A look at Apple's technical approach to AI including core model performance etc.
interconnects.ai
interconnects.ai
>This looks like 2 years after the release of GPT-4, Apple has approximately trained an original GPT-4 level model.
Are there any more details on what exactly is being sent as context to the cloud? Do they send features extracted from an image on device or the full picture? Is it capable of only sending current pic or will on-device model select what it thinks is needed for context and many pictures and/or texts will be sent?
Overly freezing the context guarantees at this early stage would be highly limiting.
It makes more sense to invest in surrounding privacy infrastructure (PCC) to provide sane guarantees without strictly bounding context.
https://fleetwood.dev/posts/a-case-for-client-side-machine-l...
Some predictions in the article:
1. Partial client side execution of the models, then main model execution in cloud. (Example: text encoder for Stable Diffusion runs local, UNet runs cloud)
2. Local LoRAs for “dynamic ML”
3. WebGPU will be new standard to unlock models in the web
Even though the article had some good calculations and overview, I am not sure word-for-word is how I would describe its relationship to the OP.
This basically says that only a tiny bit of the AI will done on NVDIA hardware for the billions of Apple users.
The market is not pricing that in.
Inference is on Apples own silicon though.
https://www.reuters.com/technology/artificial-intelligence/h...
That’s what the market is pricing into the Nvidia stock.
It is very unlikely that Apple will sell their “server” hardware to other businesses in that way. Their advantage is the vertical integration, for themselves.
While I am bullish on AAPL, I don’t expect it to have a run like Nvidia.
When the market is saturated with shovel-making machines, what then?
I expect at some point a) the training and the inference will move to even more specialised hardware and the current existing silicon go to either hobbyists or some other market that requires the current cards b) the way in which we do AI may fundamentally change in ways we can't predict, causing the required hardware to also change (move to FPGA or favouring some other aspect than VRAM)
In both scenarios NVIDIA isn't the one who win, because the sudden influx lifts a secondary market nv don't profit from directly.
I personally don't think they'll be able to sustain their current valuation past the current AI rush, and will compact back down to somewhere their levels pre-chatgpt.
That said I'm just a guy on the Internet who wouldn't mind some beefy ex-datacentre gear if I could make it work for video editing workloads.
The official specs [0] say the A100 80GB has a 300W TDP. That's pretty doable on standard domestic power... harder if you want to run multiples.
[0] https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Cent...
https://www.ebay.com/itm/335392812272?nordt=true
https://www.ebay.com/itm/335408349837?nordt=true
https://www.ebay.com/itm/335414821299?nordt=true
Too good to be true? Perhaps. For $2000 I might trust Ebay's buyer protections...
(1) MLIR, ROCm, or some other tooling that reduces CUDA's moat,
(2) AMD attracting investment to go after NVIDA
(3) ARM-based GPUs or accelerators gaining traction among cloud companies that have huge fleets of AI devices and also have the money to devote to custom chips
Can anyone who understands the industry explain why those threats (or similar ones) aren't a major issue for NVIDIA?
(1) Much development is already moving from CUDA to the LLM, so less of an issue. Nvidia is also doing more work to increase interoperability. Could be an issue I guess, but doesn't seem like it since there's nothing close to CUDA or the ecosystem.
(2) AMD has attracted significant investment looking at appreciation in its market cap, with a PE ration 3X Nvidia's. However, AMD is so far behind in so many ways, I don't believe it is an investment problem, but structural. Nvidia has just been preparing for this for so long it has a temendous head start, not to mention being more focused on this. Remember AMD also competes with Intel, etc.
(3) Hyperscalers already are building their own chips. It seems even Apple used its own chips for Apple Intelligence. It's relatively (which is doing a lot of lifting in this sentence because it's all HARD) not too hard to make custom chips for AI. The hard (near impossible) thing is making the cutting edge chips. And the cutting edge chips are what the OpenAIs of the world demand for training, but releasing the newest best model 1-3 months ahead of a competitors is worth so much.
If anything, I'd say the biggest threat to Nvidia in the next 1-3 years is an issue with TSMC or some new paradigm that makes Nvidia's approach suboptimal.
I don't think I understand your point in (1) that it's less of an issue because development is moving to the LLM. I can infer that maybe CUDA isn't a big part of the moat given your other points that the hard part is making cutting edge chips.
But since there are plenty of more advanced models now, the next level is getting built out as more developers start building applications that use the models (e.g. apps using GPT's API).
So where 5 years ago most AI dev was on CUDA, now most is on the LLMs that were built with CUDA to build applications.
Obviously there’s nothing of the sort in the works now, but it’s not totally implausible to think that the expertise and infrastructure apple is developing now might be setting themselves up to also handle inference for things like home automation systems, industrial sensors and actuators, etc which may all need edge inference capabilities. Chips like the M Series or on the other side the NPU will presumably be coming to many devices outside of the mobile/pc market and it doesn’t seem totally insane to me to think that Apple might want to get into that world.
Again, this is all just speculation about long term pathways that might be unlikely after all.
This will undoubtably prove to be temporary.
>It is very unlikely that Apple will sell their “server” hardware to other businesses in that way.
The point is Nvidia isn't required.
Training is still a different story, imo.
They kind of lucked out with apple silicon and their unified memory architecture.
You scared me. I know the market is hyping up Nvidia hard, but if it was really worth more than Apple then things would really be crazy.
AAPL market cap is still 10x NVDA’s (3.2T vs 330B), so not yet too crazy.
Update: Nvda is at 3.24T. ddg “nvda market cap” shows the wrong amount. It probably doesn’t take the stock split into account. Djeez. That’s. Well, time to short.
Apple silicon isn't magic.
Nvidia's stock is surging because they sell GPUs and big companies by a lot of their GPUs to train AI models.
Apple's stock has not surged following the AI boom because they... don't sell GPUs used to train AI models.
AI is a profit center for Nvidia and a cost center for Apple (yes, people that weren't buying iPhone won't suddenly buy an iPhone because it has an AI assistant embedded).
There's absolutely no link between the two.
> There's absolutely no link between the two.
These are remarkable levels of speculation. Already I've heard from Android friends who are intrigued by the recent AI announcements that they think it would be really cool to have that in their phones. I don't know if they will necessarily switch, but they certainly indicated they were very interested in the iPhone AI announcements.
That is of course without mentioning that the only two devices compatible right now (both versions of the iPhone 15 Pro) represent a total market share of 3.1% of smartphones.
AI in iPhones is essentially a distraction from the fact that Apple hasn't brought any software innovation to its devices in probably close to a decade (note I say this as someone who has only had iPhones since the 4).
But if you were switching platforms to iOS, you’d be getting a new phone, wouldn’t you?
Most of it has been there for Samsung already and some of them like the emoji generator look suspiciously similar.
On Android, usually some large manufacturer (here Samsung) push the platform further and then some of it is integrated in stock later.
I still know more people who don’t use ChatGPT on a semi-regular basis than people who do. Most people’s phone choices come down to:
1) is it available in a color I like
2) is the camera any good
3) can I afford it
4) *optional* does my trusted techy family member recommend it
Want to point out Google photos has been able to do this with a voice command for close to(?) ten years. Just tried it with "[person name] in a dress"
Fancy marketing tends to do that. Have they tried it?
Especially considering how Apple's Neural Engine is borderline-pointless for LLMs, neither Android or iOS feel particularly well-poised to deal with this trend. If anything I lean more towards Android, since Android manufacturers have been shipping 8GB of RAM in phones for years, and support the Vulkan accelerators that a lot of inference apps feature.
That's a big "if". Right now it's marketing.
>...I think you will see people moving platforms.
This presumes Google can't do something similar, and over-estimates how much normies care about these things.
I'm a "techie" and personally couldn't care less about having AI get me to the airport on time.
The benefit Apple has over Android (not Google) is that they're fully integrated, ie they make so much profit from their closed platform that they can toss customers bones like iMessage and "apple intelligence".
Remember, attention is all you need came from Google in the first place...
But in the long term, it totally could be a game changer.
Think of all the AI startups paying OpenAI per token for consumer app capabilities. Or having to rely on suboptimal local foundation models. The value Apple can offer here is considerable.
I’d expect Google is headed towards a similar architecture, and both platforms will include API monetization for third parties.
I wonder if Apple’s deal with OpenAI precludes the latter from attempting to build their own hybrid platform on Apple devices…
Now compare to Nvidia, their stock has been on an AI related run because their revenues have been MASSIVELY increasing over the past year, because they sell the most advanced chips for training and inferencing the most complex models.
Apple may be “doing AI” now, but it’s for consumer products, they’re not selling enterprise AI services nor are they selling the actual chips needed to train and inference.
Also, I have major doubts as to whether any of the Apple AI announcements are going to drive higher iPhone sales. There will be tech evangelists who buy a new iPhone for the AI features but the bulk of users are folk like my parents, who use their phone for calling, messaging, and taking photos. The only thing they ever noticed with new phones was either it was:
1: better camera
2: better/bigger screen
3: really obviously faster
The last few phones have been indistinguishable in terms of screen, performance, or camera, and my own personal iPhone cycle went from every year or every other year, to now roughly every 3/4 years (and I’m an engineer).
I do think that Siri’s Generative AI upgrade (which is how I think of this whole Apple Intelligence thing) might be significant enough to change the smart phone as a segment. There are so many things that I feel Siri should be able to do, but it just can’t. Maybe this unlocks those use cases.
If it does, that’s a major incentive for me to increase my investment/reliance on the Apple ecosystem. I’ll use the calendar and mail apps if it means Siri has all that context. I’ll upgrade devices to get the same answers with less perceptual lag.
If we’re focused on hardware sales, I’d be interested in an Apple version of the Rabbit R1. That thing is almost a modernized iPod Touch - great device for kids you don’t trust with an unsupervised connection to the mobile internet. I would be delighted to get a Real One from Apple.
The HomePod Mini is close to this already. It’s. Not clear to me how much Apple Intelligence will be available through those. Again, if it actually works I will buy more of them.
Why? Most play with generative AIs for a few days and then abandon them.
The actual gamechanger would be making an AI assistant to be, you know, an actual assistant maybe?
I think OS integration can provide both of those things. Mobile app selection is already pretty strong, and the Shortcuts ecosystem is a good example of providing direct access to specific functions implemented by apps. AI should use the shortcuts functions exposed by the apps I’ve already selected, installed and configured as tools.
If I can use shortcuts on demand with just my voice… that’s a huge win for me.
The competitors can and have done these things for a long time now, but all any of these things are (currently) useful for is setting a timer and playing music.
Without a rewind style access to your entire life, there is a real limit on how much useful stuff this technology can actually do.
An LLM with exclusively global context (no personal context) is already pretty powerful, but certainly not helpful as an assistant.
In theory, perfect personal context would help a lot. Certainly a smart human being who follow me around all day can be very useful, but is expensive and has human limitations.
I don’t think I want exactly that from an AI. I am not thrilled with stuff like the Humane pin. I don’t trust that it will keep all those observations secure. I don’t trust that it will sort through all the noise to identify what’s actually important.
I think a mobile OS is a good balance. It has a LOT of personal context. I already trust it to process that data securely. The context is already well-structured and actively curated. I’m optimistic that adding LLMs to this will be significantly useful.
I’m unsure I’m willing to lock myself in to the Apple ecosystem to an even greater degree though - they make it devilishly difficult to extricate yourself, and I already find the enforced storefronts built into their apps obnoxious. The Books app went from being a library to a shop - how long until the Photos app is trying to sell me things?
If anything, I may wind up bifurcating my life to an even greater degree - iPhone for general life admin (calendar, phone book, banking etc), Linux for productivity, creativity, gaming.
Historically, Apple use to have really nice functionality built into Mac iPhotos to design and order photo books and prints. Sadly they dropped this.
Works OK, but I did like the Apple interface better.
When I think about going deeper into the Apple Ecosystem, I mean stuff like:
Continue to use a 3rd party mail and calendar services, but ensure they're available through the native iOS Mail and Calendar apps.
Increase usage of reminders
Promote iCloud file storage, syncing it with other places to ensure it has a recent view of my holdings.
Increase use of Apple version of subscriptions I might get elsewhere, such as News.
From what I understand, such gimmicks are (almost) essential utilities for many.
It also is possible that small improvements in intelligence will drive sales. For example, I don’t see much use for generating entire novels on-device, but good autocorrect and prediction of a few words may drive sales. I also like the automated tagging that iOS already does on photos.
Also, even if the new features only make people upgrade a few months earlier, that will be serious amounts of money for Apple.
Last week I was in an art gallery and took some photos of the paintings, when I suddenly noticed that it was drawing a focus around the frame and there was a 'Look up Artwork' option, which work fairly well (not 100%). Just as well, as I couldn't remember the works myself the next day after looking at the photos.
If I hadn't taken those pictures I probably wold never have known the feature existed.
iPhone 15.
You can argue "Hey Apple shouldn't charge so much ram, it's better for me, it's better for you, it's better for everyone" and then you get attacked for it (usually by claims that they use "suuuuper" special flash and ram chips bla bla super high performance bla).
When I listened to the Apple Developer’s presentations this week I always filtered everything said with “they are talking about proposed product updates, and some existing working functionality.”
I have been enjoying running all of Apple’s beta iOS, iPadOS, and macOS beta releases this week, and there are nice features in Photos, Calendar, etc., but we need to wait to see what they release next fall.
I am not an Apple developer, really, I did really well selling a Mac app in 1984, and a Mac app I wrote two years ago had pathetic sales. That said, I have been trying the beta Xcode with LLM code completions, etc., and I have been experimenting with MLX and Apple Silicon for a long while. Apple is definitely an interesting player in the AI product space!
Well not really. I haven't seen anything new here, unless you consider genmoji a killer app...
The whole thing comes down to - your request may get sent to ChatGPT for advanced features.
Better doesn't mean new.
If that was your takeaway from the keynote/SOTU, then I would suggest you go back and watch it again. That's not even remotely close to being accurate.
One thing I really want is for Siri to be able to control my house with Home Assistant with natural language, e.g. "turn off all of the lights outside except for the balcony". The updates to Siri seem like it _could_ make this possible.
I would _love_ to be proven wrong, so please do show me alternatives.
As long as there are teenagers that are susceptible to peer pressure and don't know marketing tricks, I think it could last forever. Look how many other Veblen good companies have existed for hundreds of years. With Apple's cash and media, they can have rappers and celebs always using/mentioning their stuff.
Consider also that you might have made a mistake in your analysis and there are factors involved other than celebrity endorsements – for example, wouldn’t it be interesting to know why Nokia, Google, Samsung, Sony, LG, etc. weren’t more successful thanks to their own endorsements? Your own peer pressure might be hindering your understanding of the market.
Technical capability is rarely more important than product market fit.
E.g. Apple generating emojis using a diffusion model is clearly less SOTA than other image models, but could be a really neat feature from a product/user perspective. Would a better diffusion model help? Maybe, but it probably changes the product less than all the UI and tight integration work they will have done.
If I was to guess, I would guess many more emoji's will get created through this feature than images generated in Microsoft's 'Image Creator' in Paint (I could be wrong - just my gut feel).
The way to minimize hallucination is to use LLMs to manipulate text (translate, summarize, query, reword, etc) rather than generate it out of thin air, and arguably Apple are doing this by integrating LLM use into the product and therefore controlling it's application, rather than just providing a chat interface that people are wont to use an hallucinating "search engine"/oracle.
Apple relegated ChatGPT to a 3rd tier AI capability for "google answers", and "write me a poem" party tricks, and rightfully put it behind a privacy disclaimer. It looks barely better than the various ChatGPT shortcuts so many people have cobbled together. That part of the announcement stuck out like a sore thumb and looked to me like a huge L for OpenAI. The "partnership" was a nothingburger. Like someone at Apple agreed to it early on but then late in the project realized it wasn't needed at all. So much for SOTA.
It is mainly B2B market in reality. Look for the API capabilities: https://platform.openai.com/docs/introduction
It certainly is. And the Rabbit R1. How in the world did people who supposedly know anything about AI think they could make that work as a standalone device detached from rich context. The sad thing is, rich context may not be possible outside of OS-level integration that is gate kept, but I still think they're idiots for trying.