A couple months ago I tried to make ChatGPT implement a rather simple algorithm in Python:
- In a data sheet with float numbers, find each row for which col1(i) > col1(i-1) and col1(i) > col1(i+1).
- Find each of those for which there is col2(j) > col2(i) and col3(j) < col3(i).
These were explained in human terms that are not available on the Internet, or are extremely hard to find. It managed to do the first task relatively fast, but even after three hours of trying to implement the second task in a hand-holding way, basically explaining what code to write line-by-line it still couldn’t do it. It’s as if it had severe dementia and was forgetting the previous concept straight upon learning the next one. So, I’m quite skeptical about its capabilities outside of what could be solved with regular functions/macros anyways. I also do not think it will see any significant improvement in the future because IMO the low-hanging fruit of expanding the sample sizes for training are already grabbed and what’s in front of these predictive models is abstract thinking - something only available to humans. Maybe the prediction itself could be abstracted away and there is some “learning meta-pattern” behind it? If so, it would be a breakthrough much more interesting than token generation.
. My favorite analogy is that if the water in a river makes a rock move, it doesn’t mean the rock has actual locomotion the way animals do. The same way with “AI” and “thinking”.
.
