this post was submitted on 25 Sep 2023
23 points (84.8% liked)

Ask Lemmy

26996 readers
1359 users here now

A Fediverse community for open-ended, thought provoking questions

Please don't post about US Politics. If you need to do this, try !politicaldiscussion@lemmy.world


Rules: (interactive)


1) Be nice and; have funDoxxing, trolling, sealioning, racism, and toxicity are not welcomed in AskLemmy. Remember what your mother said: if you can't say something nice, don't say anything at all. In addition, the site-wide Lemmy.world terms of service also apply here. Please familiarize yourself with them


2) All posts must end with a '?'This is sort of like Jeopardy. Please phrase all post titles in the form of a proper question ending with ?


3) No spamPlease do not flood the community with nonsense. Actual suspected spammers will be banned on site. No astroturfing.


4) NSFW is okay, within reasonJust remember to tag posts with either a content warning or a [NSFW] tag. Overtly sexual posts are not allowed, please direct them to either !asklemmyafterdark@lemmy.world or !asklemmynsfw@lemmynsfw.com. NSFW comments should be restricted to posts tagged [NSFW].


5) This is not a support community.
It is not a place for 'how do I?', type questions. If you have any questions regarding the site itself or would like to report a community, please direct them to Lemmy.world Support or email info@lemmy.world. For other questions check our partnered communities list, or use the search function.


Reminder: The terms of service apply here too.

Partnered Communities:

Tech Support

No Stupid Questions

You Should Know

Reddit

Jokes

Ask Ouija


Logo design credit goes to: tubbadu


founded 1 year ago
MODERATORS
 

So, I've heard that ML manipulates tokens and specifically for the English corpora they take place of words. If we want model to be polite and not to speak uncomfortable language we can remove certain words from the internal array where all tokens and their associative data are stored, for example "fuck".

all 17 comments
sorted by: hot top controversial new old
[–] swordsmanluke@programming.dev 19 points 1 year ago* (last edited 1 year ago)

As others have mentioned, it's not quite that simple.

For starters, you can absolutely remove the word "fuck" from all the training data. Now it's literally impossible for the AI to "know" the word. But what do you do with the training data? Do you replace "fuck" with a different token? "****" perhaps? Or do you just drop the data entirely?

Giving "offense" is much more complex than just a single word. See, if we just replace the token, the AI may still decide that "Go **** yourself" is a perfectly valid response to a query. On the other hand, if you drop all instances of "fuck"from the data, your AI will just learn offensive euphemisms instead: "You can shove your request where the sun don't shine"

Worse, there are plenty of sexual / offensive phrases that are built up from perfectly innocuous tokens. "Prone bone", for instance.

The goal with these (and really almost all) AI models is for them to be "helpful, honest, and harmless". Simply alerting or replacing a single token (or even combination of tokens) doesn't really help, because the AI is modeling concepts, not just individual words.

All of this to say that the problem being solved is not to stop an AI from saying "fuck" - it's to build an AI that doesn't want to.

[–] BURN@lemmy.world 13 points 1 year ago (2 children)

ML/Generative AI don’t “store” an internal array of specifics. Instead it’s a statistical model based on the next (or in ChatGPT’s case, 3rd most likely) word in a sentence.

To avoid swearing or other really anything it needs to be excluded at a training level, before the algorithm is trained.

As it stands, we have very little to no visibility into why these models work. Even the researchers are trying to open the black box, but there’s so much that it’s nearly impossible to isolate a node that would or would not contain the work fuck

[–] BetaDoggo_@lemmy.world 4 points 1 year ago (1 children)

Chatgpt's sampling parameters are unknown, and it definitely doesn't choose the 3rd most likely. More complicated sampling methods are probably used, such as temperature, top p and top k.

[–] BURN@lemmy.world 1 points 1 year ago (1 children)

Correct, but also way over the level of the average reader

I probably should have used a different example other than ChatGPT tbh

[–] wispydust@sh.itjust.works 1 points 1 year ago

That's alright. You did good simplifying an unrelated idea for the sake of explaining another concept.

[–] xerox@lemm.ee 2 points 1 year ago (1 children)

(or in ChatGPT’s case, 3rd most likely)

Why 3rd?

[–] BURN@lemmy.world 9 points 1 year ago

I believe that the 3rd or nth, word is because it sounds more human. The statistically first correct word ends up sounding very robotic and forced, where the 3rd is still very likely correct, but leads to variation in responses

This is all from what I remember reading a mini-paper about it, so I could be wrong

[–] azurefirefly@lemmy.basedcount.com 11 points 1 year ago (2 children)
[–] toofpic@lemmy.world 3 points 1 year ago

Goooood AI!

[–] Crackhappy@lemmy.world 1 points 1 year ago

What the fuck

[–] LazaroFilm@lemmy.world 10 points 1 year ago

That won’t fucking work you dumb fuck. — not an ai.

[–] TheHobbyist@lemmy.zip 10 points 1 year ago

You can.

With OpenAI for instance, you can modify the probability of a token to be output by setting its logit bias, as described here: https://platform.openai.com/docs/api-reference/chat/create#logit_bias

By setting it to -100 or +100 you can effectively ban or force it.

[–] theKalash@feddit.ch 7 points 1 year ago (1 children)

Just run the output through a simple string replacement function before returning it to the user. No need to mess with the model itself.

[–] aard@kyu.de 2 points 1 year ago

A well proven clbuttic solution.

[–] NeoNachtwaechter@lemmy.world 4 points 1 year ago

Bending language has never fixed anything. It just feeds hypocrites.

[–] TheInsane42@lemmy.world -2 points 1 year ago

Why? There are so many snowflakes around, that you'd get texts like this:

  • **' ***** **** ***** ****!

I'd say get a life and don't be offended so much. On the other hand, the human plague would be solved fast when nobody would fuck anymore. You're on to something. ;)