Question for people who vibe-code full apps, especially blind people, as I believe that covers most of my followers who do this: What would it take for you to be comfortable coding directly in a programming language without an LLM? A high-level programming language with better tutorials and quickstart templates than are currently available, particularly for writing real apps? Or do you believe that having an LLM generate the code is the only way you could practically do what you do?
@menelion @matt as someone who actually knows how to do this already i'd say it's really just the overwhelming amount of broken that stops me from hand-making everything. The code an AI spits out is generally pretty awful quality-wise, and you do have to run around the watercooler with it a few times now and again, but particularly for static tooling that doesn't change, the speed it creates fixes with is difficult to eat
Replying to @zersiax@cupoftea.social
@zersiax @menelion @matt So much this, and everything out there is piles and piles of complexity which I guess I could learn about if it was actually something I needed regularly but as a one-off it's more a waste of time than anything. I just integrated Android's Activity Recognition API into another project which, from what I remember of my research years back, involves lots of Google Play Services-specific RPC incantations for setting up the connection. Once that's done, I'm not going to need that knowledge again, and while the time I spent learning things for a one-off task is arguably not wasted, it's time I can never get back. That plus the brokenness of everything are probably my biggest struggles, and I can't imagine being new to this and eventually hitting those.
Also, the speed at which Claude takes 2 multi-megabyte log files and constructs abstract shell pipelines to extract just what it needs is dizzying. That's something I gather sighted folks have always been able to do way more quickly by skimming a usually ANSI-color-coded log. Even if its logic and process makes perfect sense to me, I'm not capable of slinging enough awk and sed to point at a log and identify all app-specific activity around crashes or RPC failures between this and that timestamp.
All that is to say, as someone who also knows how to do this, it feels like a larger hill to climb for blind folks because we never really had great accessible tooling for ourselves in the first place. Expecting someone to learn development and eventually grapple with a huge pile of logs or some multi-billion-dollar megacorp's bespoke pile of shit to do a thing feels like a larger ask than ever before.
That said, I don't know where this leaves us. I don't know if it's just a matter of providing one set of specific language tooling, though. Also not a particularly big fan of the solution being bellowing at my tooling several times a day because it wrote War and Peace in the doc comments again...
Replying to @fastfinge@interfree.ca
@fastfinge @menelion @matt @nolan @zersiax so one thing that would be helpful is a straightforward alternative to sed/awk for finding relevant bits of logs and accessible summaries of graphs?
Replying to @wronglang@bayes.club
@wronglang @fastfinge @matt @nolan @zersiax I'll transcribe for you *one* line of logs, so you understand. Remember: one. line.
Left bracket, two thousand twenty-six dash zero seven dash twenty-eight T sixteen colon forty-seven colon fourteen point five hundred fifty-three thousand two hundred three plus zero zero colon zero zero, right bracket, doctrine.INFO colon Disconnecting left bracket right bracket left bracket right bracket
Notes:
1. It's an extremely primitive line of logs. Here it is in notation for normal people:
[2026-07-28T16:47:14.553203+00:00] doctrine.INFO: Disconnecting [] []
2. Punctuation symbols are in fact shortened for me, but I cannot turn them of because you know, a single semicolon in programming matters.
3. It would be super mega nice to be able to turn off timestamps, since, as you can see, it's the main pain in the bottom. However, they are not useless because sometimes you need to know when exactly this or that happened. Of course, I could set up several loggers, one with times, another without, and sometimes I do that (especially in Serilog-powered .NET apps). but still, if you need to sift through hundreds of kilobytes of them… you get it.
Replying to @menelion@dragonscave.space
@menelion @fastfinge @matt @nolan @zersiax yeah that's a great example of why this needs tooling! I cringed just starting to read that!
It always reminds me of this Victor Borge routine: https://youtu.be/NPc0aijY2o4
He was onto something.
YouTubeVictor Borge - Phonetic Punctuation (with Ronald Reagan!)Replying to @wronglang@bayes.club
@wronglang @fastfinge @matt @nolan @zersiax Wooow! Respect for knowing and remembering the great and mighty Victor Borge. BTW, Michel Leeb made a… remake of this in French which, in my opinion, is also very funny. https://www.youtube.com/watch?v=Fj3edfsJgGU
YouTubeMichel Leeb "La ponctuation" | Archive INAReplying to @menelion@dragonscave.space
@menelion @fastfinge @matt @nolan @zersiax that's beautiful, thank you!
Replying to @menelion@dragonscave.space
@menelion @fastfinge @matt @nolan @zersiax maybe a dumb question but do the screen readers have hooks for custom text parsing in the pipeline before it gets read out? At a minimum it should, in a better world, be capable of calling out a timestamp without reading it...
Replying to @wronglang@bayes.club
@wronglang @fastfinge @matt @nolan @zersiax It's my greatest pleasure to answer questions, so keep asking. Answer to your one is yes and no. Yes in the sense that we have dictionaries for the cases when TTS mispronounce something like a proper name (for example, a married couple of great people I know have last name Toews, but it's pronounced "Taves", to rhyme with "waves") — for such cases dictionary exists. Also, in JAWS there is a symbol descriptions file for various mathematical characters, emojis, additional letters, it stands just before the TTS, after all processing. However, more difficult hooks like you describe would have one destructive consequence — latency. Latency is the demon we fight at all costs when developing screen readers and TTS engines. Because it's like your latency when text is output to the screen. Imagine having *perceivable* latency between you typing a letter and it appearing on your monitor — it's not simply irritating, it's devastating for a working person.
Replying to @wronglang@bayes.club
Replying to @fastfinge@interfree.ca
@fastfinge @menelion @matt @nolan @zersiax right and what I'm saying is that the rolling could do a lot better here, you don't get plots without figuring out how to parse logs either, how to summarize information verbally vs visually is a challenge but it's not something (for the tooling, not for an individual, for an individual it's overwhelming)
Replying to @wronglang@bayes.club
Replying to @fastfinge@interfree.ca
@fastfinge @menelion @matt @nolan @zersiax yes and since AI gets that feature from being a specific kind of deep learning method I think this might be a good niche for an established machine learning method that a) burns less electricity and b) doesn't make shit up, you just got me thinking a little about which one that would be.
Replying to @wronglang@bayes.club
Replying to @fastfinge@interfree.ca
@fastfinge @menelion @matt @nolan @zersiax right, and log text has a simple structure with meaningful lines located in time with repeating glue text that has changing values attached so it's really structured data for analysis as a vector of words with a small vocabulary and (some) meaningful punctuation...
Uh, I'll poke/ask around if this is something that's been a target for model training and whether I can find anybody who has done it from this angle.
Replying to @wronglang@bayes.club
@wronglang @fastfinge @menelion @matt @nolan I'm halfway noodling around with making a video editor thats actually useful and THAT may honestly have to wait for a decent local model, like ... I can make an existing video editor more accessible, and I will, but to actually know what you're editing visually would take a lot more work and I don't think even ML can really help there