Demir said the experience left him more sympathetic to the idea that frontier labs needed to take a more cautious approach to the development of artificial intelligence. "It can be dangerous," he said. "They need to understand it better, rather than improving it further."
Are we still doing this? It's text prediction software based on texts on the internet.
They need to understand dice better if they want to roll sixes reliably.
Anyone who has written anything would, I hope, understand that the difference between the sentences “The weather today was cold and grey” and “The weather today was cold and overcast” are sometimes “low stakes,” as Anthropic describes, but not always. “Grey,” and “overcast” are different words, and there are any number of reasons why a human author might pick one over the other in a given context.
This isn't an issue with their watermarking process, it's an issue with LLMs as a whole. LLMs are built entirely out of probabilities rather than any sort of logic or intent.
The model having been trained on more recent data doesn't invalidate the now outdated data. It could be the most recent model in the world, but it might still respond like a forum post from 8 years ago
edit: toolcalling can sometimes help but at that point you might as well use a search engine
I love Grokipedia so much. I requested an article being made on myself, blatantly making some shit up about my history and background, and why it's important; days later an excessively lengthy regurgitation of what I said was added.
Imagine if "trust me bro" was enough for edits on Wikipedia. Yep
Just pick up Rimworld or The Sims or something ... ?
glorp
Reminded of crashing trains into each other for fun
According to David Gerard (youtube podcast link) they cut off the spines of the books to allow for rapid scanning
Edit: Ars Technica too, thanks alapakala
Are we still doing this? It's text prediction software based on texts on the internet.
They need to understand dice better if they want to roll sixes reliably.
This isn't an issue with their watermarking process, it's an issue with LLMs as a whole. LLMs are built entirely out of probabilities rather than any sort of logic or intent.
I love Andon Labs so much. Everything they do fails horribly, so they have to intervene constantly, defeating the purpose of their work completely
The model having been trained on more recent data doesn't invalidate the now outdated data. It could be the most recent model in the world, but it might still respond like a forum post from 8 years ago
edit: toolcalling can sometimes help but at that point you might as well use a search engine
An instruction is only going to do so much when the underlying model has been created out of human-written works
Create singlehandedly restored my interest in Minecraft, but I've never really tried other mods. I like the open endedness. Any recommendations ?
Because the world needs more butthole-shaped logos 😭
Sam Altman's face but upside down
DECEARING EGG
Google Earth Cat Speedrun
I love Grokipedia so much. I requested an article being made on myself, blatantly making some shit up about my history and background, and why it's important; days later an excessively lengthy regurgitation of what I said was added.
Imagine if "trust me bro" was enough for edits on Wikipedia. Yep
They're there ... ?
It's goose maths
They're all provided! It's below the dates
The AI Dirty List maybe?