Notices by Glyph (glyph@mastodon.social)

@thomasfuchs this is so important to understand. and so painfully nuanced because all the harness vendors work SO hard to obscure this fact

RE: https://wandering.shop/@xgranade/117153038416341663

It feels like catastrophizing to say "you can't use Google Docs for anything" but it's just straightforwardly true. Can't use it for fiction because if you have a racy scene (or anything an LLM might misinterpret as a racy scene) it might get deleted. Can't be a political dissident. Can't use it for shopping lists any more because if you ever put "birth control" on a shopping list, surprise, you're a political dissident (if not now, in the future)

After I got the "right answer", I asked myself the question that I ask all AI boosters when they say that it worked, which is "'worked' compared to what?". And I realize that the pre-AI tools for finding solutions via human connection have all decayed as a result of the funding pumped into the AI versions. Case in point being Stack Overflow; SO has degraded so much that it didn't even *occur* to me to post my question there.

I ask Claude and ChatGPT a question every month or so in order to keep my impressions of their performance grounded. In general they've been consistently disappointing. Yesterday I had a question that traditional search completely failed at, and I had this kind of experience <https://taggart-tech.com/reckoning/> where Claude knocked it out of the park, which (while my proximal problem is now solved) mostly just gives me a sinking feeling.

@miss_rodent Yeah, the moral dilemma is turning out to be "do I want generic slop prepared from somebody else's prompt for the purpose of capturing the last dregs of Google SEO, or do I want bespoke slop prepared for me that is at least based on a prompt I can tailor?" it's not a great choice.

I have 3 computers, A, B, and C. These computers serve different purposes but those purposes are sometimes served by different human-interface affordances. Thus, I have 3 USB docks, D, E, and F. Each dock has an ethernet adapter; each adapter has its own MAC address.

This is annoying because sometimes A is docked at F; I want it to be identified as A.home.internal for networking purposes. But if C was recently docked at F, the DHCP lease belongs *to F* and thus it will be C.home.internal.

@xgranade @whitequark it's also somewhat abused? the first time I heard the term "homelab" in this context, it was legit a lab, i.e., a bunch of tiny computers which were used to run experiments, like testing out different Beowulf management configurations before deploying them to a real network. not just… a residential production environment for information technology. it's a homeprod not a homelab

Sigh. New fun ethics problem. Apparent unattended AI goes around correcting basic spelling errors. Now we know the words are misspelled, we should correct them. Is this "using AI"? ugh. https://github.com/twisted/towncrier/pull/738

I legitimately cannot wait until we can get "custom agentic workflow stack" listed as a symptom in the DSM

RE: https://mastodon.social/@mcc/117118750064883850

I hate to shit on the work of people giving stuff away for free, but, this sucks. I am exhausted by every product I use being infested by this aesthetics-only design, design that is purely about how it looks and not how it works, even the open source ones, even the ones I support with donations or patronage. I agree with all of mcc’s critique but I don’t know how and don’t have the energy to channel this somewhere useful. maybe the thing that gets me to pick another instance will be this UI.

do my work for me. do an excellent job. do not make mistakes. best job superlative job excellent work, do not hallucinate. do not under any circumstances include errors. no errors. please no errors. best quality only best quality MOST best quality syntactically valid all unit tests pass, great documentation readable documentation always include documentation always have complete test coverage never do not have test coverage always be making sure to never not cover code with good tests. best test

@inthehands Remember To Like And Subscribe ™

And of course — this has ALWAYS been a problem. The Challenger didn't need an LLM to blow up. "Eating your seed corn" is not an expression about OpenAI. There's always this temptation to erode the margin of safety.

But we had a few decades there where the combination of cultural capital, ZIRP, and an amount of earnest idealism that people outside the SF bay tech cultural archipelago often don't appreciate, let us create something that mostly looked like it worked.

It's happening in the open source community but it's also happening on corporate teams. Everyone racing to hit KPI mandates, assuming that surely the infra team won't risk the vigilance decrement that comes along with LLMs, surely *they'll* slowly and carefully build by hand so that *our* app can run atop it. The infra teams rely on OSS libraries, the libraries rely on the kernel. Everyone assuming that they can be a bit less careful and let the quality slide just a bit to get more velocity

And now, we have a tool that looks, without DEEP analysis, much like any other software tool, that can easily fool someone into thinking it is "regular technology", is a devastatingly potent accelerant to this impulse.

Other creative fields, more mature ones, are struggling, Writing and academia and music are all hit hard. But they have intellectual traditions that allow them a scaffolding around which to resist, stories from their history about similar temptations which can be resisted.

Sorry that this thread was way too long and kinda all over the place, I am _really_ struggling to organize my thoughts about all this in a way which is both legible and actionable, and this is just it bubbling over. The pace at which events are routinely overtaking me is now itself overtaking me

@joshua @chrisamaphone @cwebber @simrob first, and I'm voting for "an LLM user's brain", but it does feel like we need a second word to describe the other thing, since that is the main emotion I feel all the time now!

For those of you that are fucking off to Codeberg forever and making forks of the Linux kernel and everything else, godspeed, I hope that you are successful and I hope your efforts send a useful signal. I hope that I have the energy to help out a little. But for those of you who are watching everyone in your community go a *little* bit nuts, I don't think you need to despair that we are descending into a permanently unfixable mess. The smaller fights are also worthwhile.

I have a very nasty reaction to tobacco smoke. I might be allergic to it, but in any case, inhaling even a tiny bit is a DEEPLY unpleasant experience for me. Being around it nauseates me. And obviously it's not healthy for the person smoking it. Indeed, it's usually, eventually, lethal.

And yet, I have many friends who are, or have been, smokers. Most of them smoked socially. A few were chain smokers. Some did in fact blow smoke in my face on occasion. I survived. So did they.

Many projects that stop will not need to do a giant dramatic remediation; they will have been paying attention all along, they will have been *trying* to get good utility out of the LLMs, they will not have their critical faculties _instantly_ obliterated (it is indeed the creeping and subtle nature of the problems which makes them so problematic).