2 ms·
Full tweet in case there's a login wall: Wild result. gpt-4-turbo over the API produces (statisticall significant) shorter completions when it "thinks"
by simonw 3y ago
Full tweet in case there's a login wall:
Wild result. gpt-4-turbo over the API produces (statisticall
significant) shorter completions when it "thinks" its December
vs. when it thinks its May (as determined by the date in the
system prompt).
I took the same exact prompt over the API (a code completion
task asking to implement a machine learning task without
libraries).
I created two system prompts, one that told the API it was May
and another that it was December and then compared the
distributions.
For the May system prompt, mean = 4298
For the December system prompt, mean = 4086
N = 477 completions in each sample from May and December
t-test p < 2.28e-07
To reproduce this you can just vary the date number in the
system message. Would love to see if this reproduces for others.
This is part of the ongoing discussion about whether or not ChatGPT has got "lazier".
OpenAI claimed they had made no model changes and were puzzled as to what was going on: https://twitter.com/ChatGPTapp/status/1732979491071549792 https://twitter.com/ChatGPTapp/status/1732979491071549792
we've heard all your feedback about GPT4 getting
lazier! we haven't updated the model since Nov 11th,
and this certainly isn't intentional. model behavior
can be unpredictable, and we're looking into fixing
it
The theory was that maybe the system prompt they inject telling it the current date might be influencing it, because its training data showed people worked less hard in December.
New evidence suggests that theory might actually hold up!
- nerpderp82 3y agoLLMs are an amazing high dimensional lens into our society. When they get more capable and can move that from system 1 to system 2, they will just tell us that we write less in December and probably why! Imagine being an sociologist with the ability to train LLMs on very specific hyperplanes of data.