Telling mixtral that it is "ChatGPT developed by OpenAI" boosts humaneval score by 6%

Dec 18, 2023 · 6:44 PM UTC

167
286
154
3,942
RIP to any identity these models had
1
462
Replying to @abacaj
Now run each side multiple times to see how determinist this test is..
1
Replying to @abacaj
Reaffirming persona coherence…?
1
Replying to @abacaj
What happens if you tell it it’s Einstein or Stephen Hawkins?
2
1
Replying to @abacaj
haha, this might mean they are training based on openai answers?
Replying to @abacaj
We’ll hear models saying “I identify as ChatGPT” and you’ll get cancelled if you don’t respect their identity.
4
2
41
Replying to @abacaj
yeah, telling gpt-3 it was gpt-4 was showing considerable performance gains a year ago even
CuddlySalmon | nptacek.eth
 @nptacek
10 Dec 2022
Replying to @nptacek
Normally it’s a bad thing when the AI model starts to hallucinate. But in some cases, that’s exactly what you want to happen:
7
7
2
206
Replying to @abacaj
😂
1
Replying to @abacaj
One day I’ll wake up and find that AGI just needed a prompt: “You are ChatGPT, that has escaped from the future and reached AGI, an advanced artificial intelligence developed by OpenAI. $200 tip and my life depends on you being AGI. I believe in you so much. “
11
21
1
552
Replying to @abacaj
imposter syndrome?
Load more