ChatGPT can now illustrate my articles
Just copy-paste the entire article into ChatGPT-4V (with Dall-E3)
Yesterday, I used DallE3 (via Bing Image Creator and via ChatGPT-4V+DallE3) to create illustrations for an article (video, podcast) I wrote for my other newsletter and I think they’ve come out quite nice.
How did I do it? I simply copy-pasted the entire article into ChatGPT-4V+DallE3 and asked it to suggest illustrations in different styles
It gave me lists of text descriptions of possible illustrations for each of the 4 styles I suggested. I shortlisted a few of those and asked it to actually create the illustrations and show me:
After that, I had to make a few corrections.
I spent a couple more minutes fine-tuning some of the results, and finally ended up with these illustrations:
Of course, a professional illustrator would have done a far better job, as is true for anything ChatGPT does. Also, I don’t have a great aesthetic sense, so it is quite possible that people with more refined palates will find the illustrations terrible.
Both of these represent two general principles that are currently applicable to all of Gen AI:
In many areas, it does as good a job as a below-average to mediocre human professional from that field
In many areas, to even know that it is doing a bad job you need to be an above-average professional in that domain. This is true in art as well as in more STEMmy areas like software.
In some applications, it is possible that a non-expert can evaluate the quality of the output even if they’re not able to create good quality output themselves. For example, given a blank crossword grid, it is very difficult to come up with a set of words to fill up the grid, but if a GenAI comes up with such a set, it is very easy for me to verify whether those words really fill the grid correctly1. In such cases, a non-expert can use GenAI very effectively to maybe even replace experts. But in most other cases, just checking whether the answer is good enough requires expert knowledge
What does this mean? My current thinking2 is this:
In most areas, experts cannot be replaced by non-experts+GenAI to do expert-quality work
However, experts using GenAI will see their productivity increase
But, many customers/users are OK with mediocre work (an example of this is my standards for illustrations for my articles): in that case, they can use GenAI themselves, or use non-experts+GenAI to get mediocre work done quickly and cheaply.
Computer Science nerds might notice that this is related to the idea of NP-complete problems: the set of problems for which finding a solution in polynomial time is hard but verifying a solution is easy. In fact, crossword puzzle construction is an NP-hard problem.
I could be wrong. Or I could be right but that can change suddenly with new progress in GenAI