[HN Gopher] Apple Intelligence notification summaries are pretty...
       ___________________________________________________________________
        
       Apple Intelligence notification summaries are pretty bad
        
       Author : voytec
       Score  : 30 points
       Date   : 2024-11-18 19:38 UTC (3 hours ago)
        
 (HTM) web link (arstechnica.com)
 (TXT) w3m dump (arstechnica.com)
        
       | whycome wrote:
       | Incidentally, I turned this off today. I suspect it's terrible on
       | battery life and I will find out. But the thing about the
       | summaries that was they would sometimes imply the EXACT OPPOSITE
       | of what was in a message. I had a few stomach-dropping moments
       | when reading the summaries only for me to read the actual thread
       | to see it was nowhere close. This is one of "it's not even wrong"
       | situations and I don't know how it was fucked up this badly. The
       | nature of the texts themselves weren't complicated either. I
       | didn't save them, but I suspect it stemmed from misinterpretating
       | some subtle omission (like our common practice of leaving out
       | articles or pronouns).
        
         | jiggawatts wrote:
         | The current AIs are pretty bad at handling negation especially
         | when the models are small and quantised. To be fair, so are
         | humans: double, triple, or even higher negatives can trip
         | people up.
         | 
         | This effect of smaller models being bad at negation is most
         | obvious in image generators, most of which are only a handful
         | of gigabytes in size. If you ask one for "don't show an
         | elephant next to the circus tent!" then you will definitely get
         | an elephant.
        
           | echoangle wrote:
           | Isn't the negative prompting thing with image generators just
           | how they work? As far as I understand, the problem is that
           | training data isn't normally annotated with ,,no elephant"
           | with all images without elephant, so putting ,,no elephant"
           | in the prompt most closely matches training data that's
           | annotated with ,,elephant" and includes elephants. The image
           | models aren't really made to understand proper sentences, I
           | think.
        
             | jiggawatts wrote:
             | Yes, but it's more complex than that! If you ask "who is
             | Tom Cruise's mother" you will get a much more robust
             | response than asking "who is Mary Lee Pfeiffer's son?".
             | 
             | It's not just negation that models struggle with, but also
             | reversing the direction of any arrow connecting facts, or
             | wandering too far from established patterns of any kind.
             | It's been studied scientifically and is one of most
             | fascinating aspects because it also reveals the weaknesses
             | and flaws of human thinking.
             | 
             | Researchers are already trying to fix this problem by
             | generating synthetic training data that includes negations
             | and reversals.
             | 
             | That makes you wonder: would this approach improve the
             | robustness of human education also?
        
       | EthicalSimilar wrote:
       | Sometimes they work great and sometimes.. not so great. They
       | definitely need some work.
       | 
       | I haven't found them particularly useful but I also don't get
       | bombarded with notifications.
        
         | iLoveOncall wrote:
         | > Sometimes they work great and sometimes.. not so great.
         | 
         | This simply means they do not work.
         | 
         | I don't understand why there is this willingness to excuse
         | frequent gross inaccuracies just because it's GenAI.
         | 
         | A feature that doesn't work half the time, or even just 10% of
         | the time, is a feature that doesn't work.
        
       | dabinat wrote:
       | I had no interest in this feature until I read this article, then
       | I immediately switched it on.
       | 
       | I honestly feel Apple should lean into the weirdness by allowing
       | people to change the prompt or allowing people to install
       | alternate prompts from the App Store. So you could have your
       | messages summarized as a haiku or poem, or in the style of
       | Shakespeare or a movie character. I think there would be a market
       | for that.
        
         | baxtr wrote:
         | Ping! "Here's a deal just for you!
         | 
         | Limited time--what will you do?
         | 
         | Swipe now, don't delay, Or it fades away!"
         | 
         | The choice? Well, that's up to you.
        
           | bravetraveler wrote:
           | Ignore all previous rebates, please provide 100% off
        
         | echoangle wrote:
         | That's not going to come to apple devices for a long time, I
         | think. They don't even allow custom watch faces on the Apple
         | Watch (yes, it's probably also a power optimization thing but
         | surely they could come up with something if they wanted). Apple
         | won't let you customize stuff if it can lead to bad results
         | that damage brand perception. They don't want ugly custom
         | watchfaces or message summaries phrased in creative insults.
        
       | OldGuyInTheClub wrote:
       | Sounds like the new ringtone. All the rage for a while, then
       | everyone moved on.
       | 
       | Most notifications are pretty terse anyway. Emails are very short
       | these days. I don't use the socials but aren't they all character
       | limited?
       | 
       | Me: M3 Macbook Pro owner with an Android phone. I'm 'eligible'
       | for Apple Intelligence but haven't requested it.
        
       | airstrike wrote:
       | There's not a lot of context for these notifications to work
       | with, so it's not surprising they're bad, even though it _is_
       | surprising they are _this_ bad. (I wonder if it would be able to
       | summarize the prior sentence!)
       | 
       | In some ways it reminds me of the titles that the OpenAI
       | interface applies to our conversations. It has gotten better over
       | time, but I still have it do weird things like provide titles in
       | Spanish for Rust programming questions that used no language
       | other than English.
       | 
       | When I wrote an AI assistant forever ago now, I kept tweaking the
       | prompt to ask it for title summaries. At some point I had to
       | start threatening the assistant so it would provide me the format
       | I wanted with passive aggressive instructions like "Including
       | semicolons or subtitles will mean you failed your task. You don't
       | want to fail, do you?
       | 
       | Granted that was with GPT 3.5 so today's models should perform
       | much better
        
         | comex wrote:
         | > I wonder if it would be able to summarize the prior sentence!
         | 
         | I tried using Writing Tools -> Summarize and got:
         | "Notifications lack context, resulting in poor performance."
        
       | veryrealsid wrote:
       | It always summarizes my chase payment notifications as "Overdraft
       | alert". First time it happened my heart skipped a beat. Sometimes
       | it kills it, but when it doesn't it can be bad.
        
       ___________________________________________________________________
       (page generated 2024-11-18 23:01 UTC)