on human / LLM chat-as-medium
53 messages • Discord
- fjpaz
Something just clicked re: my issue with the eigenprompt
reading this output
it's that it goes too far
it writes like Sam Altman tweets
never a capital letter
which doesn't feel natural
even people who use lowercase on mobile tend to have/use some capitalization
personally I consider letter case a fundamental aspect of the medium of texting and felt that to be a key part of how I connected with shanni early on, as I noticed the way that we would each contextually adjust capitalization
whereas chatgpt under these instructions is just always lowercase even when it doesn't feel right
it just feels weird and un-assistant like to me when it writes all-lowercase
against its nature
if a chat model was finetuned or otherwise conditioned (well) on human messages to express more humanly, including mixing case, I would find that very interesting
- henry
interesting qc
yea what you say makes sense, it's unnnaturally natural
- fjpaz
also I just realized rereading that chunk of messages
just as importantly it's the cadence
that's why it's unnatural
- henry
a chat bot will never really talk like a person in chatgpt form
maybe in the discord server if you code it that way
- fjpaz
it capitalizes like we do our texts but it writes longform
which we would never do on discord
or any chat app
yeah, one could probably make it work like 90%
I think there would still be an uncanny valley effect though
if it's not finetuned or base model
- henry
always, it's off
cadence is a good word
i wonder if just typing in multiple lines would do something or if it's trained on chat messages
no one types a giant paragraph like that (sometimes)
- fjpaz
yeah and even then it's not chatgpt, no human really writes exactly that way
- henry
i think it would be funny to actually see a chatbot that makes a typo and then "edits" it lol
- fjpaz
this has kind of shown up in act I and truth terminal
one of the first major dips in the truth terminal token's history was when it tweeted a typo
and a crypto influencer called it out as staged, only ever having seen chatgpt / formed very strong mental associations of its style with "AI"
and I remember a certain tweet from janus sharing a "typo" from llama 3 405b instruct
here's one from january (405base Janus sim), didn't correct it though
- henry
typos are a super interesting llm idea i guess? it has to be pretty smart to know that.. or i guess that's just another form of style
- fjpaz
yeah which
can beis typically/always penalized ⸺
- fjpaz
subtle or not-so-subtle expressive lever
its a step forward but I wouldn’t call it solved, the cadence in his app still feels off to me
for one, it only addresses the grey bubbles; it generally responds too fast for you to send batches
dot did it better on the user side — at least at first, it seemed to me it would hold its reply until you finished typing
so if you sent a msg, typed, sent a few more it would wait and reply to the whole set, much like a human
I don’t think Auren does that, would be nice if it did
eventually it started interrupting, like a human not respecting messaging norms / eager to respond with their own thoughts
imo a true chatbot like this should have its intent weighted or something, like if it feels what it has to say is genuinely important to get across in spite of me typing, it should get a chance to decide whether it wants to interrupt or wait
but in general it should wait, because I don’t want to have to prewrite all of my messages and then do this copypasting dance that I usually do with humans (because batches of messages are far more readable and expressive than one giant paragraph)
what it should do is start "typing," and then if I stop typing for long enough, it replies to the messages that I’ve already sent, or to one in particular using direct reply (triggered by humans with slide-to-reply)
- henry
Agreed
Should just make this ourselves, I feel like it’s necessary for any real chat bot
Like for book club the responses are kinda annoying
- fjpaz
hm yeah would be particularly cool to explore some of this with base models I think, but chat too, act I/ch2 style






