@glyph Did you quote post something?
Uncategorized
903
Posts
312
Posters
9.5k
Views
-
@agowa338 @tymwol @glyph I've done some experimentation and I found the tools to be extremely powerful in the way having a tracked excavator is more powerful than a hand shovel. Perhaps applying it takes some skill, but it magnifies human effort quite a lot. I've grown quite concerned that many people really underestimate the impact these tools are having and will have on a secular trend of power shifting away from labor.
I did some experiments too. I was actually quite happy with the very first version of ChatGPT. But any iteration on it just made it less capable and worse. They optimized it for cost which heavily impacted advanced capabilities that I enjoyed.
Same for codegen. I initially thought of it as advanced Intelisense. But then the more I (tried to) use it it just become more and more annoying as it didn't even manage to suggest the right APIs of the project itself.
-
Oh and all of them failed hilariously when I tasked them with doing some negative code edits. Like typical cleanup tasks where you'd replace large chunks of codes with a single API call.
-
And I think it also had some Syntax issues. But I don't remember which programming language I tried to use it with. Probably something PowerShell or C# isch but could have been something else too...
-
@agowa338 @tymwol @glyph It's interesting, this doesn't at all line up with my experience. The tools I tried were able to translate ~1 paragraph of instruction about a change into a point by point plan with references, correctly enough to need very few edits, and then to execute it in one go. Similarly, to take a pasted error (I tried a few, the interesting one was a infrequently manifesting concurrency problem), diagnose two likely causes, and patch.
-
@agowa338 @tymwol @glyph It's interesting, this doesn't at all line up with my experience. The tools I tried were able to translate ~1 paragraph of instruction about a change into a point by point plan with references, correctly enough to need very few edits, and then to execute it in one go. Similarly, to take a pasted error (I tried a few, the interesting one was a infrequently manifesting concurrency problem), diagnose two likely causes, and patch.
@mirth @agowa338 @tymwol sometimes it does amazing things, sometimes it wastes hours of your time. is it a net benefit? unclear. it sure *does stuff* but I haven't yet seen anything that indicates it reliably does enough stuff correctly to be an improvement to efficiency. c.f.: https://blog.glyph.im/2025/08/futzing-fraction.html
-
@mirth @agowa338 @tymwol sometimes it does amazing things, sometimes it wastes hours of your time. is it a net benefit? unclear. it sure *does stuff* but I haven't yet seen anything that indicates it reliably does enough stuff correctly to be an improvement to efficiency. c.f.: https://blog.glyph.im/2025/08/futzing-fraction.html
@mirth @agowa338 @tymwol more broadly speaking, there are some strong indications that on balance it isn't delivering on its promise. A recent roundup from @davidgerard: https://pivot-to-ai.com/2025/09/09/if-ai-coding-is-so-good-where-are-the-little-apps/
-
@glyph Ow, this "title butchering" is really concerning…
-
@mirth @agowa338 @tymwol more broadly speaking, there are some strong indications that on balance it isn't delivering on its promise. A recent roundup from @davidgerard: https://pivot-to-ai.com/2025/09/09/if-ai-coding-is-so-good-where-are-the-little-apps/
@glyph @agowa338 @tymwol @davidgerard Your post is very well framed. I think you can simplify the equation to C/PH because I, W, and E are each approximately zero, and as you point out it's tricky to estimate the other numbers. If we assume for a second well-framed problems have P close to 1.0 (I suspect this is true), the major variability is C/H which I expect to be highly variable dependent similar to how some areas are easier to mechanically test than others.
-
@glyph @agowa338 @tymwol @davidgerard Your post is very well framed. I think you can simplify the equation to C/PH because I, W, and E are each approximately zero, and as you point out it's tricky to estimate the other numbers. If we assume for a second well-framed problems have P close to 1.0 (I suspect this is true), the major variability is C/H which I expect to be highly variable dependent similar to how some areas are easier to mechanically test than others.
@glyph @agowa338 @tymwol @davidgerard What I am most concerned about are what in the post you categorize as harm to the humans and 2nd order effects. I think these get to be much more concerning as the underlying tools get more effective, and I'm not sure how to even mitigate those problems. If we look at how mass media and social media have melted people's brains a bit...
-
@glyph the story in four acts we are about to see (just one example, but could be Wikipedia, or anything else as well)
-
@agowa338 @glyph @tymwol @davidgerard Is there data on that? Acknowledging that this is anecdote, I see non-programmers in my extended work and social world trying to build little apps for whatever personal purpose that I don't think they would have attempted a couple years ago. I don't think most of them even know what StackOverflow is. (and their inspiration may be all about marketing rather than actual experience, so we will see how they fare)
-
@glyph I know they have a policy about retitling posts to avoid click bait in some cases, but it seems to be a failure of reading comprehension to take this as clickbait and a joke in the introductory paragraphs to be the ‘answer’. And then extremely bad faith to then take that non-answer as a dogwhistle.
-
@mirth @agowa338 @tymwol more broadly speaking, there are some strong indications that on balance it isn't delivering on its promise. A recent roundup from @davidgerard: https://pivot-to-ai.com/2025/09/09/if-ai-coding-is-so-good-where-are-the-little-apps/
@glyph @mirth @agowa338 @tymwol @davidgerard
Maybe the 10x coders are those that couldn't code before?
-
@glyph Ow, this "title butchering" is really concerning…
@krono it's not an *entirely* unfair attempt at summarization but if I wanted people to get mad at me for somebody else naming my articles I'd go work for a newspaper
-
@krono it's not an *entirely* unfair attempt at summarization but if I wanted people to get mad at me for somebody else naming my articles I'd go work for a newspaper
@glyph Fair.
-
@glyph I know they have a policy about retitling posts to avoid click bait in some cases, but it seems to be a failure of reading comprehension to take this as clickbait and a joke in the introductory paragraphs to be the ‘answer’. And then extremely bad faith to then take that non-answer as a dogwhistle.
@benjamineskola As I said in other replies on lobsters, I absolutely understand how everyone can be on a hair trigger about fascist shit these days so I am definitely not judging anyone for getting upset. They are for sure not _correct_ but also it is, a little bit, on me for not addressing this explicitly earlier.
-
@glyph I have a friend who spent the last year+ battling a kratom addiction. I get the analogy, and why you chose it, but using an addiction framing to talk about LLM usage is an analogy that risks really trivializing addiction (Is there a biological, genetic component to LLM usage? Is LLM usage linked to underlying psychological disorders like anxiety and depression?). I understand how you got here, and don’t think you’re wrong exactly, but I do wish you hadn’t made this argument.
@jacob @glyph couldn't there easily be an actual physical addiction to LLMs, same as with social media / "checking your phone" => dopamine releases => resistance => need to increased usage (cycle of addiction) ?
I would also guess from all the bad news that LLMs in relation to other areas in which it enters people's lives would produce similar cycles of addiction that also involves anxiety and depression.
-
@glyph got to love the folks going "the only right answer is one that assumes a 4k monitor" like 1080p monitors aren't still incredibly common devices.
-
@glyph if there is anything you would regret not doing, I recommend focusing on that
-
@glyph I feel so seen with the Pomodouroboros option, but also I voted for all of them so 🤷
Gli ultimi otto messaggi ricevuti dalla Federazione
-
Amazon preferisce dare la colpa ai propri ingegneri piuttosto che alla sua intelligenza artificiale
"La cosa più comune che fanno i computer è rompersi, ed essere sinceri e trasparenti su questa realtà senza far sembrare la propria piattaforma un mucchio di mattoni incoerenti che barcollano su un dirupo sopra un parco giochi è un delicato gioco di equilibri. L'affidabilità di AWS è leggendaria e, nelle rare occasioni in cui fallisce, camminano sul filo del rasoio e comunicano molto bene. Quindi sono rimasto sorpreso nello scoprire che tutto ciò che bisogna fare per spazzare via vent'anni di eccellenza e farli sembrare dei fanatici insicuri e spumeggianti è cospargerli di un po' di narrativa del tipo "forse AWS è pessima nell'intelligenza artificiale". Poi, perdono la testa."
https://www.theregister.com/2026/02/24/amazon_blame_human_not_ai/
-
@marcoboh @ju @matz @lysander @thatgiga ah si, il famoso italiano corretto dei fiorentini, «sta lí a dusgento metri».
-
@G3rt come potrei non scegliere le opzioni 1 e 3 oppure la 2 e la 4?!?
DIMMI TU COME? 😝
-
@fdroidorg no issues so far, thanks. The big question is how will the 148 update be handled. I don't want an “AI kill switch”, I want a browser without AI features in the first place. Will Fennec patch that shit out at the source level?
-
Current* conditions near Toledo, OH:
-
@glitchypixel I'm very interested to see where you go with this. I got really excited about VALA about ~15 years ago, and the unportability of glib to even things like wasm was one of the reasons I eventually fell off of it. I'd been hoping someone might come along and produce a middle ground between gobject and glib. It seems like that never shaked out, but patching over the deficits with a few choice libraries might actually work out really well.
-
@50Percent_Cooler LibreWolf Is a bit too opinionated for my taste
-
@manto ma è un pezzo di storia! Telefono salvavita per anziani, più anni '90 dei Crash Test Dummies!
