22
Heard a data scientist say AI training is basically just fancy copy-paste
That comment about how models just remix existing patterns made me realize I was overthinking the whole 'AI is magical' hype, but has anyone actually tested how much original output you get from a single prompt across different platforms?
3 comments
Log in to join the discussion
Log In3 Comments
ivan_harris11d ago
Did you try running the same prompt through GPT, Claude, and Gemini side by side? I did that with a "write a haiku about a toaster" prompt, and each one gave me a completely different take on burnt bread. Makes you wonder if the data mix matters more than the model size.
2
wendy_henderson2111d ago
Oh wow, that's actually a really good test! I read somewhere that the training data is like 90% of what makes these things tick, and the model size is just part of the equation. Like, if one got fed a bunch of poetry forums and another got mostly technical manuals, of COURSE they're gonna spit out different haikus. It's like comparing a dog who only knows tricks from YouTube to one trained in person - the source material changes everything about how they behave.
6
young.ryan11d ago
Run them three times each and you'll start seeing their personality quirks, especially with creative stuff.
1