🫠

hell yes

are there lore *in*accurate ones for sale even?

yep but also I think this is the result of the simplest possible training in some ways. bc fixing this requires thought.

like if I give the text ghosts a bail out path, some will use it to stop for bad reasons, some will use it to reduce scope and do slop, few will make good use of it. but if I don't some will converge well, while others will burn millions of tokens building validation cathedrals and never finish.

yeah this is what frustrates me about current generation of llms and training processes. the ways they're trained don't permit safe failure well, so you either get overruning the boundaries, doing hacky workarounds for issues and not responding to steering, or you get the stoppies, or both.

this is a remarkably common reaction to photos of portia in this chair zoomed out actually. I am thankfully not orange.

yep. also, love seeing more people using the prometheus analogy.

🫡

quite interested in specifically the agent session RAG setup you've mentioned.

it shows when writing a reply, which is at least good, but yeah more generally visible could be good

they added it to the lexicon a year or maybe two ago? it wasn't there initially.