Good article, Maxim, thank you. I agree with the commenter below that the AI style is a giveaway (and with you that GPT is actually extremely strong at changing style, even GPT-2 was very strong at this already).
The poetry is pretty rudimentary. I feel that, like your essay example, it is middle school level, but more like a B- effort. It mostly does vague attempts at AABB/ABAB that scan poorly. It will produce a "haiku" that violate the syllable structure, and that's about as simple as it can get. I think the best thing I saw is that it can do a villanelle of sorts.
Input: Please write a villanelle on the topic of using excess deaths to measure Covid effects uniformly across countries.
Re: Diplomacy it was immediately obvious that Austria was the AI. One issue I've seen with GPT (both in my own research and online) is that it really really wants to speak in full sentences and formal language, and is not good at casual texting
you have to give it inputs on the style of speech you want the answer to be in, I also find giving it portions of text to reference and then tell it to output that style..
it takes a few rewrites but its sufficient compared to the original answer
It´s hardcoded to have political bias which makes it wrong in hilarous ways in other areas too. If you ask it wheather a human is more intelligent than a fly, it recommend assessing the individual. Or if you ask wheather a poisonous snake is more dangerous than a babysheep, it also keeps insisting both can be dangerous.
Good article, Maxim, thank you. I agree with the commenter below that the AI style is a giveaway (and with you that GPT is actually extremely strong at changing style, even GPT-2 was very strong at this already).
The poetry is pretty rudimentary. I feel that, like your essay example, it is middle school level, but more like a B- effort. It mostly does vague attempts at AABB/ABAB that scan poorly. It will produce a "haiku" that violate the syllable structure, and that's about as simple as it can get. I think the best thing I saw is that it can do a villanelle of sorts.
Input: Please write a villanelle on the topic of using excess deaths to measure Covid effects uniformly across countries.
Output: The pandemic grips us tight
With fear and death in tow
We search for ways to fight
Excess deaths show us the light
A metric to measure and know
The pandemic grips us tight
Uniform across lands, a uniform sight
No country escapes its blow
We search for ways to fight
The data speaks with all its might
A somber tale of woe
The pandemic grips us tight
We must learn from our plight
And better prepare for the next go
We search for ways to fight
The villanelle repeats with all its might
As the pandemic continues to grow
The pandemic grips us tight
We search for ways to fight
Ha, good to know!
I agree, the poetry is somewhat "B" grade, but probably still better than what the average human would compose.
Re: Diplomacy it was immediately obvious that Austria was the AI. One issue I've seen with GPT (both in my own research and online) is that it really really wants to speak in full sentences and formal language, and is not good at casual texting
It wasn't so obvious to me.
For example, the "get you back on your feet" from Austria at the end is very colloquial. I agree that the formal grammar is the biggest "tell."
But ChatGPT can speak however you ask it to, including in broken grammar! Check out this prompt I just entered: http://covid19mapped.com/AI_Casual.png
Interesting, I had been asking for responses in the form of a text message. I'll give that a try, thank you
you have to give it inputs on the style of speech you want the answer to be in, I also find giving it portions of text to reference and then tell it to output that style..
it takes a few rewrites but its sufficient compared to the original answer
It´s hardcoded to have political bias which makes it wrong in hilarous ways in other areas too. If you ask it wheather a human is more intelligent than a fly, it recommend assessing the individual. Or if you ask wheather a poisonous snake is more dangerous than a babysheep, it also keeps insisting both can be dangerous.
Your point about critical thinking - there's some interesting early signs here:
https://openai.com/blog/critiques/