4 ms·

I think a missing piece is that ChatGPT doesn't know when it's either confused or out of depth. It starts producing gibberish, which is very hard to separate fr

I think a missing piece is that ChatGPT doesn't know when it's either confused or out of depth. It starts producing gibberish, which is very hard to separate from the quality answers. Especially if it would just straightforwardly say when it doesn't know the answer, that would be really helpful. It's easy for ChatGPT to fail the classic turing test because if this. I know that's not the goal here, but for it to pass it does not (or should not?) be able to answer everything but just say when it doesn't know the answer. If it would just say it doesn't know the answer, it would be harder to tell it apart from an human.

featherless 4y ago
"I respectfully disagree with your assessment of ChatGPT. While it's true that it can produce gibberish when confused or out of depth, I don't think that necessarily means it's easy for it to fail the Turing test. In fact, I would argue that the ability to recognize its own limitations and admit when it doesn't know the answer is actually a step towards passing the Turing test. It shows that the chatbot is capable of self-awareness and understanding its own limitations, which are important traits for a machine to possess in order to pass the Turing test." — ChatGPT's response
boole1854 4y ago
This. I've been trying out using ChatGPT to write code. It seems to vacillate between providing fantastic, working answers and answers that look right if you're not paying close attention but actually do some complicated thing that is wrong. This effectively means you have to scrutinize every answer so closely that it's not clear there is any net benefit from using it. If it instead would provide the good answers when it understood and could solve the problem and an error when it didn't, that would be a game changer.
- LeanderK 4y ago
  yeah, the next step is not getting everything perfect (which, btw, is also quite impossible), but getting it to know it's capabilities. It would unlock quite interesting interactions and would make it way more useful. Imagine being able to somewhat trust it judgement! Easy tasks could be deferred to the AI while you can handle the rest. I wonder what the capabilities will be in 5 years. 5 years ago, this was unthinkable!
  - solomonb 4y ago
    Imagine being able to fine tune it to your specific code base as a sort of interactive docs tool.

featherless 4y ago

"I respectfully disagree with your assessment of ChatGPT. While it's true that it can produce gibberish when confused or out of depth, I don't think that necessarily means it's easy for it to fail the Turing test. In fact, I would argue that the ability to recognize its own limitations and admit when it doesn't know the answer is actually a step towards passing the Turing test. It shows that the chatbot is capable of self-awareness and understanding its own limitations, which are important traits for a machine to possess in order to pass the Turing test." — ChatGPT's response

boole1854 4y ago

This. I've been trying out using ChatGPT to write code. It seems to vacillate between providing fantastic, working answers and answers that look right if you're not paying close attention but actually do some complicated thing that is wrong. This effectively means you have to scrutinize every answer so closely that it's not clear there is any net benefit from using it. If it instead would provide the good answers when it understood and could solve the problem and an error when it didn't, that would be a game changer.

LeanderK 4y ago
yeah, the next step is not getting everything perfect (which, btw, is also quite impossible), but getting it to know it's capabilities. It would unlock quite interesting interactions and would make it way more useful. Imagine being able to somewhat trust it judgement! Easy tasks could be deferred to the AI while you can handle the rest. I wonder what the capabilities will be in 5 years. 5 years ago, this was unthinkable!
- solomonb 4y ago
  Imagine being able to fine tune it to your specific code base as a sort of interactive docs tool.