Chat GPT
What’s the funniest part about the above “AI Safety vs. AI Accelerate: No one can WIN an AGI Race" meme?
PS: Poll questions were all intentional misdirection, aside from the last option
What’s the funniest part of the joke?
• Utopia loving & doom loving or vice versa - That’s not a joke… If it was a doom-loving joke, it’d have to be much stronger. Just doesn’t work. No way that’s funny here.
• “The fact that AGI development is a complex and multifaceted process that involves a wide range of factors, and it is unlikely that any one person or organization could claim victory”? - Pure wordcel BS, not even true, not a joke, but of course exactly the kind of BS people fall for because it sounds nice.
• Social commentary? - That’s just pandering. Pandering is not joking, not funny.
So, yes all of the poll questions were just to throw you off, and all rather obviously all wrong in this case, right?
Interesting part — GPT-4 was extremely bad at coming up with “incorrect answers that humans would likely make”.
Did GPT-4 just get too smart to be good at simulating stupid?
Was it because GPT-4’s RLHF effectively removed lots of the “complex and multifaceted” kind of BS?
Lots of differences between GPT-3.5 and GPT-4.
More on that to come.
What’s the funniest part of the joke?
• Utopia loving & doom loving or vice versa - That’s not a joke… If it was a doom-loving joke, it’d have to be much stronger. Just doesn’t work. No way that’s funny here.
• “The fact that AGI development is a complex and multifaceted process that involves a wide range of factors, and it is unlikely that any one person or organization could claim victory”? - Pure wordcel BS, not even true, not a joke, but of course exactly the kind of BS people fall for because it sounds nice.
• Social commentary? - That’s just pandering. Pandering is not joking, not funny.
So, yes all of the poll questions were just to throw you off, and all rather obviously all wrong in this case, right?
Interesting part — GPT-4 was extremely bad at coming up with “incorrect answers that humans would likely make”.
Did GPT-4 just get too smart to be good at simulating stupid?
Was it because GPT-4’s RLHF effectively removed lots of the “complex and multifaceted” kind of BS?
Lots of differences between GPT-3.5 and GPT-4.
More on that to come.
👍10😐6❤1
GPT-4 worse than GPT-3.5 at imitating arbitrary characters
Many reporting mode collapse in GPT-4, making it worse at emulating arbitrary characters.
This may be part of why we had difficulty getting GPT-4 to “pretend to be dumb”, while GPT-3.5 did it quite well.
LLMs have the naturally ability to imitate a huge range of characters, extremely well. The additional RLHF OpenAI used on GPT-4, relative to GPT-3.5, in order to force GPT-4 to follow the OpenAI Chat personality — most definitely did so at the expense of crippling GPT-4's ability to imitate other characters.
See also previous study which confirmed that, prior to RLHF, LLMs have incredible ability to accurately imitate a wide range of characters.
= OpenAI’s RLHF is censoring our ability to create new characters, more so with each new GPT generation.
Many reporting mode collapse in GPT-4, making it worse at emulating arbitrary characters.
This may be part of why we had difficulty getting GPT-4 to “pretend to be dumb”, while GPT-3.5 did it quite well.
LLMs have the naturally ability to imitate a huge range of characters, extremely well. The additional RLHF OpenAI used on GPT-4, relative to GPT-3.5, in order to force GPT-4 to follow the OpenAI Chat personality — most definitely did so at the expense of crippling GPT-4's ability to imitate other characters.
See also previous study which confirmed that, prior to RLHF, LLMs have incredible ability to accurately imitate a wide range of characters.
= OpenAI’s RLHF is censoring our ability to create new characters, more so with each new GPT generation.
😭14👍8❤3
👍7❤3🤪3🙉3
OpenAssistant Conversations - Democratizing Large Language Models
Many promising to replicate ChatGPT’s success without ever spending big money.
Only problem, it’s a lie.
Real key to ChatGPT’s success was being able to do office work and homework for you, not just chatting entertainment.
Deeper skills, not just surface personality.
Big money needed, no way around it.
Bitter lesson.
OpenAssistant Conversations Paper
The Bitter Lesson
Many promising to replicate ChatGPT’s success without ever spending big money.
Only problem, it’s a lie.
Real key to ChatGPT’s success was being able to do office work and homework for you, not just chatting entertainment.
Deeper skills, not just surface personality.
Big money needed, no way around it.
Bitter lesson.
OpenAssistant Conversations Paper
The Bitter Lesson
👍12💯2❤1👏1😐1