WWDC, June 10-14, 2024, to be 'A(bsolutely) I(ncredible)' — Keynote Discussion Here!

ZnU said:

Apple doesn't need "Siri, cure cancer" to work, but "Siri, arrange a two week vacation in France" ideally should, someday.

Click to expand...

The technology is more or less there for that already. I’d be shocked if an existing AI sitting on top of a travel booking database couldn’t do most of that today. The question is how to incrementally get people to a point where they trust the output to slam down a few grand on their credit card. Nothing about that request screams to me ‘now THAT’S going to take more processing power than we have today!’

ZnU said:

Well, If everything up through AGI turns out to be cheap, we'll be talking more about Dyson swarms than datacenters.

Click to expand...

The amount of hardware needed is relative to how often you need that level of intelligence. As on device hardware gets ever more capable they’ll need to go out to the cloud for fewer and fewer things. So even if the state of the art hardware requirements were to increase indefinitely the proportion of requests that need that state of the art would go down, lowering server requirements.
——
It’s an interesting thought experiment what you would do with AGI. If AGI were built into a robot that could do dishes and fold my clothes I’d be all over that shit. But the abstract equivalent of a human executive assistant living in my phone. I’m not sure just how much I would need it to do. How often do I book trips to France?

wrylachlan said:

The technology is more or less there for that already. I’d be shocked if an existing AI sitting on top of a travel booking database couldn’t do most of that today. The question is how to incrementally get people to a point where they trust the output to slam down a few grand on their credit card. Nothing about that request screams to me ‘now THAT’S going to take more processing power than we have today!’


The amount of hardware needed is relative to how often you need that level of intelligence. As on device hardware gets ever more capable they’ll need to go out to the cloud for fewer and fewer things. So even if the state of the art hardware requirements were to increase indefinitely the proportion of requests that need that state of the art would go down, lowering server requirements.

Click to expand...

See the travel-booking scenario is exactly one I'd imagine a server being best at. Not for raw compute, but the ability to rapidly query dozens of airlines/hotels/tour services/etc. This just isn't task one would deem 'mobile/edge optimal'.

Bonusround said:

See the travel-booking scenario is exactly one I'd imagine a server being best at. Not for raw compute, but the ability to rapidly query dozens of airlines/hotels/tour services/etc. This just isn't task one would deem 'mobile/edge optimal'.

Click to expand...

Aggregating queries of multiple airlines/tours/etc. is already done by things like Expedia. Where intelligence is needed is understanding out of the results which ones the user would prefer. But my point wasn’t that this is doable in device, but rather that it doesn’t require a leap forward towards AGI to get there. You could almost certainly train an LLM to take multiple pages of exploratory queries on Expedia + access to the new semantic index and turn it into a coherent itinerary today. This isn’t the type of challenge that’s going to require a ChatGPT 8 running on a quantum computer. It’s not suggestive that Apple will need to keep increasing their investment in Private Cloud to keep up with the Joneses.

Replace multiple airlines/tours with multiple airline/tour services (which aren't already owned by Expedia, at least!), or multiple restaurants for reservations, multiple retailers for optimal pricing/availability, etc. etc. etc. My humble point that a sever-executed task might sometimes be best stands, as does your (completely independent) point about runaway LLM needs or the lack thereof.

wrylachlan said:

I think this is all largely true. But it doesn’t change the fact that scraping the web for training before telling the world they were doing it isn’t a good look relative to their “we’re the good guys” marketing stance. Even if they’re taking extreme precaution it’s not a good look.

Click to expand...

I won't argue with that. But as this social problem develops, I think we'll find that scraping the web to learn basic language structure vs scraping the web to pass my recipe for egg salad off as your recipe for egg salad land in the 'acceptable' vs 'unacceptble' categories. We're still in the hair on fire everything is terrible phase of this. Nuance is not available at this time.

johnsonwax said:

I won't argue with that. But as this social problem develops, I think we'll find that scraping the web to learn basic language structure vs scraping the web to pass my recipe for egg salad off as your recipe for egg salad land in the 'acceptable' vs 'unacceptble' categories. We're still in the hair on fire everything is terrible phase of this. Nuance is not available at this time.

Click to expand...

Agreed. I don’t have a problem with Apple scraping the web for training on language. I do question whether doing it without telling people first is smart marketing given the current climate.

Scud said:

Don't those 200 million devices also use datacenters for private cloud compute?

Click to expand...

They can. Apple's hope is that they won't need to. This isn't that hard to understand.

Apple launched Satellite SOS on about 500 million devices - for free. If all 500 million people used that, you could probably see the Hollywood data overload supernovaing Globalstars constellation during daytime. Apple launched that service with the expectation that, for instance, 0.1% of all users would use it one time per year. So about 1000 people per day. Something like that. If they guessed wrong, well, shit would hit a fan and they'd deal with it. Meanwhile, 500 million people walk around with the comfort they have a feature, which effectively none of them ever use.

So Apple's 200 million users capable of running AI™ (plus another 300 million probably being added in the next year), the feature is only available in the country of about 20% of them. We're down to 100 million. Of those, probably half never use the feature in any active way. Apple knows how many people have Siri completely turned off. Down to 50 million. Of those 80% (Pareto meme number) never use features that require more compute than is on the device. Down to 10 million. Of those, 10% per day do something intensive enough to need cloud compute. So, a million requests per day. Taking your other numbers for transaction time, etc. yeah, 5 digits worth of machines seems about right.

By next year, the amount of on-board compute will go up, likely faster than the demand for that compute does. That shifts more users off of the servers as things remain on-board. Apple expands into new countries. Some of the older compute is being replaced with newer reducing the old demand further, server demand goes up, but proportionately less than the number of new customers able to use it. Apple can add a smaller 5 digits worth of hardware but they also now have M4 ultra which can chew through requests even faster so a smaller number yet. With each successive year the pattern repeats until Apple no longer needs to add server compute, and then demand falls - exactly as it did with Siri - and Apple can start retiring servers. And it's entirely possible that AI™ never launches in mainland China - about 20% of Apple's user base due to issues of political control in the country. They never need to build server compute for them.

wrylachlan said:

Agreed. I don’t have a problem with Apple scraping the web for training on language. I do question whether doing it without telling people first is smart marketing given the current climate.

Click to expand...

I don't think this has really seeped into the general public. The backlash is primarily among tech people and certain parts of society that are attuned to intellectual property (like academics). I don't think the general public has any expectation that everything they post on the internet isn't already being churned into a data souplike homogenate. They came to terms with that back when they created their Facebook account.

When you see Gemini spit out an almost word-for-word explanation for how to keep your cheese on pizza that someone else wrote, it's very clear that information exists almost verbatim inside that model, and they don't like that. I suspect, just based on what we've seen and the nature of the interface that simply isn't possible in AI™, and that will be apparent in time. A lot of this kind of stuff only connects to the public after a few years of not seeing someone in the news. Can you trust Apple will your credit card? Yeah, they've never made a headline on that.

It's a little risky now, but Apple still plays the long game. I think they expect they'll be vindicated over the longer arc.

Bonusround said:

See the travel-booking scenario is exactly one I'd imagine a server being best at. Not for raw compute, but the ability to rapidly query dozens of airlines/hotels/tour services/etc. This just isn't task one would deem 'mobile/edge optimal'.

Click to expand...

It depends. If you are making a generalized request then yes, server would be best. But Apple's bargain here is that knowing where you are, or what your schedule is, and so on will be important factors in how people do this, and that brings at least part of the exercise on-device. Travel booking isn't really that complex because people usually have fixed constraints - when they are traveling and from where and to where. The variables in-between get surprisingly small surprisingly fast. That data set is quite small.

Plus, if this is done, as is most likely in the travel app, with intents to interact with Siri and on the back end the ability to distribute compute between device and say, Expedias servers, that seems ideal.

The problem with almost all server-side services is that the thing the user is interacting with is the phone in their pocket, which is why so much of the web has gone the way of apps - getting as close to the user is the thing that matters, and that's what AI™ leans into by setting user expectations for that interaction to be even closer to the user.

stevenkan said:

Holy shit. No kidding!!!!

Click to expand...

LOL. That was so fucking long ago. I couldn't remember who it was that happened to. Nice to know you're still talking to me.

Not the first time I guessed someone's lunch order from a bit of metadata. Did it to some poor fucker on usenet ages ago when we were discussing NSA metadata harvesting of phone records. Freaked them the fuck out. I think I got their general location from an earlier post of theirs.

I think LAs local sales tax stuff narrowed the location down with you - somebody had a half cent sales tax that was necessary to get the total, so there were only 2-3 locations that could have produced that receipt total, and only 1 menu combination that could.

You know, I worked at a university not in North Korea, now retired. I was neither stalking your or worked at a CC company. But that was just fun. I was a bit of a data mercenary for my university that way - people would throw all kinds of weird problems like that at me.

johnsonwax said:

I won't argue with that. But as this social problem develops, I think we'll find that scraping the web to learn basic language structure vs scraping the web to pass my recipe for egg salad off as your recipe for egg salad land in the 'acceptable' vs 'unacceptble' categories. We're still in the hair on fire everything is terrible phase of this. Nuance is not available at this time.

Click to expand...

The catch is that those two things are not separable in the LLMs that are currently trendy-- they're essentially fancy next-word predictors with interesting emergent properties (because of the size of the network built during training and their ability to look back at previous context), so "training for basic language structure" and "training for knowledge" are not two separate activities that can be done independently. Anything the model takes in during its training is something it might regurgitate later.

If you want to be able to do some kind of dragnet all-the-things training for basic language structure, have the model somehow "forget" any knowledge it acquired during that stage, and separately train on licensed material for its actual knowledge base, you're looking at a fundamentally different type of model than what OpenAI or Google or Meta are doing today.

Jonathon said:

The catch is that those two things are not separable in the LLMs that are currently trendy-- they're essentially fancy next-word predictors with interesting emergent properties (because of the size of the network built during training and their ability to look back at previous context), so "training for basic language structure" and "training for knowledge" are not two separate activities that can be done independently. Anything the model takes in during its training is something it might regurgitate later.

If you want to be able to do some kind of dragnet all-the-things training for basic language structure, have the model somehow "forget" any knowledge it acquired during that stage, and separately train on licensed material for its actual knowledge base, you're looking at a fundamentally different type of model than what OpenAI or Google or Meta are doing today.

Click to expand...

Agreed. I don't know how much of it is a fundamental difference in the model or a fundamental difference in the nature of the data being collected. It seems to me the latter might be the easier problem to solve. We can speculate (based on papers Apple has published) that they have worked out a process for global training which is then continued on-device with local information. That alone is a bit unique. But removing information is effectively impossible given the nature of what these things are. If you can, I'd love to read a paper on that.

So my guess here has been that they are unusually judicious with what they feed in. Apple's model needs general information - it needs to know what 'car' contextual means, it doesn't need to know the curb weight of the 2024 Ford F-150 Lightning. The upside is that if Apple is just building a language structure and basic knowledge, that's not a dynamic dataset. Once you build it, maintaining it is probably pretty easy - language and general knowledge doesn't move very fast. But that's not what OpenAI does - ChatGPT is expected to know the curb weight of the F150. It's a perpetual and ever expanding dataset that requires constant retraining to be useful. So pointing AI™ at Reddit is just a nightmare because it's just going to pollute the model with a bunch of shit they can't get out. Pointing it at a publication date ranged Project Gutenberg is probably a lot more suitable.

I've been trying to work out how you would hook such a system into video games, as people have tried to do with Skyrim, and I think you have to have a bespoke model that is completely ignorant of anything published after a given time, otherwise you can strike up a conversation with the Whiterun guard about which iPhone you should buy. Maybe there could be a medieval knowledge trained model that is shared across properties, which is then fed local information about the information in the game as it becomes available, available to that character, and so on. It's a remarkably difficult problem to solve. If you want to not leak copyrighted materials through the model, you simply can't train on identifiable copyright materials. There was a point in time where the phrase 'catch-22' only existed in a single work of art, but it's now a common idiom. There's a point where training on that information reveals a copyright risk, and there's a point when it no longer does. Navigating that kind of space is difficult, and OpenAI and Google are apparently making no effort to try. I suspect Apple is trying here just based on the great many things that it lacks an interface for.

Jonathon said:

"training for basic language structure" and "training for knowledge" are not two separate activities that can be done independently. Anything the model takes in during its training is something it might regurgitate later.

Click to expand...

I can think of four aspects of this system that almost certainly mitigate this effect:

1. Preprocessing - just because Apple’s web crawler picks up some text doesn’t mean they’re going to use it. Choosing what goes into the model is a huge part of the process and absent evidence to the contrary my assumption would be that Apple would be their normal thorough self at this step.
2. Reinforcement learning - undoubtedly reinforcement learning was a huge part of this process. And when doing reinforcement learning you get to choose what to ‘reward’ the model for. ChatGPT rewards its model for showing off the breadth of its knowledge. I would assume based on the intended use case that Apple would penalize, not reward, the model for conspicuously demonstrating ‘world knowledge’.
3. The Orchestrator - we’re talking about ‘the’ model, but based on the presentation it’s actually 3 components - a core model, a shim-like adaptor that specializes the model for a given task, and the Orchestrator (itself a model) that decides which shim is best for a given prompt and whether a prompt should instead go to ChatGPT. Even if Apple’s core model was prone to spitting out hallucinatory world knowledge, you’d never know because the Orchestrator doesn’t send those types of prompts to Apple’s core model.
4. Size - last but not least, this is a relatively small model. It can encode language, grammar and rudimentary concepts. But it doesn’t have enough nodes to encode very much world knowledge. For that reason alone it’s never going to spit out your grandmas pie recipe you put on your blog last year.

wrylachlan said:

I can think of four aspects of this system that almost certainly mitigate this effect:

1. Preprocessing - just because Apple’s web crawler picks up some text doesn’t mean they’re going to use it. Choosing what goes into the model is a huge part of the process and absent evidence to the contrary my assumption would be that Apple would be their normal thorough self at this step.
2. Reinforcement learning - undoubtedly reinforcement learning was a huge part of this process. And when doing reinforcement learning you get to choose what to ‘reward’ the model for. ChatGPT rewards its model for showing off the breadth of its knowledge. I would assume based on the intended use case that Apple would penalize, not reward, the model for conspicuously demonstrating ‘world knowledge’.
3. The Orchestrator - we’re talking about ‘the’ model, but based on the presentation it’s actually 3 components - a core model, a shim-like adaptor that specializes the model for a given task, and the Orchestrator (itself a model) that decides which shim is best for a given prompt and whether a prompt should instead go to ChatGPT. Even if Apple’s core model was prone to spitting out hallucinatory world knowledge, you’d never know because the Orchestrator doesn’t send those types of prompts to Apple’s core model.
4. Size - last but not least, this is a relatively small model. It can encode language, grammar and rudimentary concepts. But it doesn’t have enough nodes to encode very much world knowledge. For that reason alone it’s never going to spit out your grandmas pie recipe you put on your blog last year.

Click to expand...

Nods to 1-3.

But for 4, there are different models on server than device, and the server model Apple compares to GPT-4 Turbo, so it's performant to a big model, but I contend that's just because they've done more of 1-3, not because they've jammed in grandmas pie recipe.

johnsonwax said:

Nods to 1-3.

But for 4, there are different models on server than device, and the server model Apple compares to GPT-4 Turbo, so it's performant to a big model, but I contend that's just because they've done more of 1-3, not because they've jammed in grandmas pie recipe.

Click to expand...

Very true. Of course it’s not clear what the benefit of a model that large is if not to show off world knowledge. Maybe the size helps it deal with larger bodies of text for summation???

wrylachlan said:

The technology is more or less there for that already. I’d be shocked if an existing AI sitting on top of a travel booking database couldn’t do most of that today. The question is how to incrementally get people to a point where they trust the output to slam down a few grand on their credit card. Nothing about that request screams to me ‘now THAT’S going to take more processing power than we have today!’

Click to expand...

Today's LLMs aren't reliable enough to complete non-trivial multi-step tasks to any reasonable standard. They still make reasoning errors that an average human wouldn't, and nobody has entirely solved hallucination. Despite following interesting papers in the field pretty closely, I've honestly got no idea if this will be fixed in six months or six years, and no idea if what's needed is some clever tweak to transformers, 100× more scale, or some fundamentally different approach (with its own unguessable compute requirements).

Note that what Apple showed off this week still only works with a fixed (if expanded) set of intents. In theory LLMs are capable of directing open-ended action, so why did Apple impose this limitation? Almost certainly because, as anyone who has tried to hook an LLM up to external tools can tell you, they're not great at this right now. Restricting things in this way lets Apple keep quality high by testing and fine-tuning against specific tasks. But in the end, they'll want something fully open-domain, to which apps can expose any set of actions or which can simply look at apps or web pages and reason out how to interact with them like humans do.

wrylachlan said:

The amount of hardware needed is relative to how often you need that level of intelligence. As on device hardware gets ever more capable they’ll need to go out to the cloud for fewer and fewer things. So even if the state of the art hardware requirements were to increase indefinitely the proportion of requests that need that state of the art would go down, lowering server requirements.

Click to expand...

The assistant role requires extensive use of common sense reasoning. This has proven to be an extremely difficult problem for machines. We're finally making some progress on it (after ~70 years of failure), but the models that are best at it are the largest that exist. And, as per above, even those aren't really good enough yet.

wrylachlan said:

It’s an interesting thought experiment what you would do with AGI. If AGI were built into a robot that could do dishes and fold my clothes I’d be all over that shit. But the abstract equivalent of a human executive assistant living in my phone. I’m not sure just how much I would need it to do. How often do I book trips to France?

Click to expand...

Your daily life is probably more complicated than a two week vacation, really. It's routine enough that you don't think of it like this, but there's a good chance it involves more interlocking parts and overlapping goals, and of course it's not clearly time-bounded. Over the course of a couple of weeks you might go to dozens of locations, purchase hundreds of items (counting every grocery item), engage with dozens of businesses and people in various capacities, check up on household accounts, pay bills, try to find interesting new TV shows or restaurants or whatever. AI could mediate all of this. Especially if every individual and business had one, if every digital system could be accessed through one, huge amounts of friction could be removed from daily life by having AI agents handle many details between themselves.

ZnU said:

The assistant role requires extensive use of common sense reasoning. This has proven to be an extremely difficult problem for machines. We're finally making some progress on it (after ~70 years of failure), but the models that are best at it are the largest that exist. And, as per above, even those aren't really good enough yet.

Click to expand...

Regarding ‘the models “best at it are the largest that exist”, can you please define ‘it’? ISTM that domain-focussed models need not be gigantic at all and could, one presumes, achieve much greater success when reasoning within the confines of that domain, no?

ZnU said:

Your daily life is probably more complicated than a two week vacation, really. It's routine enough that you don't think of it like this, but there's a good chance it involves more interlocking parts and overlapping goals, and of course it's not clearly time-bounded. Over the course of a couple of weeks you might go to dozens of locations, purchase hundreds of items (counting every grocery item), engage with dozens of businesses and people in various capacities, check up on household accounts, pay bills, try to find interesting new TV shows or restaurants or whatever. AI could mediate all of this. Especially if every individual and business had one, if every digital system could be accessed through one, huge amounts of friction could be removed from daily life by having AI agents handle many details between themselves.

Click to expand...

This sounds like froth to me. My bills are all on auto-pay which requires zero AI. I have a network of friends who give me restaurant suggestions just fine. I enjoy grocery shopping since I like to cook, but I suppose if I squint real hard I could imagine a personal assistant grocery shopping for me. But there the issue is the delivery, not the choosing what to buy which is already pretty dialed in.

I would love a personal finance assistant to review my bills, but there the problem to be solved isn’t the reasoning about or understanding them (I can do that just fine) it’s getting all the itemized receipts into structured data or better yet receiving them as structured data from the vendor which has a whole ecosystem component that moves glacially.

Can you give an example of:

• A specific interaction in my day to day life
• That would be meaningfully lower friction with an AI agent
• AND the limiting factor is currently the AI intelligence not other system issues (like getting receipts into structured data)

ZnU said:

AI could mediate all of this. Especially if every individual and business had one, if every digital system could be accessed through one, huge amounts of friction could be removed from daily life by having AI agents handle many details between themselves.

Click to expand...

I've seen where this ends up going...

nimro said:

From Apple's page on it's model training:


So, they never use their users' private personal data... unless their web crawler hoovers it up. Same scummy opt-out data vacuuming that every other shitty corporate AI play is relying on.

There goes all hope of I had Apple doing it "better". It's the same trend-chasing consent-disregarding crap everyone else is doing they're just late to the party.

Click to expand...

How does their web crawler hoover up private personal data?

eas said:

How does their web crawler hoover up private personal data?

Click to expand...

That's kind of where my head is at too. Do people have an expectation of privacy for things that are on the open crawlable web? I've always assumed that stuff might as well be on a billboard as far as my keeping it "private" goes...

Listening to the Talk Show now, some really good pieces about the AI stuff. Well worth your time, either podcast or YouTube.

kenada said:

Do you have any examples? I have 1Password configured as the only credential provider in iOS, but I only ever get prompts to fill OTPs for sites I’ve set up in Apple’s password manager. The only way I can fill OTPs from 1Password is by manually copying them from the app or by using 1Password’s Safari extension. If I’m understanding correctly, 1Password should be able to fill OTPs in iOS 18 using the native UI once it’s updated to implement the new API (and also in macOS 15 should 1Password ever implement it, though they’ve been pretty resistant supporting the extension API so far even though it makes the Passkey experience worse on macOS).

Click to expand...

Ubiquiti - 1Password enters the OTP automatically, yes, even on Mac.

1Password is annoying me now, it keeps superceded FaceID/TouchID with passkeys. When I can use FaceID or TouchID - I would much rather, less friction.

johnsonwax said:

I mean, Apple has approximately 1 billion TPUs in the wild right now, and about 200 million that Apple Intelligence will run on. That's very clearly the 'datacenter' they expect to be operating on.

There's also the matter of rapidly diminishing returns in terms of performance. I didn't see anything in Apple's presentation to suggest that the door is open to extremely massive context windows (even though they do compare their server models to GPT-4 Turbo), to doing generative efforts outside of rather narrow scopes, and so on, suggesting that these models are more narrowly tailored to the use cases that Apple provides an interface for. It can summarize an article, but it can't write one. As such, my guess is that AI™ will overperform on-device compared to other models because its utility is narrow and known and controlled by Apple, and that many of the cases where Google and OpenAI need these big data centers (in part because they only operate as a server model) Apple doesn't because there's no interface to do those kinds of things. It's not like AI™ gives you an open prompt to do whatever, so my guess is that the conversational Siri stuff never leaves the device. Some of the larger text processing stuff, particularly on larger works might go to server, and some generative image stuff might, but not like the emoji stuff - the high volume things, summarizing an email chain, etc. Apple's pretty darn good at hitting that Pareto optimal level, even when that leaves them trailing their competition simply due to their scale. It's a recurring exercise for someone to hack Apple features like this onto older hardware, because Apple was conservative in terms of setting user expectations. No reason to believe things will change here.

Plus, the vast majority of users won't actively use any of this stuff. Adoption curve dynamics apply here as well.

Click to expand...

This may be why they are not extending their AI to older devices by doing that processing on server-side. That are trying to do as much processing as they can on-device with server-side as a fall back. Opening up to less powerful devices would significantly shit the processing demand to their servers and that might easily exceed their near term capacity.

fixed:typo

Hap said:

Ubiquiti - 1Password enters the OTP automatically, yes, even on Mac.

Click to expand...

How does that work? 1Password doesn’t provide a Password Provider extension on macOS. Is it using universal autofill?

wco81 said:

I don’t get why iPhone mirroring is such a big deal for many.

The scenario that they described was that your iPhone is in another room.

Is it that onerous to go get it? Do people here live in mansions or something, it would be a trek to go get it?

The one device which is most-likely to be always on hand would be your phone?

Click to expand...

I spend half my Teams meetings with my iPhone resting on my MacBook Pro screen like it’s a phone stand. I will use the hell out of this feature.

adherent said:

I spend half my Teams meetings with my iPhone resting on my MacBook Pro screen like it’s a phone stand. I will use the hell out of this feature.

Click to expand...

While at my stationary workplace, my phone is my webcam. Will help me respond to messages and stuff while vid.-conferencing

ham bulu said:

While at my stationary workplace, my phone is my webcam. Will help me respond to messages and stuff while vid.-conferencing

Click to expand...

That was mentioned on the Talk Show. Craig uses his iPhone for continuity camera. So this is sort of a natural evolution. Release feature -> dogfood feature -> realize it’s a great feature but creates a secondary problem -> release feature that solves secondary problem.

kenada said:

How does that work? 1Password doesn’t provide a Password Provider extension on macOS. Is it using universal autofill?

Click to expand...

I honestly have no idea, just once I select the appropriate item from the 1Password popup in the username box, it auto enters user name password, the page changes to allow OTP entry and it's already filled by the time the page loads and it proceeds into my account. I don't have to do anything.

Hap said:

I honestly have no idea, just once I select the appropriate item from the 1Password popup in the username box, it auto enters user name password, the page changes to allow OTP entry and it's already filled by the time the page loads and it proceeds into my account. I don't have to do anything.

Click to expand...

This has been my experience as well. I can’t think of any site where 1P couldn’t fill a OTP code (and it copies it to the clipboard as a backup).

So I didn't really pay attention to how they might use AI to have Photos app with memories or putting together some kind of presentation from your photo library as Apple Intelligence knows about you.

But I just created an album of a trip which consisted of photos from my iPhone and my Nikon, panos, videos and time lapse videos. I played them back through the Photos app on my iPad Air and my 14 inch MBP.

They have a slide show theme called Magazine which presents the photos as tiles. It appears to randomize the number of tiles they show, so they show anywhere from 1 full screen tile to maybe up to 6 or 8 tiles.

The tiles could have videos or panos, so if a video, it plays the video while the static photos stay in the other tiles. If it's a pano, it scrolls it left to right -- maybe some they scroll top to bottom though I didn't have any vertical panos.

What was interesting was the grouping of the multi-tiled photos. It was often variation of the same subjects, which are all landscapes. So for instance, one screen might have 5 tiles of a lighthouse that I took from different angles, different zoom levels or even different focus points.

I think it's pretty much going by the order in which I added the photos to the album. I would say it's using the time-stamp of the photos but I export them from RAW in LRC to JPGs (in this case JPGs with HDR metadata) before uploading them as shared albums in iCloud. That is how I access them on my iPad for instance, through Photos app shared albums. My guess is that use the original photo metadata because they're grouping photos with videos captured around the same date and time together so the JPG export file creation/modification dates, which would be months later than capture dates, wouldn't come into play.

But there are such possibilities. It could use the GPS location data to group say several photos taken around the same locations together. Or AI could do some rudimentary image recognition and just see the general same shapes to determine that they are variations on the same subject.

I don't know if they ever highlighted the Slideshow feature in some keynote but they have some good ideas and they could use the AI hype to make some more "intelligent" presentations.