Hard_Space

Previous discussion - https://news.ycombinator.com/item?id=43678144

I understand the anger. But I wish I were better able to resist fixing the world with code in this way, as I really am supposed to be working.

Yes, LLMs certainly learned these tropes somewhere. I'm just surprised that different models and different pre-processing methods arrived at the same dominant cliches.

I don't know - the 'great enshittification' of Substack seems an inevitable event, for which reason I have stayed away from using it as a platform, except for a few experiments. It was initially populated by refugees from Medium, itself enshittified, and there seems no reason that Substack refugees won't eventually become a thing for the same reason. Reminds me of that Twilight Zone (or was it Night Gallery..?) episode with the guy who keeps moving from one sinking ship to another (Titanic, Lusitania, etc.).

Yes, it was polluted with multiple 'Not just [x] but [y]' examples - the most obvious tell of AI-generated text, among other aggravating examples. It was hard to read, but I was interested enough in the subject to slog through it.

Agreed. I always hated the 'two part', 'payoff'-based drama of titles like these, even before the LLM era. If it was lazy before (it was), it now comes off as 'one-click' lazy. Sadly, The Guardian has become infested with this style lately.

This is interesting. Until autumn of 2024, when the company was subsumed into a better-heeled AI-VFX concern, I worked for probably the best-known and earliest all-AI VFX house, whose ethics department was headed by a philosopher, though the company struggled to place him to practical advantage.

The only comment I can make on the general trend is that it's apparently good PR for cash-saturated startups. Ultimately what AI 'means' is certainly not the business of those making it (who are arguably least-qualified to comment); and insider insights offer no benefit that I can discern.

Reminds me of 1990s multi-CD Mac games.

This is a chronic LLM style, particularly the multipart dramatic flourishes and the Haiku paragraphs. But at this stage, there's a good chance that LLM re-diffusion of this style is feeding back into human culture. God save us.

Well, neither am I - I am talking about my own peer group, non-tech types just fine with computers.

60 year people can't user your fancy site because then don't have an internal model of how a computer works.

I think this is a bit outdated. I'll be 60 in a month, and have been practicing and writing about machine learning, for money, for a straight 10 years now; and I was a young man (and a full stack developer) during the digital revolution.

If anything, GenX had to work harder to get into these brittle emerging technologies and paradigms. There's no-one of my age group, at least that I know of, who is remotely as tech-illiterate as your comment depicts.

Truth is that it took so long for smartphones to dumb down everyone's tech acumen that those of my generation had already learned to do it the hard way.

But these figures measure exposure, not integration.

Ach, it's probably a mostly human-generated piece, but any time I see the 'Not [x] but [y]' formulation, I tune out.

Nothing to see here. The title suggests some hidden trick or tech wizardry, but the two people referenced were just...well, dumb: using Gmail in the first case, and retaining his icon of his own real face for an avatar in the second.

I used to see AI generated images with lots of unintelligible writing or misspelled words in slides, but the speaker left them in anyway. “Good enough” is not customer obsession.

This enforced adoption of immature GenAI reminds me of Milo Minderbinder trying to make people eat cotton in Catch 22, because he had inadvertently obtained a huge amount of it.

I need Photoshop for an almost vanishingly small subset of all the things it is capable of, and this holds true for nearly all 'full-fledged' software that I use. So what may not be surfacing, in the absence of vibe-coded Photoshops, is the growing local script collections of many users.

Since I have had AI to knock up Python scripts and workflows incorporating local ImageMagick and FFMPEG, I have devolved a lot of tedious Photoshop work to scripts and routines of some kind. Likewise with text and data manipulation that I might have turned to software for before.

I don't have the slightest urge to incorporate this ad hoc collection of scripts into a central program, and I certainly don't intend to share them in any way, considering the growing hostility to sharing vibe-code.

So this particular iceberg may be 99% underwater.

Sorry for the dumb question, but why is there an archive.today link for this? Is BBC news now paywalled in some regions or under some conditions? I'm guessing that the Beeb can handle the traffic, and that the site is not hugged to death.

In the UK in the 1980s (and with more difficulty in the 90s) I would phone the theater and ask when the film actually starts, even though it was almost always 20 minutes after the advertised programme time. Now there are no humans to ask, and my wife wants to see the entire programme anyway.

I use subdomains on an OVH VPS, since I want to access the services outside the network, so I can use freshrss.mydomain.com. But anything that can rationalize port number sprawl is welcome.

Not so for this port, which, besides not remembering sessions, crashed very early, and made me return to BBEdit.

Not really understanding the negative trend of comments. As someone who accesses multiple Windows machines on a LAN via a MacBook Air, I'm glad to have as many common GUIs as possible. I found it a bit hard to get used to BBEdit when I started using a Mac again, and have been a Notepad++ lover for many years. So, thanks to the dev for this.

Drove a no-tech tractor working on a farm in Tuscany in the early/mid-90s. Best driving experience ever.

In fairness, real-world researchers are expert at selective emphasis too.

Yes - the two-part title tells it all. I wonder of the likes of The Guardian, who are addicted to this style, will be forced to abandon it to be less associated with LLM-produced text.

A spinal injury three years ago made me above-averagely aware of ergonomics, as I created a dedicated office (chairs, floating supports for monitors, keyboards, etc.) that would support the injury.

I have to say that I now access the 8 or so machines across my LAN and VPS via a Macbook Air rested on my lap, either through RDP or SSH. I probably use this 80% of the time. Because the machine is on my lap, held up at about 30 degrees on my lap, I can't say I ever noticed the sharp edges, until this post!

Wow, I’m surprised. Of all my self-hosting solutions, this needs least maintenance, for me. Recently had to move to a fork of SyncTrazor, because a new project picked up support, but that’s the first time I had to think about it in years. Wish NextCloud and Immich were that easy!

Also see https://arxiv.org/pdf/2604.00025 ('Brevity Constraints Reverse Performance Hierarchies in Language Models' March 2026)

Incredible that we are regressing back to webrings and hand-curated lists like this, both of which I remember well. That's not a criticism! I guess that the quality-drop in search wasn't quite enough to make it happen, but the advent of AI content predomination will be.

If you turn on showdead, you'll see OP's explanatory post.

I am not sure why the title got truncated from 'I Created My First AI-assisted Pull Request and I Feel Like a Fraud', which conveys a very different impression, and makes more sense when browsing the comments here.

How does this improve on Floccus?