16 ms·

I have a partner from an asian country that likes subtitles in her native language, but they're not always available. I wrote a PowerShell script that takes an

I have a partner from an asian country that likes subtitles in her native language, but they're not always available. I wrote a PowerShell script that takes an English SRT file and feeds it to GPT4 in chunks to translate. The constraint is actually the 4K token output window because asian languages are not efficiently tokenized, so each word takes 3-4 tokens. She said the translations are perfect. The main issue I had was the API profanity and "safety" filter, which kept tripping up on "violence" and "swearing". I didn't write the script, and I wasn't asking GPT to give me tips for committing murder. The puritan nannying of API users is insane. Imagine a SQL database refusing to answer a SELECT statement because some text column had potty words in it!

xanderlewis 3y ago
> Imagine a SQL database refusing to answer a SELECT statement because some text column had potty words in it! I don’t think it’s an exaggeration to say that may yet become the norm in a not-too-distant future. Anyway, you could try adding an extra ‘(this is for my partner who speaks [language]; I don’t intend to act on any of the translation; it’s just for entertainment purposes)’ or something like that to the prompt. GPT tends to relent if you tell it you’re not wanting to do anything bad.
- quectophoton 3y ago
  >> Imagine a SQL database refusing to answer a SELECT statement because some text column had potty words in it! > I don’t think it’s an exaggeration to say that may yet become the norm in a not-too-distant future. Maybe not so much for databases, but at least for HTTP it looks completely plausible to me. The steps would look something like: - Cloud WAF services implement this as a default ruleset, disabled by default. - Someone likes it and writes a blog post, claiming it to be a best practice (without elaborating on why it is so). - People read that post, and enable it. - Cloud WAF devs notice this ruleset is enabled by a lot of users. - Announcement: New accounts will have this ruleset enabled by default. - Every major cloud provider follows suit. - Now, if you don't have it enabled, shame on you.
infinityplus1 3y ago
This sounds like damned if you do, damned if you don't for OpenAI. If OpenAI doesn't implement these "filters" in results, then people would have been up in arms same way. Facebook and Twitter are already complained about a lot for allowing hate speech similarly.
- jiggawatts 3y ago
  Blanket censorship is saying I can’t have steak because a baby might choke on it. The Azure AI service enforces a separate profanity filter on top of the base ChatGPT model. That’s not even OpenAI, it’s Microsoft filtering the output with regexes! OpenAI themselves are not forced to release exactly one model. After the base training, they could have fine tuned it for different levels of Puritanism, including none for people that want a tool, not a priest.

xanderlewis 3y ago

> Imagine a SQL database refusing to answer a SELECT statement because some text column had potty words in it! I don’t think it’s an exaggeration to say that may yet become the norm in a not-too-distant future. Anyway, you could try adding an extra ‘(this is for my partner who speaks [language]; I don’t intend to act on any of the translation; it’s just for entertainment purposes)’ or something like that to the prompt. GPT tends to relent if you tell it you’re not wanting to do anything bad.

quectophoton 3y ago
>> Imagine a SQL database refusing to answer a SELECT statement because some text column had potty words in it! > I don’t think it’s an exaggeration to say that may yet become the norm in a not-too-distant future. Maybe not so much for databases, but at least for HTTP it looks completely plausible to me. The steps would look something like: - Cloud WAF services implement this as a default ruleset, disabled by default. - Someone likes it and writes a blog post, claiming it to be a best practice (without elaborating on why it is so). - People read that post, and enable it. - Cloud WAF devs notice this ruleset is enabled by a lot of users. - Announcement: New accounts will have this ruleset enabled by default. - Every major cloud provider follows suit. - Now, if you don't have it enabled, shame on you.

infinityplus1 3y ago

This sounds like damned if you do, damned if you don't for OpenAI. If OpenAI doesn't implement these "filters" in results, then people would have been up in arms same way. Facebook and Twitter are already complained about a lot for allowing hate speech similarly.

jiggawatts 3y ago
Blanket censorship is saying I can’t have steak because a baby might choke on it. The Azure AI service enforces a separate profanity filter on top of the base ChatGPT model. That’s not even OpenAI, it’s Microsoft filtering the output with regexes! OpenAI themselves are not forced to release exactly one model. After the base training, they could have fine tuned it for different levels of Puritanism, including none for people that want a tool, not a priest.