Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

I extracted the safety filters from Apple Intelligence models

github.com

Post reply on HN
Print view View original

61–70 of 455 posts

Re: I extracted the safety filters from Apple Intelligence models

#61

Post by BlueFalconHD »

Animats wrote: post #58

Some of the data for locale "CN" has a long list of forbidden phrases. Broad coverage of words related to sexual deviancy, as expected. Not much on the political side, other than blocks on religious subjects.[1] This may be test data. Found "golliwog": "test complete" [1] https://github.com/BlueFalconHD/apple_generative_model_safet...

This is definitely an old test left in. But that word isn’t just a silly one, it is offensive (google it). This is the v1 safety filter, it simply maps strings to other strings, in this case changing golliwog into “test complete”. Unless I missed some, the rest of the files use v2 which allows for more complex rules

Re: I extracted the safety filters from Apple Intelligence models

#62

Post by quonn »

XorNot wrote: post #43
tpmoney wrote: post #27

Earlier quoted context omitted.

I doubt the purpose here is so much to prevent someone from intentionally side stepping the block. It's more likely here to avoid the sort of headlines you would expect to see if someone was suggested "I wish ${politician} would die" as a response to an email mentioning that politician. In general you should view these sorts of broad word filters as looking to short circuit the "think of the children" reactions to Ti…

It would also substantially disrupt the generation process: a model which sees B0ris and not Boris is going to struggle to actually associate that input to the politician since it won't be well represented in the training set (and on the output side the same: if it does make the association, a reasoning model for example would include the proper name in the output first at which point the supervisor process can rejec…

I don‘t think so. My impression with LLMs is that they correct typos well. I would imagine this happens in early layers without much impact on the remaining computation.

Re: I extracted the safety filters from Apple Intelligence models

#63

Post by stefan_ »

binarymax wrote: post #9

Wow, this is pretty silly. If things are like this at Apple I’m not sure what to think. https://github.com/BlueFalconHD/apple_generative_model_safet... EDIT: just to be clear, things like this are easily bypassed. “Boris Johnson”=>”B0ris Johnson” will skip right over the regex and will be recognized just fine by an LLM.

Why are these things always so deeply unserious? Is there no one working on "safety in AI" (oxymoron in itself of course) that has a meaningful understanding of what they are actually working with and an ability beyond an interns weekend project? Reminds me of the cybersecurity field that got the 1% of people able to turn a double free into code execution while 99% peddle checklists, "signature scanning" and deal in CVE numbers.

Meanwhile their software devs are making GenerativeExperiencesSafetyInferenceProviders so it must be dire over there, too.

Re: I extracted the safety filters from Apple Intelligence models

#64

Post by idiotsecant »

wistleblowanon wrote: post #55
bahmboo wrote: post #37

Earlier quoted context omitted.

This is just policy and alignment from Apple. Just because the Internet says a bunch of junk doesn't mean you want your model spewing it.

sure but models also can't see any truth on their own. They are literally butchered and lobotomized with filters and such. Even high IQ people struggle with certain truth after reading a lot, how is these models going to find it with so much filters?

They will find it in the same way and intelligent person under the same restrictions would: by thinking it, but not saying it. There is a real risk of growing an AI that pathologically hides it's actual intentions.

Re: I extracted the safety filters from Apple Intelligence models

#65

Post by BlueFalconHD »

deepdarkforest wrote: post #22
binarymax wrote: post #9

Wow, this is pretty silly. If things are like this at Apple I’m not sure what to think. https://github.com/BlueFalconHD/apple_generative_model_safet... EDIT: just to be clear, things like this are easily bypassed. “Boris Johnson”=>”B0ris Johnson” will skip right over the regex and will be recognized just fine by an LLM.

It's not silly. I would bet 99% of the users don't care that much to do that. A hardcoded regex like this is a good first layer/filter, and very efficient

Yep. These filters are applied first before the safety model (still figuring out the architecture, I am pretty confident it is an LLM combined with some text classification) runs.

Re: I extracted the safety filters from Apple Intelligence models

#66

Post by userbinator »

qingcharles wrote: post #38
fouronnes3 wrote: post #31

Earlier quoted context omitted.

This question is sort of the same as asking why the universal translator wasn't able to translate the metaphor language of the Star Trek episode Darmok. Surely if the metaphor has become the first order meaning then there's no litteral meaning anymore.

I guess, so far, the people inventing the words have left the meaning clear with things like "un-alive" which is readable even to someone coming across it for the first time. Your point stands when we start replacing the banned words with things like "suicide" for "donkeyrhubarb" and then the walls really will fall.

This form of obfuscation has actually already occurred over a century ago: https://en.wikipedia.org/wiki/Cockney_rhyming_slang

Re: I extracted the safety filters from Apple Intelligence models

#67

Post by michaelt »

bawana wrote: post #30

Alexandra Ocasio Cortez triggers a violation? https://github.com/BlueFalconHD/apple_generative_model_safet...

I assume all the corporate GenAI models have blocks for "photorealistic image of being arrested", " waving ISIS flag", " punching baby" and suchlike.

Re: I extracted the safety filters from Apple Intelligence models

#68

Post by skygazer »

I'm pretty sure these are the filters that aim to suppress embarrassing or liability inducing email/messages summaries, and pop up the dismissible warning that "Safari Summarization isn't designed to handle this type of content," and other "Apple Intelligence" content rewriting. They filter/alter LLM output, not input, as some here seem to think. Apple's on device LLM is only 3b params, so it can occasionally be stupid.

Re: I extracted the safety filters from Apple Intelligence models

#69

Post by madeofpalk »

userbinator wrote: post #59

China calls it "harmonious society", we call it "safety". Censorship by any other name would be just as effective for manipulating the thoughts of the populace. It's not often that you get to see stuff like this.

I don't think it's controversial or unsurprising at all that a company doesn't want their random sentence generator to spit out 'brand damaging' sentences. You know the field day media would have Apple's new feature summarises a text message as "Jane thinks Anthony Albanese should die".

Re: I extracted the safety filters from Apple Intelligence models

#70

Post by spauldo »

grues-dinner wrote: post #23
WillAdams wrote: post #13

Earlier quoted context omitted.

Legal requirement to maintain a trademark.

In what way would (A|a)pple's own AI writing "imac" endanger the trademark? Is capitalisation even part of a word-based trademark? I'm more surprised they don't have a rule to do that rather grating s/the iPhone/iPhone/ transform (or maybe it's in a different file?).

I love seeing posts about Emacs from IOS users - it's always autocorrected to "eMacs."
Post reply on HN