Turning arbitrary text input into something structured can be a nice UI. For example, being able to say "meeting 2pm Thursday" is sometimes nicer than form input. I expect it's going to be easier...
Turning arbitrary text input into something structured can be a nice UI. For example, being able to say "meeting 2pm Thursday" is sometimes nicer than form input. I expect it's going to be easier to make UI's like that from now on.
I'd want some kind of confirmation that it did it what I wanted, though. There's also a question of what's the best way to make a correction: tell it was the mistake was or edit it yourself?
I really like this idea! I don't have an OpenAI API key, can't try it myself, but it seems like a brilliant approach for generating structured data. This could go horribly wrong with user input in...
I really like this idea! I don't have an OpenAI API key, can't try it myself, but it seems like a brilliant approach for generating structured data. This could go horribly wrong with user input in production though. shudder
Prompt injection is the real danger; if you’re feeding user input into an LLM, you’re opening the door to bad actors crafting jailbreaks that circumvent whatever rails you’ve put in place. Also...
Prompt injection is the real danger; if you’re feeding user input into an LLM, you’re opening the door to bad actors crafting jailbreaks that circumvent whatever rails you’ve put in place.
Also I’m generally wary of the randomness they produce. You might have 99% certainty that it’s going to give you well-formatted JSON but that still means 1% of the time it’ll hallucinate a curly brace in the wrong place and whatever system you’re feeding it into will fail to parse it correctly.
Maybe I’m needlessly concerned and it’s more reliable than I’m giving it credit for, I don’t know. I usually err on the side of caution with stuff like this. Latent space is inherently unpredictable. So I tend to view this tech as a great tool for supervised engineer use, but not for public-facing unsupervised data processing.
Turning arbitrary text input into something structured can be a nice UI. For example, being able to say "meeting 2pm Thursday" is sometimes nicer than form input. I expect it's going to be easier to make UI's like that from now on.
I'd want some kind of confirmation that it did it what I wanted, though. There's also a question of what's the best way to make a correction: tell it was the mistake was or edit it yourself?
I really like this idea! I don't have an OpenAI API key, can't try it myself, but it seems like a brilliant approach for generating structured data. This could go horribly wrong with user input in production though. shudder
What are you imaging could go wrong?
Prompt injection is the real danger; if you’re feeding user input into an LLM, you’re opening the door to bad actors crafting jailbreaks that circumvent whatever rails you’ve put in place.
Also I’m generally wary of the randomness they produce. You might have 99% certainty that it’s going to give you well-formatted JSON but that still means 1% of the time it’ll hallucinate a curly brace in the wrong place and whatever system you’re feeding it into will fail to parse it correctly.
Maybe I’m needlessly concerned and it’s more reliable than I’m giving it credit for, I don’t know. I usually err on the side of caution with stuff like this. Latent space is inherently unpredictable. So I tend to view this tech as a great tool for supervised engineer use, but not for public-facing unsupervised data processing.
Makes a lot of sense, thanks for explaining