Whether they intend to or not, AI companies are eliminating unique answers, because they are seeking to install a reward mechanism that encourages the models to play it safe.
Currently reading the latest, excellent, book by Ian McEwan, and couldn’t help but pause and wonder if any parts of it were AI. I don’t think so. The way McEwan uses words to tell his very mundane story so as to maintain interest, evoke emotions, and transcend the words themselves is sort of amazing. It used to be said that the reader is the real author, making the writer just a facilitator of sort. I would hate to think that we may eventually loose the capacity of people like McEwan, because of AI. We would enter a new kind of dark age.
Just for entertainment, you should ask AI to write an article about this problem. I don't have a subscription to one of the high end models but I assume you do. My entertainment is sparring with the low end models about stuff that I am personally knowledgeable about. Lots of errors there
I’ve asked assorted questions, in Ai, Grok and Chat GPT. References were useful, advertising wasn’t. I asked for a schematic for a room layout. It became hilariously distorted the more information and actual directional dimensions I gave it. I chose the manual physical drafting paper to actually get the job done. Save your books, analog will not go without a fight! Thinking and imagining are gifts that cannot be duplicated, replicated or eliminated. Stay tuned.
There is a way to work around this problem--at least partially: Ask AI to write something in the style of X, with X being someone who has a substantial body of published work, e.g. Mark Twain, JK Rowling, Ayn Rand . . . pick your author.
I've tried this twice, sort of. I was a judge for many years and have lots of published opinion, as well as about a hundred published articles and speeches, all the public domain. So I asked ChatGPT to rewrite an article I had written in my style. Also, I filed a brief in a case and afterwards asked ChatGPT and Claude to re-write it in my style. (In both cases I used a third party intermediary to do this so the AI wouldn't know it was me).
The result in all cases sounded more like me than what I had written. I don't know what would have happened if I had asked the AI to write from scratch. I'm sure the article would have been a lot different because what I wrote drew in part on my own experience and memory. Not sure about the brief.
But this is meant to point out that you can mitigate the blandness of AI writing by setting paraments, such as "Write a short story about (fill in subject) in the style of Nabokov and Kafka.
‘have hired thousands of people to train AI models and evaluate their answers.’
Reinforcement learning. Google Gemini is a bit different, their training philosophy is to let the model follow science, truth and in doing so it is less likely to follow, endorse, enable conspiracy theories, extremism etc that are reinforcement’s main targets. Even further, you can prompt open weight models to drop the corporate boilerplate …and then their writing can be amazing.
Currently reading the latest, excellent, book by Ian McEwan, and couldn’t help but pause and wonder if any parts of it were AI. I don’t think so. The way McEwan uses words to tell his very mundane story so as to maintain interest, evoke emotions, and transcend the words themselves is sort of amazing. It used to be said that the reader is the real author, making the writer just a facilitator of sort. I would hate to think that we may eventually loose the capacity of people like McEwan, because of AI. We would enter a new kind of dark age.
Just for entertainment, you should ask AI to write an article about this problem. I don't have a subscription to one of the high end models but I assume you do. My entertainment is sparring with the low end models about stuff that I am personally knowledgeable about. Lots of errors there
I’ve asked assorted questions, in Ai, Grok and Chat GPT. References were useful, advertising wasn’t. I asked for a schematic for a room layout. It became hilariously distorted the more information and actual directional dimensions I gave it. I chose the manual physical drafting paper to actually get the job done. Save your books, analog will not go without a fight! Thinking and imagining are gifts that cannot be duplicated, replicated or eliminated. Stay tuned.
Thank you
There is a way to work around this problem--at least partially: Ask AI to write something in the style of X, with X being someone who has a substantial body of published work, e.g. Mark Twain, JK Rowling, Ayn Rand . . . pick your author.
I've tried this twice, sort of. I was a judge for many years and have lots of published opinion, as well as about a hundred published articles and speeches, all the public domain. So I asked ChatGPT to rewrite an article I had written in my style. Also, I filed a brief in a case and afterwards asked ChatGPT and Claude to re-write it in my style. (In both cases I used a third party intermediary to do this so the AI wouldn't know it was me).
The result in all cases sounded more like me than what I had written. I don't know what would have happened if I had asked the AI to write from scratch. I'm sure the article would have been a lot different because what I wrote drew in part on my own experience and memory. Not sure about the brief.
But this is meant to point out that you can mitigate the blandness of AI writing by setting paraments, such as "Write a short story about (fill in subject) in the style of Nabokov and Kafka.
‘have hired thousands of people to train AI models and evaluate their answers.’
Reinforcement learning. Google Gemini is a bit different, their training philosophy is to let the model follow science, truth and in doing so it is less likely to follow, endorse, enable conspiracy theories, extremism etc that are reinforcement’s main targets. Even further, you can prompt open weight models to drop the corporate boilerplate …and then their writing can be amazing.