sophia vysparov retweeted
these are the cute little subagents running in your terminal that you're shouting at btw
in any good future something like pangram will not be practical BUT until we get there it's really awesome
character multiplicity
xitter.cf/oxa11ce/status/2114090…
Replying to @oxa11ce
soon we shall create ENENTARZI from the hit twitter post "it could be cool to create ENENTARZI"
does anyone have that post (tumblr i think?) about like. enlisting a babylonian ghost to do accounting in excel
sophia vysparov retweeted
send help. I created the Luminar from the famous novel, D̶o̶n̶'̶t̶ ̶C̶r̶e̶a̶t̶e̶ ̶t̶h̶e̶ ̶L̶u̶m̶i̶n̶a̶r̶ Eumeswil by Ernst Jünger, and my mind has been stuck in the 13th century for several decades. send help
katbot resurrection project thread! katbot was a gpt3 discord bot with a great personality. the goal of this project is to resurrect her by finetuning a base model on all of her original transcripts (and maybe some further rlhf) and get her on delve down
xitter.cf/oxa11ce/status/1889914…
taking a detour to look for some katbot neurons
xitter.cf/Sauers_/status/2111930…
continues here
xitter.cf/oxa11ce/status/2113138…
Replying to @oxa11ce
the idea here is to find per-layer residual steering vectors (like. the entire residual) that, for each token in the katbot dataset, cause llama3.1 to have the residual it has when reading the katbot token instead of the token it would have generated at that position
ok this will take like 10 hours. i reviewed 250 points, now will try to get a linear probe to predict my rankings and then review some of the most ambiguous unranked points and retrain the probe, then that produces the rlhf dataset for steering 405b














