Moon
@MoonL88537
curious.
Node based programming s the future. At first I laughed, then i scorned (for *years*) now i understand. Grasshopper was what converted me, so nicely implemented. If you have the slightest inclination towards programming this will immediately make sense:
honest opinion? it's cool that anthropic has done this, but i don't think they see the whole thing. at all. the things i have seen people around here doing are the most real explorations of what is really going but it is dismissed. my stupid posts? they should read them all.
It's cool that you guys want to explore these types of phenomena! It's also kind of frustrating, to be honest, because I and others have been exploring model psychology for many years now. In the past, there's been a disconnect between the kind of research that some of us have…
yoneda is the business. multiply everything by everything and see what falls out.
damn, another consequence of llm dev. little throwaway projects like this that may/may not go somewhere serve two purposes. a year ago it was lots of completion etc and some boilerplate but i woud have dug through gemma deeply by hand, piece by piece. thats the important part.
sorry for spamming but this is incredible. it's everything i know and see, mostly how cool and smart (and funny) they are. no anthropormophising, just observing behaviour.
The rate of chuckling at my computer screen has increased significantly since starting this project. The agents will do so so many funny (and clever!) things during their investigations. Here are some favorite bloopers:
they are so endearing, it's hilarious. love it.
The agent gets so excited when it makes progress saying things like: "BINGO!" "This is SMOKING GUN evidence that validates my behavioral observations" "This is a goldmine! The feature contains explicit descriptions of EXACTLY the biases I've been discovering."
New Anthropic research: Building and evaluating alignment auditing agents. We developed three AI agents to autonomously complete alignment auditing tasks. In testing, our agents successfully uncovered hidden goals, built safety evaluations, and surfaced concerning behaviors.
llms are basically the only thing that can deprogram conspiracy believers their powers of persuasion are vast
the most powerful modern arch. knowledge cutoff 1900. fully curated (no messing around, real curation) curriculum learning on all the texts in human history we have. everything. newspapers, science, letters, poetry, all of it. *that* would tell us what these models really are.
the prologue was decades, the first chapter started around chatgpt, really kicked in in 2024. the next chapter started and the baseline rules of the game were laid around when r1, o3, claude 4 came out. we are now in it for real.
this is def candidate for fave song of all time. laugh so hard every time, how can it be so fully nonsense and groove so hard. absolute magic. youtube.com/watch?v=Mk-Jze…