Posted by retskrad 10/25/2024
Smartphones are an absolute graveyard of fads; remember the 3D screen phones, the phones with projectors, and so forth? They generally go nowhere. I suspect 'AI' on phones will be similar.
Overwhelmingly, what people want out of phones is "like my current phone, but with better battery life and maybe a better camera." Previously 'faster' was also a concern, but modern phones are largely Good Enough.
Technically, I understand the difference in the technology, I just don't know who needs that vs who gets excited about new features for a brief moment.
I would love to take a photo using my smartphone that doesn't look pixelated, blurry and or over-processed. Maybe asking too much considering smartphone sensors can't compete with DSLRs in some situations, but I'm always baffled with how dark and desaturated some of my photos turn out on my smartphone, as well.
Now pendulum has swung so all go for 'realistic look', but I expect people actually want rather milder version of above.
Phone photos look OK on phones, but enlarged even the top contender from current dxomark show very much how hardware limits work. Its just not presentable, maybe apart from very bright scenes. Now I wouldn't go bashing phones per se, its marvel what they achieve from those tiny plastic lenses and some CPU time. And they are always there. But any low hanging fruit in phone photography was picked up long time ago by whole market, what lies ahead are slow computational improvements, coupled with very slow increase in size and thickness of camera section of phones to capture more light.
Even things like "[my dog name] beach" isn't reliable.
Or things like I use photos as notes. It doesn't reliably recall things like cheese when I take pictures of cheese in various stores to remember what is sold where. Not even the name of the cheese; just cheese. Ditto spices.
I've found google photos search to be pretty good, and if it can't find something usually the map-mode is enough to pin it down (e.g. go to the beach where you took the photo and it shows you the photos from there)
I did just check, and "dog at beach" generates sub 20% recall for me. I go to the beach weekly with my dog, take lots of photos because I'm a dork, and that first query skips many weeks.
Also, I did add my dog as a known / named pet under the explore tab, which is why I thought the name should work.
I can make it work by picking out the beach via geo, but I think the whole thing illustrates how much better this could be. I'd like to be able to get responses to queries like
* [pet name] at [beach X]
* [pet name] with sand on face
* dead seal, or even just dead animal (pics on beach)
* seaglass (recall is poor there too until I manually added to a photo album)
* dent in car
* [spice name] (I take pics of spices to know which stores offer what)
etc etc. The only way I manage the thousands of photos I have now is by carefully sorting into hundreds of albums, which google also doesn't support well.
Amongst the many many deficiencies of the app (which, tbf, does work extremely well as a read-through cache and seems to back things up very well), it likes to surface spotlights of dead people and pets. Which is not at all what I want proactively surfaced.
It was also hobbled quite hard near the start as they had a scare with searches for "gorilla" accidentally returning pictures of black people so they have probably turned all the safety knobs they have up to 11, even if that impacts recall.
It's not bad, it's really really bad.
The image clean-up feature is utterly useless, and I think it's one of the areas where we can clearly see the difference between a company for whom Generative AI is an afterthought (Apple), and the competition. I paid $1500 for the new iPhone Pro Max and a great part of the deal was the Apple Intelligence support, but frankly, I might as well switch to Android at this point because I'm really disappointed at Apple's take on AI. I'll probably wait until the official Apple Intelligence is introduced but tbh I don't think there will be much improvements over this version.
And as for Siri: It's as stupid as it ever was. I ask it to convert something from lbs to kg and it responds "there's no music playing". If anything, its natural language comprehension has degraded.
Currently, there's no context awareness, so I still can't ask Siri "how do I respond to this email?".
Really, the only thing that "works" is the ChatGPT feature that describes an image you send it. Anything Apple-related is bonkers. It's really embarrassing.
As far as the future of compute goes, Meta seems like it has a much more compelling argument with the Orion glasses and their investment in AI.
Yes, but Apple fans always respond by saying "...but Apple has been using ML in much of the OS for years, you just don't see it..."
I really want Apple to be better at this, but Generative AI is too uncontrollable for a control-freak company.
“Siri get me directions to McDonald’s” ‘I found this on the web for you:’ <serp for “Direct In Donald”>
Or it would get directions to a McDonalds in a different state. Or say “I’m sorry, I don’t see a McDonald in your contacts.”
All this AI marketing push has got to be because they think investors are stupid, and they can fool the market into thinking they're doing an AI.
That said, I think the tradeoffs being made right now are probably the right ones. Apple's latest devices have gone to an electrically released sort of adhesive (versus the older pull-strip removeable adhesive, which is a big step up from the "glue it in" approach many vendors take), and for a given volume, you get more battery if you can rely on the phone to protect it from damage - which is why almost everything with an internal battery uses some variety of pouch cell. They're quite a bit more fragile than the hard-cased batteries, but you get a lot more battery in the volume than you do with the hard cased ones.
As long as it's not incredibly irritating to replace the battery, I'm fine optimizing the daily use thing (battery life in a given phone size) over the once-every-few-years thing (replacing the battery).
Building a privacy aware cloud is a lot harder than you think it is. Apple is going the whole way with remote attestation, hardware root of trust, software BOM, etc.
But are you suggesting I will be able to flip a switch to use their cloud for all the same AI stuff instead of my battery-powered phone? That would be really welcome news! I haven’t used any of the betas so I know nothing about how “Apple Intelligence” works.
Yup! That's the idea. Compute negligible tasks run locally, but for requests that would hog battery and compute (or even be impossible to run on your iPhone), the data center will be used!
> though suspect that using their weird Mac hardware in the DC can’t possibly be a truly superior option to the NVIDIA hardware
So they are doing this for two reasons:
1. They get remote attestation and hardware root of trust when using their own hardware. No need to load proprietary blobs from Nvidia, introduce a new unvetted third-party supply chain, etc
2. Apple Silicon is actually much more cost effective than Nvidia's GPUs for inference. You can get a 48GB Apple Silicon Mac for like $4k and run large models - it's $40k for similar VRAM from Nvidia. I imagine Apple's costs are much lower than $4k as well.
The common response to this is to point out that I'm trusting Apple or Google anyway just by using their phones, which is true, but (1) since phones are directly accessible to security researchers, there's more potential to find out about malfeasance if data is exfiltrated from the phone in a way that it shouldn't be, and (2) even if I have complete trust in Apple or Google, I also have to trust that they'll be able to protect my data from malicious actors, and, the less data I give them, the less I have to worry when they fail.
Providers want to drip feed and charge for 'improvements' indefinitely.
AI is the next drip feed.