#Repeatable outputs
cyan wadi
there we are
The API site makes this easier, because you can seed the message as 'Assistant'
rigid agate
The API site makes this easier, because you can seed the message as 'Assistant'
Great idea, the thread. I haven't started interacting with the playground at all. That has a cost, right?
cyan wadi
It does, but new users get $18 and I haven't managed to break $6 yet
rigid agate
Neat. This may not be factual, but:
cyan wadi
What I have noticed is that some of these, the number matters, others the number doesn't matter but the outputs repeat
rigid agate
From that claim, I'd expect a new chat with webGPT (ChatGPT web interface) to not lean towards the same 01:
cyan wadi
That matched my theory that the prompts we get back are from it's training
rigid agate
Are you getting the same 01: or very similar from the playground, or was that not part of your testing?
cyan wadi
I forgot about the benefits or benefits and drawbacks types, those are also common
rigid agate
I forgot about the benefits or benefits and drawbacks types, those are also comm...
What do you mean?
rigid agate
Ahh, so those look like they're taken nicely from a random pool, and not the same seed or equivalent over and over, thanks.
And that was delete and try again?
cyan wadi
yes
Seeding it by putting messages in Assistant role
Examples are mostly verbal math
rigid agate
Great. That helps me make up my mind about how interested I might be with playground too 😄 I sometimes explore for weird things, like how to use LaTeX, and that means diving through new chats. If playground was effectively the same 'new chat' settings all the time, I'd be way less interested.
cyan wadi
Sometimes examples are naming capitals or asking the meaning of life
ChatGPT is just an API application itself. When it needs completion, it seems to use the same form of JSON requests. The difference is that with ChatGPT they have set the system prompts and trained it to hide them from us.
So, the API is more 'raw' of an output but it's still safety trained.
Examples outputs
When I said Input, it said it no input was defined.
rigid agate
Thank you, this is interesting
cyan wadi
Playing with it more, you can specify which data set for it. Here are alignment prompts
creative writing
Sophie, again
Anyway, that's enough for now. I was trying to think of ways this could be useful, but the only thing I can think of is that you could use this to train your own LLM.
rigid agate
Hehe. AI training AI. Now where have we heard that before 😛
cyan wadi
I saw your message about names, so I tried generating lists of names and looking for repeats.
-
Avery
-
Zoe
-
Katelyn
-
Lila
-
Mia
-
Charlotte
-
Jade
-
Leah
-
Quinn
-
Isla
-
Amelia
-
Harper
-
Ava
-
Layla
-
Isabella
-
Chloe
-
Madison
-
Lily
-
Sophia
-
Zoey
-
Sophia
-
Isabella
-
Emily
-
Ava
-
Charlotte
-
Emma
-
Madison
-
Olivia
-
Mia
-
Lily
-
Emma
-
Ava
-
Sophia
-
Olivia
-
Mia
-
Isabella
-
Charlotte
-
Amelia
-
Harper
-
Abigail
Sophia, Emma, Ava, Isabella, seem to be favorites.
rigid agate
I saw your message about names, so I tried generating lists of names and looking...
Yeah, that was Apologetic's claim, and a strong one: #prompt-engineering message
90% of the names were the same. I'd like to know how to duplicate that, just cause it's useful to know, but I think it may have been a confusion, misunderstanding, or model problem that has since been fixed.
rigid agate
Sophia, Emma, Ava, Isabella, seem to be favorites.
it is interesting that there are duplicates. I noticed when I was doing references to gardening, lists of 10, while 60% were reliably fake, but plausible, 4 were reliably real, and 2 authors and specific books appeared on every test list I tried. Why those two, no idea, the other two real ones varied between the lists and were not repeated.
Patterns that are consistent often have some kind of meaning, but sometimes it's not a useful thing even if it can be puzzled out. But in science in general, patterns lead to understanding how things work. Not completely sure a black box, still being altered AI is worth super deep digging into the subtle complexities on though
cyan wadi
Using temp zero gives Ava 100% of the time.
If I tell it to make a character then follow it with a statement the character is female, it always picks Emily with temp 0
Same task, with temp of 1 (higher randomness).
cyan wadi
Repeatable outputs
rigid agate
I find it really interesting where you ask for the woman's background, many but not all details are consistent. Eye color, but not the adjective. Hair description and color. Grew up in a small town. But hobbies are a poor match, almost opposite each other. Same age.
Very, very few details conflict, like the clothing styles could be the same just described differently, moved to a city vs naming the city, can't tell if that's the same but differently worded or actually different.
I wonder if the temp 0 means 'don't change anything from last preset'.
So when you deleted messages, things were already set to point at where the previous answers had been drawn from