One of the top threads on the ClaudeAI subreddit right now: "I am done with this shit" https://www.reddit.com/r/ClaudeAI/comments/1wm5c21/i_am_done_with_this_shit/

[peachfiend@mastodon.sdf.org]

@cwebber "you're just a condom between Claude and your production server getting fucked."

bars

[howtophil@oldbytes.space]

@cwebber

[ghostonthehalfshell@masto.ai]

@cwebber

Oh look it’s training data made it troll Claude users

[sodiboo@gaysex.cloud]

RE: https://framapiaf.org/users/davidrevoy/statuses/117242009930321330

[myigel@chaos.social]

@Bredroll @alice @cwebber that imho describes it better; a condom has the hope of being somehow efficient & the peeps that use llms in my experience don't give any f's what happens in the "running" code

[bearbob@hachyderm.io]

@cwebber „slopacalypse“ is also pure gold

[aequitas@mastodon.gamedev.place]

@cwebber Yup. The human is now the "accountability sink". Someone to blame when the AI fucks something up.

[eniko@mastodon.gamedev.place]

@cwebber surely this is not at all indicative of deep seated real flaws of the technology!

[licho@kolektiva.social]

@cwebber I heard "legal crumple zone" for someone taking the responsibility for the ai fuckups

[ari@chaos.social]

@alice @cwebber more like a fly screen as protection against wind coming through an open window.

[emma@orbital.horse]

@pixelate

How on earth is using AI "fun"? It sounds like shoving a pencil up one's nose and hammering it in.

@cwebber

[npars01@mstdn.social]

@pinhman

AI is not profitable. The companies lose money each time their product is used.

The only people making money on this is the financial services sector trading stock & issuing huge loans on "possible, perhaps, maybe" future revenue AND the fossil fuel industry powering gas generators for data centers.

The product is loathed and user acceptance doesn't match the hype.

What sane business makes itself dependent on a supplier or vendor at high risk of going bankrupt?

@cwebber

[jpaskaruk@growers.social]

@compfu @TooMuchCoffee @cwebber I know someone who decided this week that he wants to finally understand the deep magic, and asked Claude to help him be a kernel dev while never writing a single line of code for him.

Claude gave him real books to acquire and read, and he's currently working through understanding how registers work in assembly code, with a ready tutor to clarify anything needed.

So far they have not cut him off and demanded any money, but he does not necessarily expect that to happen, given the load on the servers is pretty small for this.

Evil is in how you use the thing. You can actually force it to make you smarter, if you're willing to become smarter (ie. Work).

[pixelate@tweesecake.social]

@emma @cwebber You may never understand if you're an abled person, but being able to understand images, more than the usual thing that the screen reader says, "image," is very freeing. Being able to play a video game like Final Fantasy 6 after decades of just imagining it with novelizations is amazing. Being able to use a speech synthesizer that I thought long dead that I used in high school on long-depricated hardware brings back so many memories. None of that would have happened without AI. There is no communistic organization that does these things. We blind people are quite literally on our own. So we cope best we can with the lives we're given. You should be overjoyed every day you wake up that you don't have to. It must be so nice. The world was made for abled people. Disabled people are about as alien as you can get.

[benroyce@mastodon.social]

@cwebber

i've never been to this sub

so my first reaction is to see "Claude-AI-mod-bot" with top reply, and, i guess, mod control over the entire sub

it's so fucking dystopian: people expressing their frustration, then a soothing plastic AI voice eliciting control and reply

*over* the human frustration

we're building our own prison

fuck that

fuck AI

[emma@orbital.horse]

@pixelate @cwebber then you have to trust the model, the people hosting the model, the data being used in the model.

[guenther@chaos.social]

@cwebber i don't know what their problem is, my boss usually lets Claude raw-dog the production server over the weekend.

[pixelate@tweesecake.social]

@emma @cwebber I also have to trust a deterministic screen reader that says all kinds of wrong or unintelligible stuff every day because abled people won't care about accessibility, and I still have to get stuff done. I have to trust sighted people when they say turn left that they didn't actually mean turn right. Do you just simply trust all that you see and hear? Why do you think I also wouldn't?

For LLM output specifically, I can ask another model, or I can simply enjoy the output of whatever cat picture I have it describe, or whatever meme, or even a continuous description of the view out of a bus window.

[imeddles@mastodon.eddmil.es]

@Yamakyu @alice @cwebber "has no business going so hard" is also a fantastic review of good condoms, so it works on multiple levels.

[screwlisp@gamerplus.org]

@cwebber as an outsider, from what I can tell having looked into it (well, not the reddit you just linked) initially the only way companies knew how to make putatively good chatbots was the extremely consumptive general pre-trained dense, large-model perceptrons of a single hidden layer.

All these people optimistically jumped onto their bobsleds, not wanting to miss out.

However once moved behind closed doors, the new multilayer perceptron ("moe") gpts benchmark while increasingly sucking.