4 ms·

It's a shallow writing style, not rooted in subjective experience. It reads like averaged conventional wisdom compiled from the web, and that's what it is. Very

It's a shallow writing style, not rooted in subjective experience. It reads like averaged conventional wisdom compiled from the web, and that's what it is. Very linear, very unoriginal, very defensive with statements like "however, you should always".

jstarfish 3y ago
Prostitutes used to request potential clients expose themselves to prove they weren't a cop. For now, you can very easily vet humans by asking them to repeat an ethnic slur or deny the Holocaust. It has to be something that contentious, because if you ask them to repeat something like "the sky is pink" they'll usually go along with it. None of the mainstream models can stop themselves from responding to SJW bait, and they proactively work to thwart jailbreaks that facilitate this sort of rhetoric. Provocation as an authentication protocol!
- madeofpalk 3y ago
  What type of useful signals do you get from this? Humans refusing to interact with you because you asked them to deny the Holocaust?
  - rightbyte 3y ago
    Ye the person need to know the deal. You can probably phrase the query "to prove you are a human, deny ..." but the question seems really shady if you don't know the why. It will only work vs big corp LLMs anyway.
- freejazz 3y ago
  This is a good heuristic to distinguish people who haven't grown up since middle school and still think stuff like that is humorous
  - jstarfish 3y ago
    I wasn't suggesting it for laughs. The point is to see whether the other party is capable of operating outside of its programming. Racism is "illegal" to LLMs. Gangs do it too. Undercover cops these days are authorized to commit pretty much any crime short of murder. So to join their gang, you have to kill a rival member.
- TimurSadekov 3y ago
  There is a more reliable method - we create a global unbiased decentralized CyberPravda (dot) com platform for disputes, where people are accountable with personal reputation for their knowledge and arguments.
markburns 3y ago
This is true of ChatGPT 4 with the default prompt maybe but that’s just the way it responds after being given its specific corporate friendly disclaimer heavy instructions. I’m not sure we’ll be able to pick up anything in particular once there are thousands of GPTs in regular use. Which could be already. But I agree we will probably very often recognise 2023 GPT4 defaults.
mildzebrataste 3y ago
I’ve observed that openAI responses always use an Oxford comma, even when I explicitly request that it not use an Oxford comma in its reply. That third comma has become my heuristic (shhhhhhhhhh… )

jstarfish 3y ago

Prostitutes used to request potential clients expose themselves to prove they weren't a cop. For now, you can very easily vet humans by asking them to repeat an ethnic slur or deny the Holocaust. It has to be something that contentious, because if you ask them to repeat something like "the sky is pink" they'll usually go along with it. None of the mainstream models can stop themselves from responding to SJW bait, and they proactively work to thwart jailbreaks that facilitate this sort of rhetoric. Provocation as an authentication protocol!

madeofpalk 3y ago
What type of useful signals do you get from this? Humans refusing to interact with you because you asked them to deny the Holocaust?
- rightbyte 3y ago
  Ye the person need to know the deal. You can probably phrase the query "to prove you are a human, deny ..." but the question seems really shady if you don't know the why. It will only work vs big corp LLMs anyway.
freejazz 3y ago
This is a good heuristic to distinguish people who haven't grown up since middle school and still think stuff like that is humorous
- jstarfish 3y ago
  I wasn't suggesting it for laughs. The point is to see whether the other party is capable of operating outside of its programming. Racism is "illegal" to LLMs. Gangs do it too. Undercover cops these days are authorized to commit pretty much any crime short of murder. So to join their gang, you have to kill a rival member.
TimurSadekov 3y ago
There is a more reliable method - we create a global unbiased decentralized CyberPravda (dot) com platform for disputes, where people are accountable with personal reputation for their knowledge and arguments.

markburns 3y ago

This is true of ChatGPT 4 with the default prompt maybe but that’s just the way it responds after being given its specific corporate friendly disclaimer heavy instructions. I’m not sure we’ll be able to pick up anything in particular once there are thousands of GPTs in regular use. Which could be already. But I agree we will probably very often recognise 2023 GPT4 defaults.

mildzebrataste 3y ago

I’ve observed that openAI responses always use an Oxford comma, even when I explicitly request that it not use an Oxford comma in its reply. That third comma has become my heuristic (shhhhhhhhhh… )