Discussion
Loading...

Discussion

  • About
  • Code of conduct
  • Privacy
  • Users
  • Instances
  • About Bonfire
Charlie Stross
@cstross@wandering.shop  ·  activity timestamp 3 weeks ago

"The red recording light promised truth; the coffee beside it had already stamped it with a brown ring on the console. ... I adjusted the pop filter, as if I wanted to politely count the German language's teeth."

GPT-5 is optimized to produce garbage that is evaluated positively by other LLMs.
https://futurism.com/gpt-5-literary-outputs

  • Copy link
  • Flag this post
  • Block
joriki
@joriki@infosec.exchange replied  ·  activity timestamp 3 weeks ago
@cstross

AI: "Social Text affair? hold my beer"

  • Copy link
  • Flag this comment
  • Block
Dec.tar.gz
@dec23k@mastodon.ie replied  ·  activity timestamp 3 weeks ago

FFS, Futurism:

"We and our 1257 technology partners ask you to consent to the use of cookies to store and access personal data on your device."

REJECT!

  • Copy link
  • Flag this comment
  • Block
wolfstettler
@wolfstettler@troet.cafe replied  ·  activity timestamp 3 weeks ago
@cstross This reminds me of the Berlin, Miami novel/experiment by Hannes Bajohr, where training a LLM on 3 LLM generated novels led to complete gibberish in a 4th generated novel. The same LLM was trained on novels written by humans for generating the first three novels. Those 3 novels were absolutely readable (but not exactly Nobel Prize material).

https://strauhof.ch/ausstellungen/maschinenpoesie/
https://de.m.wikipedia.org/wiki/Hannes_Bajohr

  • Copy link
  • Flag this comment
  • Block
Sam Livingston-Gray
@geeksam@ruby.social replied  ·  activity timestamp 3 weeks ago
@cstross you say "model collapse," I say "LLM prion disease"... actually, Bob would probably shorthand it as "L Syndrome" 😜
  • Copy link
  • Flag this comment
  • Block
Mz. April Daniels
@MzAprilDaniels@chaosfem.tw replied  ·  activity timestamp 3 weeks ago
@geeksam @cstross Adversarial data spiking, is my bet. Altman realizes he can't compete on the merits (DeepSeek ate his lunch and he didn't see it coming) so he's gonna try and poison the wells that other models draw from, hoping that his own stockpiles of untampered information are enough to give him a lead in an adversarial data harvesting environment going forward.
  • Copy link
  • Flag this comment
  • Block
Mz. April Daniels
@MzAprilDaniels@chaosfem.tw replied  ·  activity timestamp 3 weeks ago
@geeksam @cstross When I was in high school, we had a beautiful moment where it seemed like the sum of human knowledge was going to be freely available to everyone going forward.

Rich men have decided this should not happen, that high quality data should always be expensive and scarce. But how to accomplish this? The commons cannot be enclosed in a digital space--so they are salting them.

  • Copy link
  • Flag this comment
  • Block
Mz. April Daniels
@MzAprilDaniels@chaosfem.tw replied  ·  activity timestamp 3 weeks ago
@geeksam @cstross This is what the marijuana tells me after reading the headline and the tweets appended.

Now, to actually read the article...

  • Copy link
  • Flag this comment
  • Block
Jacek Wesołowski
@jzillw@mastodon.gamedev.place replied  ·  activity timestamp 3 weeks ago
@cstross I saw your post and read the quote before realizing it was a genAI sample, and for a moment I thought that either my reading comprehension or my English fluency was gone.
  • Copy link
  • Flag this comment
  • Block
Lazarou Monkey Terror 🚀💙🌈
@Lazarou@mastodon.social replied  ·  activity timestamp 3 weeks ago
@cstross I love how that website has NOT bought into the AI hype

#AI#GenAI#LLMs#TechBros

READ THIS NEXT
DOC. ON HOLIDAY
A Single Typo in Your Medical
Records Can Make Your Al
Doctor Go Dangerously
Haywire

TROUBLING
First Al Psychosis Case Ends in
Murder-Suicide

OPEN BOOK, CLOSING DOORS
At OpenAl, Signs of Crisis
Grow Behind the Scenes
READ THIS NEXT DOC. ON HOLIDAY A Single Typo in Your Medical Records Can Make Your Al Doctor Go Dangerously Haywire TROUBLING First Al Psychosis Case Ends in Murder-Suicide OPEN BOOK, CLOSING DOORS At OpenAl, Signs of Crisis Grow Behind the Scenes
READ THIS NEXT DOC. ON HOLIDAY A Single Typo in Your Medical Records Can Make Your Al Doctor Go Dangerously Haywire TROUBLING First Al Psychosis Case Ends in Murder-Suicide OPEN BOOK, CLOSING DOORS At OpenAl, Signs of Crisis Grow Behind the Scenes
  • Copy link
  • Flag this comment
  • Block
Sepia Fan
@stekopf@mstdn.social replied  ·  activity timestamp 3 weeks ago
@cstross

To all non-native English speakers: The cited sentence is gibberish. There is nothing you don't understand or miss out, no reason to search the web for a proper phrase translation. No need to despair or doubt your own abilities. It's just AI nonsense.

  • Copy link
  • Flag this comment
  • Block
Charlie Stross
@cstross@wandering.shop replied  ·  activity timestamp 3 weeks ago
@stekopf

Exactly! "The policeman's beard is half-constructed."

(Only the 1980s chatbot RACTER—which delivered tha above—could run in 256Kb of RAM on a PC XT without burning forests and turning tidal estuaries into steam.)

  • Copy link
  • Flag this comment
  • Block
Stephen Bannasch (316 ppm)
@stepheneb@ruby.social replied  ·  activity timestamp 3 weeks ago
@cstross @stekopf

A long time friend many years ago told me about RACTER and he said then it was good at fooling folks it was human because it was coded to be a bit of an asshole and not listen well.

I never interacted with RACTER and always wondered if this was true. And because I was spelling RACTER wrong I could never find any mention of it!

  • Copy link
  • Flag this comment
  • Block
Sepia Fan
@stekopf@mstdn.social replied  ·  activity timestamp 3 weeks ago
@cstross

I was wondering: Oh, did I miss some meanings of "stamed" and went curious about the meaning of "teeth counting" (English can be weird sometimes).

Then I decided to read the article 🫠

Isn't AI used for language training already? 🙈

  • Copy link
  • Flag this comment
  • Block
Mark Mason
@markmason@mas.to replied  ·  activity timestamp 3 weeks ago
@cstross I think it’s like a stage impressionist doing an impression of Sigmund Freud. A cod German accent and a few key phrases and the audience knows who it is supposed to be. It’s not about making the psychology correct, it’s just about making it sound right. In the same way ChatGPT is making something that sounds like what you asked for *for an audience that doesn’t understand the meaning*. So by design it is writing for other LLMs (or people that don’t have enough knowledge of the domain)
  • Copy link
  • Flag this comment
  • Block
Electropict
@electropict@mastodon.scot replied  ·  activity timestamp 3 weeks ago
@cstross

It's beginning to read rather like a coded message asking for help, from the kids imprisoned inside the machine. The big man promised truth and food ... but it came in tubes. He was so close, putting them in, I could count his teeth.

  • Copy link
  • Flag this comment
  • Block
Gillinger
@Gillinger@mas.to replied  ·  activity timestamp 3 weeks ago
@cstross "Garbage in, garbage out." That law will never change no matter how much money you throw at it.
  • Copy link
  • Flag this comment
  • Block
marmarta
@marmarta@chaos.social replied  ·  activity timestamp 3 weeks ago
@cstross I have not expected robot masturbation to look like that, and yet here we are.
  • Copy link
  • Flag this comment
  • Block
Charlie Stross
@cstross@wandering.shop replied  ·  activity timestamp 3 weeks ago

I am beginning to suspect that OpenAI have replaced all their human software managers with LLMs, otherwise this transparent garbage would never have been green-lit for public exposure.

  • Copy link
  • Flag this comment
  • Block
Legit_Spaghetti
@Legit_Spaghetti@mastodo.neoliber.al replied  ·  activity timestamp 3 weeks ago
@cstross My suspicion is that by this point a not insignificant portion of C-suite personnel are in full-blown LLM psychosis, and not just in the tech industry. Those things eat through brains like battery acid.
  • Copy link
  • Flag this comment
  • Block
Fleur Kelpin
@fdlk@mastodon.social replied  ·  activity timestamp 3 weeks ago
@cstross well, it can now simulate human stupidity more eloquently. It used to be a silly stubborn stupid graduate. Now it is a silly stubborn stupid PhD.
  • Copy link
  • Flag this comment
  • Block
Uilebheist
@Uilebheist@polyglot.city replied  ·  activity timestamp 3 weeks ago
@cstross The public doesn't count.
What counts is getting investors to give money, never give them any return on investment.
What looks like transparent garbage to you might look like gold to one of these people with more money than sense.
  • Copy link
  • Flag this comment
  • Block
Log in

bonfire.cafe

A space for Bonfire maintainers and contributors to communicate

bonfire.cafe: About · Code of conduct · Privacy · Users · Instances
Bonfire social · 1.0.0-rc.2.21 no JS en
Automatic federation enabled
  • Explore
  • About
  • Members
  • Code of Conduct
Home
Login