4 ms·

chatGPT seems amazing at taking corrections, iterating on the requirements, fixing as it goes. So the badly-specified requests are going to be sorted out by mul

chatGPT seems amazing at taking corrections, iterating on the requirements, fixing as it goes. So the badly-specified requests are going to be sorted out by multiple rounds of interaction.

Widdershin 4y ago
Try to get chatgpt to solve a basic math problem using BEDMAS instead of PEMDAS, even with multiple interactions and reminders it still gets incredibly confused.
- wcoenen 4y ago
  ChatGPT cannot correctly perform step by step operations within strict formal systems. On the other hand, it can provide a high level sketch of how tomathematically prove something[1]. I think this is amazing, because this is the opposite of how we normally think about computers. ChatGPT is highly intuitive instead of a formal reasoning system. No doubt some future AIs will combine this intuition with checks against formal verification engines to check correctness. [1] Section 6 of https://yaofu.notion.site/How-does-GPT-Obtain-its-Ability-Tracing-Emergent-Abilities-of-Language-Models-to-their-Sources-b9a57ac0fcf74f30a1ab9e3e36fa1dc1 https://yaofu.notion.site/How-does-GPT-Obtain-its-Ability-Tr...
sgerenser 4y ago
I tried giving ChatGPT an extremely easy “warmup” interview question, and it passed it with flying colors. But then I added the catch that it has to handle UTF-8 strings and not just ASCII, and it confidently spewed out the same program, claiming that by using the c library functions isspace and ispunct, unicode is handled because they let you pass a locale. This is, of course completely wrong. I even gave an example of an input it would fail on and it shrugged and returned roughly the same function that would still fail on that input.
- wwweston 4y ago
  On the plus side, this sounds a lot like the kind of mistake a human engineer would make.
  - sgerenser 4y ago
    Sort of, but I wouldn’t expect a human engineer to be so confidently incorrect. In fact, it would be a big red flag. I’d much rather have a “you know, I’m not very familiar with Unicode or UTF-8, so I’d have to do some research before I could come up with an answer.”
- rpep 4y ago
  I did similar, I asked it to compute the median on some filtered time series data, but as I added more requirements to the filtering it couldn’t seem to adapt the program.
krageon 4y ago
ChatGPT is an interactive prompt meant to amuse you with blatant lies, not something meant for any sort of reasonable purpose.

chatGPT seems amazing at taking corrections, iterating on the requirements, fixing as it goes. So the badly-specified requests are going to be sorted out by multiple rounds of interaction.

Widdershin 4y ago

Try to get chatgpt to solve a basic math problem using BEDMAS instead of PEMDAS, even with multiple interactions and reminders it still gets incredibly confused.

wcoenen 4y ago
ChatGPT cannot correctly perform step by step operations within strict formal systems. On the other hand, it can provide a high level sketch of how tomathematically prove something[1]. I think this is amazing, because this is the opposite of how we normally think about computers. ChatGPT is highly intuitive instead of a formal reasoning system. No doubt some future AIs will combine this intuition with checks against formal verification engines to check correctness. [1] Section 6 of https://yaofu.notion.site/How-does-GPT-Obtain-its-Ability-Tracing-Emergent-Abilities-of-Language-Models-to-their-Sources-b9a57ac0fcf74f30a1ab9e3e36fa1dc1 https://yaofu.notion.site/How-does-GPT-Obtain-its-Ability-Tr...

sgerenser 4y ago

I tried giving ChatGPT an extremely easy “warmup” interview question, and it passed it with flying colors. But then I added the catch that it has to handle UTF-8 strings and not just ASCII, and it confidently spewed out the same program, claiming that by using the c library functions isspace and ispunct, unicode is handled because they let you pass a locale. This is, of course completely wrong. I even gave an example of an input it would fail on and it shrugged and returned roughly the same function that would still fail on that input.

wwweston 4y ago
On the plus side, this sounds a lot like the kind of mistake a human engineer would make.
- sgerenser 4y ago
  Sort of, but I wouldn’t expect a human engineer to be so confidently incorrect. In fact, it would be a big red flag. I’d much rather have a “you know, I’m not very familiar with Unicode or UTF-8, so I’d have to do some research before I could come up with an answer.”
rpep 4y ago
I did similar, I asked it to compute the median on some filtered time series data, but as I added more requirements to the filtering it couldn’t seem to adapt the program.

krageon 4y ago

ChatGPT is an interactive prompt meant to amuse you with blatant lies, not something meant for any sort of reasonable purpose.