It sort of makes sense but can be risky for OpenAI. I suspect that there are limits to this deal. For example, Apple will pay x amount if conversion rate is lower than x.
You're adding a few billion devices to do inference for if you're OpenAI without knowing the conversion rate from free to paid ChatGPT accounts from Apple users.
Also, can OpenAI really handle that much more traffic? Sometimes my paid ChatGPT account stalls or is slow even as of June 2024. Now they're going to add a few billion more users/devices at once. That's insane. How are they going to do this?
> As a result of the company's partnership with OpenAI, Apple Intelligence includes an integration with ChatGPT, allowing Siri to determine when to send certain complex user requests to ChatGPT.
OpenAI does not power Apple Intelligence. Apple Intelligence will pass of queries to OpenAI (with explicit user permission) in certain circumstances where it feels OpenAI would have a better answer.
Email, etc, is all Apple inference, the presentation strongly implied that was on-device. The system always explicitly asks “Ok to send this query to OpenAI?” or user initiated request.
I think the Apple server inference is handling things like image generation where you need a beefy guy to get results quickly, and don’t need to transmit as much user context to service the request.
There isn't anything like that currently - The integration that's been announced/shown is more like the current fallback in Siri of "Would you like me to Google that for you?"
No, there is - from official Apple Intelligence announcement:
> With ChatGPT from OpenAI integrated into Siri and Writing Tools, you get even more expertise when it might be helpful for you — no need to jump between tools. Siri can tap into ChatGPT for certain requests, including questions about photos or documents. And with Compose in Writing Tools, you can create and illustrate original content from scratch.
They’ve been explicit that only some knowledge related requests get sent to OpenAI, and only after the user has been asked “ok to send this to ChatGPT?”
If they allowed your private email contents or any other private context to leak outside their encrypted enclave (local or their new online hardware) then the entire set of privacy claims would go out the window.
Eventually, but not initially. I read this thread as if it was about the initial release, not the end-state.
(Given how rapidly AI changes, anything more than a year away isn't easily planned for, even assuming no surprises on the anti-trust front for every Big Tech company nor any surprise new legal obligations for AI).
Does it need to be sold in the US or simply chaning os settings to use the en_US language is enough to activate this ?
I'm in europe but working mostly in english and therfore I set my mac in english. Could I use Apple Intelligence ?
Per Apple's site [1], it's just the language setting:
"* Apple Intelligence will be available in beta on iPhone 15 Pro, iPhone 15 Pro Max, and iPad and Mac with M1 and later, with Siri and device language set to U.S. English, as part of iOS 18, iPadOS 18, and macOS Sequoia this fall."
I'm pretty sure Apple publicly said that they're also considering Google Gemini.
It's possible that Apple sees it as a future app store-like revenue. You can use any AI model you want, and if you want to subscribe to Gemini Pro or ChatGPT Plus, Apple takes 30%.
I was certain that Apple wants to create their own cutting edge LLM to compete with GPT4 or Gemini or Claude or Mixtral. But they've decided to instead, open up all their platforms to 3rd party models.
I suppose Apple may still want to create an Apple Intelligence Pro in the future but their models aren't leading at the moment.
They are getting the data through their wires. That's why Apple make so much effort about anonymization - it's not guarantied by contract.
best free RHLF OpenAI can get
my 2c.
I wonder if VCs will be the ones left holding the bag at the end of this AI cycle.
You have hyperscalers "investing" in AI startups with compute credits, Nvidia "investing" in AI startups (who spend most of it on buying/renting their GPUs) and now this Apple deal.
What is different here is that the non-VC investors investments are being recycled back into their goods&services see: cloud credits, buying NVDA compute, etc.
I don't think we had a similar pattern in the mobile era. Did Apple ever invest in an app in exchange for App Store fee credits or some dubious scrip like that? I don't think so.
From the article, it sounds like Apple aims to create an "AI App Store", where the users can choose the AI Supplier thmeselves instead of only beeing locked in to "Apple Intelligence". I think its a smart move to build an AI platform instead of beeing an AI supplier by themselves.
Can't read the article, but the way they described it in the WWDC Keynote, you _are_ locked into Apple Intelligence on eligible devices for all the stuff Apple Intelligence actually does. The "AI App Store" only comes into play for specific questions that Apple Intelligence punts on and then tells you, "I can't answer this, do you want to query an AI Supplier?" Apple is not giving you the option of selecting a back-end AI Supplier for all AI-related stuff on the device.
I wrote in "Apple Strategy in a Nutshell"[1] that this deal is awful for OpenAI.
Not getting paid obviously makes it worse for them, and more favorable for Apple.
I think that what OpenAI is getting in return is their brand name being featured prominently in one of the top selling smartphone brands in the world.
The other wildcard here is that OpenAI will likely exit in some way, most likely by acquisition. Who knows, maybe Apple would buy them.
If I were to defend OpenAI's decision to work with Apple, I would say that Apple's ability to do gap analysis won't necessarily help them catch up with OpenAI's capabilities. If this was the case, Apple would have figured out how to make a better modem than Qualcomm by now.
I don’t think OpenAI company structure would allow an acquisition per se, although how much they bend the rules (such as with Microsoft) has not been tested
It's so bizarrely awful for OpenAI that you almost start to wonder what the auxiliary value of a few million iPhone user's data is.
But as Scarlett Johansson taught us; OpenAI will always ask first. And then they will do whatever they want anyways, regardless of what you respond with.
Apple was being pressured by Wall Street to make move here, so you'd think they'd have the weak hand. However, by the time they moved, there were 3-4 vendors Apple could play off each other.
It seems like Apple played this very well and are going to treat LLM the way they do search. At some point the LLMs will be bidding for placement on iOS devices, like the Google search deal.
In conversation with reporters after the WWDC keynote, Apple's senior VP of software engineering Craig Federighi revealed that as Apple Intelligence evolves, the company eventually wants to give its users a choice between different AI models, and suggested that Google Gemini could be an option in the future.
"We think ultimately people are going to have a preference for certain models they want to use, maybe one that's great for creative writing or one that they prefer for coding," said Federighi. "Maybe Google Gemini in the future. I mean, nothing to announce right now, but that's our direction."
> Why even list Chinese vendors? Those would never happen in a western market.
Because the Western vendors are banned in China, and Apple sells a lot of iPhones in China.
And even in a Western market, you might have Chinese immigrants choosing the Chinese vendor. That's probably not enough to justify the integration in and of itself, but if you already have to do it for the Chinese market, why not offer it in the West too?
This is tangential, but I have been wondering: is there anything to prevent Apple from reading both sides of the exchanges between the user and OpenAI, and training its own models based on that data?
At the bottom of the article, they allude to a future "AI App Store":
> Eventually, Apple aims to make money from AI by striking revenue-sharing agreements whereby it gets a cut from AI partners that monetize results in chatbots on Apple platforms, according to the people. The company believes that AI could chip away at the billions of dollars it gets from its Google search deal because users will favor chatbots and other tools over search engines. Apple will need to craft new arrangements that make up for the shortfall.
Still curious if Apple "at least" has to pay the cloud-compute for the "free" OpenAI services they use (with OpenAI just licensing them their product for free) or if Apple really gets free AI-features AND 30% RevShare from handling sales of Premium OpenAI accounts...
I can't imagine that the profit of sold Premium subscriptions for OpenAI is sufficient to pay for the infrastructure hundreds of million iPhone users will consume using the free account.
Moreover, with Apple obviously executing their usual playbook, being the gatekeeper and total owner of all statistics, then slowly diverting traffic to their own products and cutting out the supplier...
Sam has gotten OpenAI well capitalized, and is playing a bit of a ponzi game. The big moves he has made are almost certainly not sustainable, and he's betting the farm that OpenAI can snowball into a monopoly position, bring costs way down and take over the world before investors start giving him the cold shoulder. OpenAI is far ahead of pretty much everyone but Meta in the race, but in the long run I think Mark and Yann can shut them down if they keep releasing free models.
That’s not gonna work. The Chinese pricing is already 1/20th of oai. Not as good sure but that very much limits scope for a pure cost play. You can’t win that against state backed entities
a) Microsoft pays for compute. OpenAI never intended to run their own cloud.
b) Apple's 15%/30% is a "cost of channel" for developers. It is the fee they pay to get free users delivered to their app. And the decision to do this is purely a business one that can be changed at will and does not apply at all to this situation.
c) Calling Apple a gatekeeper for implementing additional features to their OS is odd to the say the least. They have made it clear during the keynote/WWDC presentations that they plan to add additional providers than OpenAI.
> a) Microsoft pays for compute. OpenAI never intended to run their own cloud.
No, Microsoft INVESTED in OpenAI and in exchange became OpenAI's exclusive cloud provider. That doesn't mean they pay for the compute of all customers.
> b) Apple's 15%/30% is a "cost of channel" for developers. It is the fee they pay to get free users delivered to their app. And the decision to do this is purely a business one that can be changed at will and does not apply at all to this situation.
Not sure what you meant to say here. Purchasing of OpenAI "premium" would be handled via Apple's payment platform, for which Apple will likely take a cut.
Regardless, OpenAI's gamble would be to onboard ~100M iPhone users under a "free" tier in a single rollout step, in hopes that the conversion-rate to an OpenAI premium will be high enough to offset the (infrastructure!) cost of serving all those users for free
> c) Calling Apple a gatekeeper for implementing additional features to their OS is odd to the say the least. They have made it clear during the keynote/WWDC presentations that they plan to add additional providers than OpenAI.
Apple is the gatekeeper because they are in control of the decision to offload AI-tasks to a 3rd party service or do them "in-house" (on-device/Apple-cloud),
This doesn't change regardless whether they stated this upfront at the keynote/WWDC or not.
Because of this structure, Apple is are aware of every single AI-task that is to be done and will use this information to influence their in-house roadmap (a natural decision, because on-device is better than cloud, and Apple-cloud is better because <insert tech of Apple cloud here>).
At some point in the future, the user will be able to select WHICH 3rd party he wants to use, but so far I don't see any indication that Apple considers themselves as a 3rd party.
And even if they would, they hold more information than all other providers because they FIRST decide whether a task can be handled on-device or not.
Do they actually hold more information, though? If it’s handled on-device, Apple never knows about it. If it’s sent to something other than ChatGPT, Apple can’t even read it. This seems like a pretty level field for anybody who wants to step in.
That's the misleading part of all of Apple's statements about privacy.
Apple only ever talks about your "personal data", they let the audience conclude what that means.
A picture is personal data. Information about the content of that picture anonymized in a way that the user/subject cannot be identified, is not personal data anymore.
A profile generated from all your usage-habits combined is not "personal data" anymore.
Not a popular statement, but I believe the main reason for Apple pushing on-device ML-models is to extract non-personal information from its devices without the need for the "personal information" to be seen by Apple. There's even a paper on that [1]
Compute and storage is paid by the user, only the extracted data is delivered to Apple.
Beyond that, Apple's entire effort on privacy conveniently ensures that only they know everything about their users, and ensure that the customers don't (accidentally) tell someone else something about themselves.
Ok so let’s play it out. I look at a picture of a cat, and Apple receives a completely anonymized log saying someone somewhere looked at a picture of a cat. This log has no information about me and cannot be tied back to me. Am I supposed to be concerned this is an affront to my privacy?
People want big data tech, most people think this is fine and not a privacy invasion. Apple’s approach to privacy is leagues better than e.g. Meta or Google. If you want no record, no accounts, no logs for anything ever, don’t buy a smartphone.
You're adding a few billion devices to do inference for if you're OpenAI without knowing the conversion rate from free to paid ChatGPT accounts from Apple users.
Also, can OpenAI really handle that much more traffic? Sometimes my paid ChatGPT account stalls or is slow even as of June 2024. Now they're going to add a few billion more users/devices at once. That's insane. How are they going to do this?