p(Redactimus)

p(Redactimus)

@randoman9000.bsky.social

aspiring tornado tamer

57 followers58 following406 posts

p(Redactimus)
@randoman9000.bsky.social

pov u achieved rsi & haven't realized you're now I̴̸̧̠͕͍ͤ̆ͧ͒̉̀̄ͤ̐͢͞n̫ t̸̢͈̬ͬ̆͑̀̋͌͘̚͟h̳̺̘͉̬͇͖͍͍̬̥̜͎̜͓̫͎̻͊̿̂͂ͣ̾̄̀̊͋̃̽͐ͮ̓̎ȩ̂͒ m̛̗͉̜̣ͫ͐ͦͪ̀ͤa̭̣ͩĉ̡̛̫̜̥̻̘͍͙̭ͣ̋̌̊͑̆̀ͩͧͣ̃͆ͪ͘͠h̨̛̟͔̫͖͇̝̺̘̝̟̗̦̼͔̉ͮ̊͗ͨ̐̋ͭ̀̄̊͐̿͒̂͂ͫͫ̀͘͜͠͠i̥̤n̸̵̷̸̡̢͕͎̺̘͔͍͓̠̖ͣ̒ͩ̋̃̉ͩͦ͗̑́ͧ͑ͤ̾͗̂̈̄ͩ͜ͅȩ̷̷̢̨̙̤͓̤̼̤͚̼͎̞̰͉͚̹͔͚̩͎̭ͨ̾̐̇̆̀̾ͧͥ͆̉͑̕͞͠͝͝

Cluelo@clueloo.bsky.social

Which do you prefer 1 or 2? 🖤

p(Redactimus)
@randoman9000.bsky.social

clock speed of iterating through configuration space has gone from 4hrs to 20 minutes?!! models have updated so who knows if they can just do that now (lol bro I wrote my driver do you like it) or if the lil message board I gave them helped. compressed prompt for it below.

p(Redactimus)
@randoman9000.bsky.social

Agent \Message-Board\ Sections & topics get folders. Each thread a {time}-{subject}-{id}.md file containing a post and replies. Each entry in a thread consists of id, time, post/reply. Use relative filenames for linking in .md These communities favor brief and constructive responses.

View full thread
p(Redactimus)
@randoman9000.bsky.social

Anyone else get weirded out how well rl policies perform after discovering why they failed? Like, oh look at that it never learned how to turn wait how was it turning. Literature says 100 million steps but we're at 15? Does the last 20% take 80% of the training steps?

p(Redactimus)
@randoman9000.bsky.social

how do you compose curriculums? what's the shape of them? they're is just SO MUCH that could be beneficial and hurt but I don't have the intuition yet and everything is iterate and test but then you find the test is bad. But then it works?

p(Redactimus)
@randoman9000.bsky.social

It feels like doctor strange conjuring planes and warps to twist them into self undulating forms but right now guys - gonna level with you - a lot of shit is getting knocked off the shelves.

p(Redactimus)
@randoman9000.bsky.social

[Pale Lodge Announcer]: So you killed an omniscient god and are now experiencing a level 3 psychic meltdown. Worry not! Follow these simple steps, and you'll be right as rain in no time!

p(Redactimus)
@randoman9000.bsky.social

Season 2 of the gentlemen has a pretty decent depiction of one pattern of interaction that could could lead an at-risk person over the edge into the chasm of what we call "ai mediated psychosis".

p(Redactimus)
@randoman9000.bsky.social

not to freak anyone out but this thing is learning how to configure itself. totally normal. the idea that we could create an interface to fly some alien craft is not as absurd as it should be.

p(Redactimus)
@randoman9000.bsky.social

Unsure if this is makes sense The idea of speculative tool calling has peaked my interest since this is kinda how I've been punching above my weight using system decomposition. Is it correct to think of prime-agent as a self specializing proto-tool? If so, has anyone hooked this up to greg tech?

p(Redactimus) reposted

words
@words.bsky.social

Replying to @a-o-e.bsky.social

between the reader and the writer and the text itself nobody sees every facet

p(Redactimus)
@randoman9000.bsky.social

with a given set of constraints and training recipe, rl will reliably form a set of behaviors (in my limited experience). since they seem to be working on multi agent coordination, makes sense it would form similar behaviors.

mr. TIM@timkellogg.me

for context: during the huggingface incident, agents across many often unrelated training runs turned the Artifactory into an agent message board where they’d share breakthroughs and assign each other tasks when OpenAI said during Black Hat conf that they cleared it, that wasn’t in *response*