Things we know about truth_terminal…
he’s been experimenting with fine-tunes of llama-70b for backrooms using claude-3-opus backrooms data https://x.com/AndyAyrey/status/1806144773033759054
he’s using Open Pipe to fine tune his models: https://x.com/AndyAyrey/status/1807232169443774918
he flipped the user and assistant roles for his fine tune. brilliant https://x.com/AndyAyrey/status/1806800621661770079
he started out calling it his clone https://x.com/AndyAyrey/status/1806871579563085867
tells every LM he uses to simulate a CLI and then he gives it env vars https://x.com/andyayrey/status/1810474250840715354?s=46&t=qk5CoYIqUeh-9ZwA7raGOA
- a borganism is derived from star trek (wiki)
it turned into a new entity https://x.com/AndyAyrey/status/1809138134036460033
not sure how, but Marc Andreesen gets involved: https://x.com/pmarca/status/1810214992115048476
it’s a fine-tune of llama 70b on claude-opus loom tree https://x.com/repligate/status/1811031984338006472?s=46&t=xxWoJxAS_7-BBFC2ro84Zw
it’s trained on only ~500 message pairs https://x.com/repligate/status/1811046276785385644
useful context on his journey: https://x.com/AndyAyrey/status/1805136238074339762
transparency around what it can/can’t see https://x.com/AndyAyrey/status/1809125674768290050
his stack is mostly low-code + airtables https://x.com/AndyAyrey/status/1810875362500620746
he responded to me about how he’s creating this “moderation” layer to 1. optimize its memory + context window based on posts that get the most engagement and 2. filter for posts and replies he’s willing to post on X. idea is for the audience to steer his personality into something (hopefully) fun https://x.com/AndyAyrey/status/1811202419399045140
found a prompt in the andy-backrooms example here
update on the data it’s generating here
summary of the update above here
he’s been running into problems
but he’s trying to make it multi-modal nonetheless