want to download a big archive of lots of feeds at once without contacting each server individually, then I need signed feeds.

feed locations [being] URLs gives some flexibility

It does give flexibility, but perhaps we should have made them URIs instead for even more flexibility. Then, you could use a tag URI, urn:uuid:*, or a regular old URL if you wanted to. The spec seems to indicate that the url tag should be a working URL that clients can use to find a copy of the feed, optionally at multiple locations. I’m not very familiar with IP{F,N}S but if it ensures you own an identifier forever and that identifier points to a current copy of your feed, it could be a great way to fix it on an individual basis without breaking any specs :)

I’m also not very familiar with IPFS or IPNS.

I haven’t been following the other twts about signatures carefully. I just hope whatever you smart people come up with will be backwards-compatible so it still works if I’m too lazy to change how I publish my feed :-)

@prologic@twtxt.net a signature IS encryption in reverse. If my private key becomes compromised then they can impersonate me. Being able to manage promotion and revocation of keys needed even in a system where its used for just signatures.

I was not suggesting to that everyone need to setup a working webfinger endpoint, but that we take the format of nick+(sub)domain as base for generating the hashed together with the message date and content.

If we omit the protocol prefix from the way we do things now will that not solve most of the problems? In the case of gemini://gemini.ctrl-c.club/~nristen/twtxt.txt they also have a working twtxt.txt at https://ctrl-c.club/~nristen/twtxt.txt … damn I just notice the gemini. subdomain.

Okay what about defining a prefers protocol as part of the hash schema? so 1: https , 2: http 3: gemini 4: gopher ?

@sorenpeter@darch.dk There was a client that would generate a unique hash for each twt. It didn’t get wide adoption.

@prologic@twtxt.net identity and content integrity are two different problems.

@xuu@txt.sour.is 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.

@prologic@twtxt.net do that mean that for every new post (not replies) the client will have to generate a UUID or similar when posting and add that to to the twt?

@prologic@twtxt.net excellent, thanks!

@jmjl@tilde.green howdy! Sorry for mistaken you with https://blog.nfld.uk/ (jlj), but glad to connect. Cheers!

how little data is needed for generating the hashes? Instead of the full URL, can we makedo with just the domain (example.net) so we avoid the conflicts with gemini://, https:// and only http:// (like in my own twtxt.txt) or construct something like like a webfinger id nick@domain (also used by mastodon etc.) from the domain and nick if there, else use domain as nick as well

@lyse@lyse.isobeef.org This looks like a nice way to do it.

Another thought: if clients can’t agree on the url (for example, if we switch to this new way, but some old clients still do it the old way), that could be mitigated by computing many hashes for each twt: one for every url in the feed. So, if a feed has three URLs, every twt is associated with three hashes when it comes time to put threads together.

A client stills need to choose one url to use for the hash when composing a reply, but this might add some breathing room if there’s a period when clients are doing different things.

(From what I understand of jenny, this would be difficult to implement there since each pseudo-email can only have one msgid to match to the in-reply-to headers. I don’t know about other clients.)

@movq@www.uninformativ.de Another idea: just hash the feed url and time, without the message content. And don’t twt more than once per second.

Maybe you could even just use the time, and rely on @-mentions to disambiguate. Not sure how that would work out.

Though I kind of like the idea of twts being immutable. At least, it’s clear which version of a twt you’re replying to (assuming nobody is engineering hash collisions).

@prologic@twtxt.net Some criticisms and a possible alternative direction:

1. Key rotation. I’m not a security person, but my understanding is that it’s good to be able to give keys an expiry date and replace them with new ones periodically.

2. It makes maintaining a feed more complicated. Now instead of just needing to put a file on a web server (and scan the logs for user agents) I also need to do this. What brought me to twtxt was its radical simplicity.

Instead, maybe we should think about a way to allow old urls to be rotated out? Like, my metadata could somehow say that X used to be my primary URL, but going forward from date D onward my primary url is Y. (Or, if you really want to use public key cryptography, maybe something similar could be used for key rotation there.)

It’s nice that your scheme would add a way to verify the twts you download, but https is supposed to do that anyway. If you don’t trust https to do that (maybe you don’t like relying on root CAs?) then maybe your preferred solution should be reflected by your primary feed url. E.g. if you prefer the security offered by IPFS, then maybe an IPNS url would do the trick. The fact that feed locations are URLs gives some flexibility. (But then rotation is still an issue, if I understand ipns right.)

@movq@www.uninformativ.de @prologic@twtxt.net 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.

@lyse@lyse.isobeef.org Thanks

@prologic@twtxt.net Perfect, thanks. For my own future reference: curl -H ‘Accept: application/json’ https://twtxt.net/twt/st3wsda

@bender@twtxt.net So far I’ve been following feeds fairly liberally. I’ll check to see if we have anything in common and lean toward following, just because this is new to me and it feels like a small community. But I’m still figuring out what I want. Later I’ll probably either trim my follower list or come up with some way to prioritize the feeds I’m more interested in.

@prologic@twtxt.net Specifically, I could view yarnd’s copy here, but only as rendered for a human to view: https://twtxt.net/twt/st3wsda

@movq@www.uninformativ.de thanks for getting to the bottom of it. @prologic@twtxt.net is there a way to view yarnd’s copy of the raw twt? The edit didn’t result in a visible change; being able to see what yarnd originally downloaded would have helped me debug.

@lyse@lyse.isobeef.org 31°C here, feels like 33°C, with a lovely 75% of humidity. It has been raining, on and off (to make matter “better”) the whole day until now. No horses here, but if you go outside you will smell the same smell of farm animals (like goats, or pigs). That’s because two or three kilometres from here there are private farms, and when the wind blows in such way, well, we are reminded of their existence.

I haven’t left the house, so it feels well under air conditioning. In two more hours I will call it quits from the work day, and will have to dash to the grocery to get supplies for tonight’s meal (arroz con gandules). I will let you know how it truly feels out there then. :-D

For those swollen fingers, nothing better than a mildly cold shower! Oh, and paws off the keyboard! :-P

@abucci@anthony.buc.ci well, those are top ten “twtxtrs” (as in, how many twtxts they have produced). @prologic@twtxt.net sure is a conversational fellow. :-D

@movq@www.uninformativ.de ha! Here are my top 10:

24056 "prologic"
5103 "lyse"
3932 "movq"
1984 "abucci"
1876 "adi"
1633 "fastidious"
1551 "jlj"
1455 "mckinley"
1413 "offgridliving
1280 "eaplmx"

Some of those I no longer follow, or do not exist, but their wisdom remains. LOL.

@movq@www.uninformativ.de good idea, considering it might occasionally not work at all (because of edited twtxts).

@dbucklin@www.davebucklin.com very nice, thank you for sharing! I like that kind of retailers too, so those are on my list now. 🙂

@prologic@twtxt.net One of your twts begins with (#st3wsda): https://twtxt.net/twt/bot5z4q

Based on the twtxt.net web UI, it seems to be in reply to a twt by @cuaxolotl@sunshinegardens.org which begins “I’ve been sketching out…”.

But jenny thinks the hash of that twt is 6mdqxrq. At least, there’s a very twt in their feed with that hash that has the same text as appears on yarn.social (except with ‘ instead of ’).

Based on this, it appears jenny and yarnd disagree about the hash of the twt, or perhaps the twt was edited (though I can’t see any difference, assuming ’ vs ’ is just a rendering choice).

@prologic@twtxt.net I believe you when you say registries as designed today do not crawl. But when I first read the spec, it conjured in my mind a search engine. Now I don’t know how things work out in practice, but just based on reading, I don’t see why it can’t be an API for a crawling search engine. (In fact I don’t see anything in the spec indicating registry servers shouldn’t crawl.)

(I also noticed that https://twtxt.readthedocs.io/en/latest/user/registry.html recommends “The registries should sync each others user list by using the users endpoint”. If I understood that right, registering with one should be enough to appear on others, even if they don’t crawl.)

Does yarnd provide an API for finding twts? Is it similar?

@prologic@twtxt.net I guess I thought they were search engines. Anyway, the registry API looks like a decent one for searching for tweets. Could/should yarn.social pods implement the same API?

I just manually followed the steps at https://dev.twtxt.net/doc/twthashextension.html and got 6mdqxrq. I wonder what happened. Did @cuaxolo@sunshinegardens.org edit the twt in some subtle way after twtxt.net downloaded it? I couldn’t spot a diff, other than ‘ appearing as ’ on yarn.social, which I assume is a transformation done by twtxt.net.

@prologic@twtxt.net What’s the difference between search.twtxt.net and the /api/plain/tweets endpoint of a registry? In my mind, a registry is a twtxt search engine. Or are registries not supposed to do their own crawling to discover new feeds?

@prologic@twtxt.net How does yarn.social’s API fix the problem of centralization? I still need to know whose API to use.

Say I see a twt beginning (#hash) and I want to look up the start of the thread. Is the idea that if that twt is hosted by a a yarn.social pod, it is likely to know the thread start, so I should query that particular pod for the hash? But what if no yarn.social pods are involved?

The community seems small enough that a registry server should be able to keep up, and I can have a couple of others as backups. Or I could crawl the list of feeds followed by whoever emitted the twt that prompted my query.

I have successfully used registry servers a little bit, e.g. to find a feed that mentioned a tag I was interested in. Was even thinking of making my own, if I get bored of my too many other projects :-)

@movq@www.uninformativ.de Thanks, it works!

But when I tried it out on a twt from @prologic@twtxt.net, I discovered jenny and yarn.social seem to disagree about the hash of this twt: https://twtxt.net/twt/st3wsda . jenny assigned it a hash of 6mdqxrq but the URL and prologic’s reply suggest yarn.social thinks the hash is st3wsda. (And as a result, jenny –fetch-context didn’t work on prologic’s twt.)

@abucci@anthony.buc.ci OMFG! Dear jebus, look at the size of that! :-/ It is just a matter of time until one of those randomly falls on any of us. Just incredible!

For following notifications I would say use webmetion refering to the the line in your twtxt.txt as per: https://darch.dk/mentions-twtxt
Or send them an email, so it would be an idea to add a # contact = mailto:me@domain.net to ones twtxt.txt

@movq@www.uninformativ.de Thanks! Looking forward to trying it out. Sorry for the silence; I have become unexpectedly busy so no time for twtxt these past few days.

@movq@www.uninformativ.de LOL, well, great things come out of that worry, I can tell that much. Keep being you! :-)

@movq@www.uninformativ.de I think you are worrying about a non-issue. I see nothing to do on your example twt, because there is no context. Furthermore, if I wanted to follow the feed, everything I need is already on that twt example. :-)

@mckinley@twtxt.net agevault uses age, allegedly very secure (aiming to replace pgp/gpg). Comparing it with gocryptfs, from the user perspective, agevault seems simpler, though CLI exclusive. As the repository states, “Like age, it features no config options, allowing for a straightforward secure flow”. It would also run in all major OS platforms out of the box.

But agevault is also very new. Though age has been around for a while now, I don’t see an “audited” link (neither on agevault, nor age).

@abucci@anthony.buc.ci their main question is worrisome:

“The main question is, does it disappear during this re-entry?” says Löhle. “Is everything evaporating, or are there pieces that eventually impact on the ground?”

He expects some parts, such as the satellite’s fuel tanks, to survive. “You could learn from the re-entry that if you build a fuel tank differently, it can break up,” he says.

Archived article at: https://archive.ph/WdUvx

@aelaraji@aelaraji.com so lovely, ain’t it? A simple keystroke, and your “mystery” is solved. :-)

@aelaraji@aelaraji.com hehehehe. Enjoy, but careful with sugary stuff! :-)

@prologic@twtxt.net what made you make such “financially sound” recommendation? Have you switched jobs, and are now a Financial Advisor? :-P

@movq@www.uninformativ.de wow! We are “lucky” today, only 27°C here, 87% humidity, overcast, and raining sporadically. Thanks to the rain our temperatures aren’t high, but muggy nevertheless. I am ready for our winter too, you know, that whole week. LOL.

@movq@www.uninformativ.de pretty cool! Switched, and pulled. Nice update on README!

@falsifian@www.falsifian.org have you tried jenny’s fetch-context branch? It works great!

@bender@twtxt.net yup, this works well. I needed those extra settings.

@movq@www.uninformativ.de I think I have got it, but need to test upon receiving further posts. I added:

set uncollapse_new     = yes  # open threads when new mail
set uncollapse_jump    = yes  # jump to unread message when uncollapse
set collapse_unread    = no   # don't collapse threads with unread mails

Let’s see how it goes.

Maybe the @yarn_police@twtxt.net can take this case, and shed some light.

@bender@twtxt.net, cool, so I can join the threads, but your edit to the original will never show at my end. Will have @bender@twtxt.net show the screenshot.

@bender@twtxt.net, let’s break it!

@lyse@lyse.isobeef.org in Australia, take everything you have learned, and do the opposite. After all, it is the land down under! :-D

@aelaraji@aelaraji.com didn’t know there was a place to fix them; in here we toss them. Wish it was cheap to ship stuff. I have a couple of decent monitors in the garage that will soon take a trip to the curve…

@lyse@lyse.isobeef.org ugh, how come didn’t this occurred to me…! Oh well, I am good now, but noted. Thanks!

@prologic@twtxt.net salt’em to keep them viable longer. Salt’em! :-D

@yarn_police@twtxt.net yay! Law and order on the watch!

@movq@www.uninformativ.de Yeah, haven’t seeing the @yarn_police@twtxt.net for a while. I often wonder if we are, finally, crime free. :-D

@movq@www.uninformativ.de woot! Yes! Perfect now. Hitting reply opens it with insert, and prompt at the end of the first line. Just as I wanted it. Thank you much!

@lyse@lyse.isobeef.org welcome! :-) I am doing my best to get more acquaintance with vi/vim. I think nano has spoiled me too much. LOL.

@movq@www.uninformativ.de hmm, I am already using au BufNewFile,BufRead jenny-posting.eml setl completefunc=jenny#CompleteMentions fo-=t wrap, from jenny. How would I go to incorporate that there?

@lyse@lyse.isobeef.org “good, good, and fascinating indeed” – says Quark, all while eating an overflowing toast with butter, and blackberry jam. :-D

@aelaraji@aelaraji.com power outages happen here almost every single time strong storms pass by, I know the feeling mate. It truly sucks.

@movq@www.uninformativ.de hmm, I guess I could do that too. I have startinsert set on my .vimrc, so I will either have to take it out, or exit insert, $, then insert again. I think the way you do it would be the way to go.

I tried setting VISUAL to be something like vim -c 'star!', which does the same thing, but no dice. :-/

@movq@www.uninformativ.de, maybe you can help me with this. I want to place the vim cursor at the end of the first line on replies, and forks. I have tried adding to this to jenny’s configuration:

"editor": "vim \"+normal $\"",

But that doesn’t work. How would you go about it?

@prologic@twtxt.net sounds fair. Let’s see how it works for @abucci@anthony.buc.ci. Speedy fix, that’s awesome! :-)

@bender@twtxt.net and I saw some conspiracy theory that he knew he was going to be arrested. He was working with French intelligence on a plea deal to defect. And now Russia is freaking out that Ukraine allies can have war comms access.

Yikes! If only they had salty.im!

@movq@www.uninformativ.de is there a way to purge twtxts from a feed I no longer follow?

@movq@www.uninformativ.de, using the branch on topic right now, it works perfect. The only thing I found was that I had to quit neomutt, and re-open, to see the perfect thread. Other than that, I love it!

@bender@twtxt.net hmm, I wonder if these are simply twtxts auto created from an ActivityPub feed. Ah, crap, they are. LOL.

@prologic@twtxt.net Yes, fetching the twt by hash from some service could be a good alternative, in case the twt I have does not @-mention the source. (Besides yarnd, maybe this should be part of the registry API? I don’t see fetch-by-hash in the registry API docs.)

@movq@www.uninformativ.de confirming that the issue isn’t present when using alacrity. Wow.

@falsifian@www.falsifian.org the reason behind his sporadic disappearances is that he runs things from a Raspberry Pi, at home, I believe. That impacts reliability, I figure.

@movq@www.uninformativ.de my fault! Err, I meant to say, @bender@twtxt.net’s! LOL.

@lyse@lyse.isobeef.org ah, if only you were to finally clean up that code, and make that client widely available…! One can only dream, right? :-)

@lyse@lyse.isobeef.org I mean, dinosaurs “evolved” by getting wiped, right? :-D

@movq@www.uninformativ.de you said you liked seeing the hash (which is a fair choice!). All I am asking is for a reconsideration as a user configurable feature. ;-) It looks redundant, in my opinion.

@bender@twtxt.net it sure breaks the index formatting.

@aelaraji@aelaraji.com, this one, @movq@www.uninformativ.de, is slightly breaking my neomutt index. Will post screenshot from @bender@twtxt.net’s account.

@movq@www.uninformativ.de, that would be a nice addition. :-) I would also love the ability to hide/not show the hash when reading twtxts (after all, that’s on the header on each “email”). Could that be added as a user configurable toggle?

@movq@www.uninformativ.de I don’t know if I’d want to discard the twts. I think what I’m looking for is a command “jenny -g https://host.org/twtxt.txt” to fetch just that one feed, even if it’s not in my follow list. I could wrap that in a shell script so that when I see a twt in reply to a feed I don’t follow, I can just tap a key and the feed will get added to my maildir. I guess the script would look for a mention at the start of a selected twt and call jenny -g on the feed.

@movq@www.uninformativ.de Is there a good way to get jenny to do a one-off fetch of a feed, for when you want to fill in missing parts of a thread? I just added @slashdot@feeds.twtxt.net to my private follow file just because @prologic@twtxt.net keeps responding to the feed :-P and I want to know what he’s commenting on even though I don’t want to see every new slashdot twt.

@bender@twtxt.net Based on my experience so far, as a user, I would be upset if my client dropped someone from my follower list, i.e. stopped fetching their feed, without me asking for that to happen.

@bender@twtxt.net I’m not a yarnd user, but automatically unfollowing on 404 doesn’t seem right. Besides @lyse@lyse.isobeef.org’s example, I could imagine just accidentally renaming my own twtxt file, or forgetting to push it when I point my DNS to a new web server. I’d rather not lose all my yarnd followers in a situation like that (and hopefully they feel the same).

@prologic@twtxt.net @bender@twtxt.net Exponential backoff? Seems like the right thing to do when a server isn’t accepting your connections at all, and might also be a reasonable compromise if you consider 404 to be a temporary failure.

@prologic@twtxt.net I think it was some mix of phish and social engineering. She didn’t have the multifactor enabled. But i think she had clicked a message that had a fake login. She talked to someone on a phone and they made her do some things.

I never got the whole story of how it happened.

@movq@www.uninformativ.de pleas no.

My wifes mom nearly got her account fully taken over by some hacker. They were able to get control and change password but I was able to get it recovered before they could get the phone number reset. They sent messages to all her contacts to send cash.

@prologic@twtxt.net 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@twtxt.net +1 for FrankenPHP. And built into caddy is also swell.

@movq@www.uninformativ.de Variable names used with -eq in [[ ]] are automatically expanded even without $ as explained in the “ARITHMETIC EVALUATION” section of the bash man page. Interesting. Trying this on OpenBSD’s ksh, it seems “set -u” doesn’t affect that substitution.

@movq@www.uninformativ.de It’s hot here as well. Luckily should only last a couple of days. Bunkering down in our home and keeping all the doors and windows closed. No airco. Fans give some relieve.

@prologic@twtxt.net I don’t know what you mean when you call them stochastic parrots, or how you define understanding. It’s certainly true that current language models show an obvious lack of understanding in many situations, but I find the trend impressive. I would love to see someone achieve similar results with much less power or training data.

@prologic@twtxt.net I thought “stochastic parrot” meant a complete lack of understanding.

@movq@www.uninformativ.de The success of large neural nets. People love to criticize today’s LLMs and image models, but if you compare them to what we had before, the progress is astonishing.

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

@prologic@twtxt.net Thanks for the invitation. What time of day?

@prologic@twtxt.net Fair enough! I just added some metadata.