I have found that in my experience, Claude is really good at following orders and fixing its own mistakes, but when it comes to showing any sort of ingenuity or creativity with problem-solving, it can be pretty useless.
For example, I used it to help me build a web image scraper. I wanted de-duping, so Claude used a hash on the file. This isn't a "wrong" way to do de-duping, but doing something like a perceptual hash made much more sense in my situation. I asked it to use phash for the image hashing and it kept both the file hash and the phash (??). I had to ask it to stop doing the file hash and stick to just the phash. I feel like Claude is like an ultra-literal developer who can program and read at 1000x the speed of a human but lacks any sort of ingenuity or problem-solving skills which is so essential to my job.
Nonetheless I have found it super useful for throwing stuff out and getting making me demos or one-time scripts, but tends to solve problems in the most convoluted and complicated word-vomitty way possible.
That's a good example... I definitely think that Claude has trained my "slop-dar" more than my actual coding skill... And a risk is that, in areas that I am inexperienced, that I won't be able to distinguish the slop, or even worse, actual poor solutions, from the good.
It seems pretty likely to me though that this is a short-term limitation, and newer models (even Fable, or Opus 5.5), are becoming good enough for more than just prototypes or demos.
I have found that in my experience, Claude is really good at following orders and fixing its own mistakes, but when it comes to showing any sort of ingenuity or creativity with problem-solving, it can be pretty useless.
For example, I used it to help me build a web image scraper. I wanted de-duping, so Claude used a hash on the file. This isn't a "wrong" way to do de-duping, but doing something like a perceptual hash made much more sense in my situation. I asked it to use phash for the image hashing and it kept both the file hash and the phash (??). I had to ask it to stop doing the file hash and stick to just the phash. I feel like Claude is like an ultra-literal developer who can program and read at 1000x the speed of a human but lacks any sort of ingenuity or problem-solving skills which is so essential to my job.
Nonetheless I have found it super useful for throwing stuff out and getting making me demos or one-time scripts, but tends to solve problems in the most convoluted and complicated word-vomitty way possible.
I appreciated your article though. Keep it up.
That's a good example... I definitely think that Claude has trained my "slop-dar" more than my actual coding skill... And a risk is that, in areas that I am inexperienced, that I won't be able to distinguish the slop, or even worse, actual poor solutions, from the good.
It seems pretty likely to me though that this is a short-term limitation, and newer models (even Fable, or Opus 5.5), are becoming good enough for more than just prototypes or demos.
Really enjoyed your thoughts!
Thanks so much!