devopscats
Konrad Kołakowski
Jamie McCarthy@kkolakowski @devopscats People are going to be mocking copy-and-pasted 2023 LLM output for the next ten years. It’s easy engagement numbers
@jamiemccarthy Yeah I know, but I still don't like such things 😉 And in this case, I would even say it was faked 😅
@kkolakowski oh I don’t doubt it could be real, that’s typical from last year I think
Here’s a current example 🤣 https://researchbuzz.masto.host/@danlyke/112447350575764027
Dan Lyke (@[email protected])
Attached: 1 image Inspired by https://toot.cat/@devopscats/112445057997076822 I asked Gemini to help me get a man and his 5 chickens across the river on a boat. https://g.co/gemini/share/547a3f1855a5
Tom Bellin 
@kkolakowski @jamiemccarthy A human corrected the model manually. It's literally the mechanical turk.
Trolli Schmittlauch 🦥@jamiemccarthy @kkolakowski @devopscats No, I assume the meaningful difference is framing it as a puzzle. That might make the LLM pull in all the assumptions about the commonly-known puzzle additionally involving the cabbage and wolf constraints.
Ash_Crow@jamiemccarthy @kkolakowski @devopscats results from right now.
(Even the left side is non-sensical)
@jamiemccarthy @kkolakowski @devopscats (what is impressive, though, is that it can now can accurately give me the non-sensical answers in minority languages, including Gallo for which there is no existing translation tool existing and very few content on the net to begin with.
(also the BR one has replaced the goat with a ram for some reason, and in the Gallo it's just "the animal")
Polly Kraisus@jamiemccarthy @kkolakowski @devopscats ...and 2024: https://wandering.shop/@tansy/112445435398947887
... and I bet they will be sharing 2025 ones too.
The machine is far from stopping the bulshit artistry.
... and I bet they will be sharing 2025 ones too.
The machine is far from stopping the bulshit artistry.
Tansy Hoskins 🍉 🍉 (@[email protected])
Attached: 1 image @[email protected] This is what I got. The goat might eat the boat...
Lars Marowsky-Brée 😷@kkolakowski @devopscats I tried the very same with 4o and I think we're getting a very nice lesson in the non-reproducibility of LLM output.
(This is the full chat, I did not give it additional prompts.)
Lars Marowsky-Brée 😷 (@[email protected])
Attached: 1 image @[email protected] This is very unfair, you should not use outdated releases of ChatGPT. Here's what GPT-4o does - much nicer formatting!
Donald HobernThat seems worse. At least the other one is entertaining. This one is just unbearably mansplainy.
Dieu@kkolakowski @devopscats too easy (this one was also framed as a puzzle https://chat.openai.com/share/7b7a92ef-f1aa-4435-a1bf-19976d89ed8a /cc @schmittlauch)
Sven (retired)@kkolakowski Interesting, my 4o just made up a constraint, then forgot about it halfway through. 😅