3 ms·

I realize this wasn't ChatGPT, and in fact the self-censorship is a big selling point of ChatGPT/OpenAI (to the chagrin of quite a few people), but this does th

by mjr00 4y ago

I realize this wasn't ChatGPT, and in fact the self-censorship is a big selling point of ChatGPT/OpenAI (to the chagrin of quite a few people), but this does throw a wrench into people claiming that it's not a big deal if an LLM gives a bad response to a prompt. Much like a self-driving car AI making a "mistake" can kill a passenger or pedestrian, a LLM mistakenly including slurs in your weekly ChatGPT-powered marketing emails can kill your business.

CGamesPlay 4y ago
Totally true. Which is why these tools should assist with this kind of generation and not completely take it over.
phillipcarter 4y ago
This just calls for more sophisticated systems for dealing with AI generated content.
lm28469 4y ago
If you don't even read the emails you send I think it's fair
- Spivak 4y ago
  If you can't actually use GPT to build even slightly autonomous systems then a lot of the potential value is lost.
  - paulmd 4y ago
    small degree of supervision is the same thing as saying "highly autonomous", just not fully autonomous. "can't do systems with even a slight degree of autonomy" is a complete inversion of the statement here lol. it's like saying it's not even worth having junior coders or underling copywriters because you have to proof-read their output. Sure but that's still a huge force multiplier. IDEs don't generate correct output if you misuse them either, they'll happily do the wrong thing.
    - mjr00 4y ago
      The point, I think, is that a lot of the speculation about what ChatGPT will do is operating under the assumption that it's fully autonomous, not highly autonomous. Which is quite a gulf. A highly autonomous LLM used to generate code can be used as a productivity enhancer by software developers. A fully autonomous one could be used to eliminate software developers entirely. A fully autonomous LLM lets me send out daily personalized marketing emails to every user who signs up for my service. A highly autonomous one requires me to spend time reviewing, which limits how many emails I can send out, though it still boosts my productivity.
  - lm28469 4y ago
    It's a text generator, not an oracle of Truth. If you remove the hype there already is very little potential value, or at least high quality value. It will automate what we pay humans to do for 20ct per article
xeromal 4y ago
The article states that the show is driven by ChatGPT and that the Davinci model had an issue so they put up on older model and the joke came out of that. The latest model does have those safe guards you mentioned.
- mjr00 4y ago
  Sure, but ChatGPT isn't infallible; we've seen tons of posts on HN about various jailbreaking methods to bypass the ChatGPT censorship, and the censorship itself has only done so well because OpenAI is constantly playing whack-a-mole with what I can only imagine are hardcoded kludges in the censorship code on top of the LLM.
marcusverus 4y ago
The existence of a single use case for censorship might serve as a handy excuse for broader, mandatory censorship if you're already in favor of it. But it isn't a very convincing argument for it, otherwise.

by mjr00 4y ago

CGamesPlay 4y ago

Totally true. Which is why these tools should assist with this kind of generation and not completely take it over.

phillipcarter 4y ago

This just calls for more sophisticated systems for dealing with AI generated content.

lm28469 4y ago

If you don't even read the emails you send I think it's fair

Spivak 4y ago
If you can't actually use GPT to build even slightly autonomous systems then a lot of the potential value is lost.
- paulmd 4y ago
  small degree of supervision is the same thing as saying "highly autonomous", just not fully autonomous. "can't do systems with even a slight degree of autonomy" is a complete inversion of the statement here lol. it's like saying it's not even worth having junior coders or underling copywriters because you have to proof-read their output. Sure but that's still a huge force multiplier. IDEs don't generate correct output if you misuse them either, they'll happily do the wrong thing.
  - mjr00 4y ago
    The point, I think, is that a lot of the speculation about what ChatGPT will do is operating under the assumption that it's fully autonomous, not highly autonomous. Which is quite a gulf. A highly autonomous LLM used to generate code can be used as a productivity enhancer by software developers. A fully autonomous one could be used to eliminate software developers entirely. A fully autonomous LLM lets me send out daily personalized marketing emails to every user who signs up for my service. A highly autonomous one requires me to spend time reviewing, which limits how many emails I can send out, though it still boosts my productivity.
- lm28469 4y ago
  It's a text generator, not an oracle of Truth. If you remove the hype there already is very little potential value, or at least high quality value. It will automate what we pay humans to do for 20ct per article

xeromal 4y ago

The article states that the show is driven by ChatGPT and that the Davinci model had an issue so they put up on older model and the joke came out of that. The latest model does have those safe guards you mentioned.

mjr00 4y ago
Sure, but ChatGPT isn't infallible; we've seen tons of posts on HN about various jailbreaking methods to bypass the ChatGPT censorship, and the censorship itself has only done so well because OpenAI is constantly playing whack-a-mole with what I can only imagine are hardcoded kludges in the censorship code on top of the LLM.

marcusverus 4y ago

The existence of a single use case for censorship might serve as a handy excuse for broader, mandatory censorship if you're already in favor of it. But it isn't a very convincing argument for it, otherwise.