Hacker News

Hacker News

Can my fine-tuned model pass Pangram?
1 points by meyerdav •                         
Recently, I fine-tuned a small local model for 15 minutes on my own writing. In different instances it is able to have Pangram detect it as 100% human. What are your thoughts? Video: https://lnkd.in/p/dSwieQey
verdverm •                         
Who cares? I don't care what Pangram says, I care about quality writing, with Ai usage or not. Only super anti-Ai people use that site and they ignore all the false positives it produces. It's a site the dog pilers use
meyerdav •                         
Well difficult discussion I feel. On the one hand you could say it is like people writing assembly code => not really needed for the majority of programmers anymore. On the other hand I do feel writing text yourself means also you took more time about the content. So honestly more on the critical side of things in regards to fully AI generated texts
verdverm •                         
It seems like this fine-tune is to use full-auto writing but not get caught, is that right?

Good writing is about organizing your thinking, structuring your arguments, and making a compelling read for your audience. The writing time can be very quick if you have been thinking about something for a long time.

The problem with Ai writing is the outsourcing of thinking

meyerdav •                         
So you believe I cannot train a model to organize text like I would? This is about given people an AI-identity rather than not getting caught. In fact, I believe people will not need to write much in the future
verdverm •                         
No, I don't believe transformers are up to the task, just like for coding. It takes someone who knows about a thing for an Ai to be good at a thing. In other words, Ai does not make experts, Ai enhances experts. You still have to put in the human work to become expert.