Apple’s Craig Federighi Explains Apple Intelligence Delays, Siri’s Future and More | WSJ

Privacy is a tremendously important consideration for Apple Intelligence because the AI must access personal information stored on the user's device to provide personalized understanding.

“Most importantly, we view it as as personal. The bulk of the focus in chat bots historically have been around products that don't know you at all... but I think the real power of intelligence, as it pertains to how our customers would use our products, are intelligence that understands you. And with that comes a great deal of responsibility because to do that, that intelligence has to traffic in information that you've stored on your device. And that means privacy is a tremendously important consideration.”

For long-form summarization tasks, running AI in the cloud can both be faster and provide higher quality results than an on-device model could achieve.

“There are some cases, let's take a long-form summarization. An on-device model could try to do that, but would probably not provide the best result. If we run that instead in the cloud, it can both be faster in that case and also provide a higher quality result.”

Siri's improvements from Apple Intelligence will arrive in waves over the coming year rather than all at once, with the foundation of personal context understanding and device action capability being adopted in stages.

“Coming. It's coming in waves. So with Apple Intelligence, we've created this foundation that understands personal context and the ability to take action. Siri is adopting that in stages and will benefit in stages over the coming year.”

Apple will release features where Siri can answer deeper questions based on access to personal data like Calendar and Messages, which is currently not yet implemented.

“Apple's promised that Siri will be able to answer deeper questions based on your Calendar, Messages, and more.”

Apple's Cleanup tool uses AI to remove extraneous details from photos without generating new content, differing from competitors who have taken more extreme approaches to generative photo editing.

“Select Cleanup, and you can tap to remove items you don't like in the shot, and AI will fill in the rest. I can edit the mic out, I can edit that water bottle out. Some companies have chosen to go in a pretty extreme route, being able to generate new parts of the photos.”

ChatGPT and other large language models send every prompt to powerful cloud servers, whereas Apple prioritizes privacy by running AI models directly on the device and only tapping encrypted cloud-based models when more processing power is needed.

“ChatGPT and other large language models sends every prompt to their powerful cloud servers. Apple, however, is prioritizing privacy by running AI models directly on the device only when more processing power is needed.”

Siri processes approximately 1.5 billion requests every day and successfully performs tasks like opening and closing garages, turning on lights, sending messages, and setting timers.

“Siri processes today, I think it's something like 1.5 billion requests every day. And Siri gets things done for me. With Siri, I'm opening my garage, I'm closing my garage, I'm turning on my lights, all kinds of things. I'm sending messages, I'm setting timers.”

Photography has a long history of being viewed as something users can rely on as indicative of reality, and Apple prioritizes maintaining that trust by refusing to enable generation of entirely new photo content and by updating metadata on edited photos.

“But we are concerned that there's a great history to photography and how people view photographic content as something they can rely on is indicative of reality. And our products, our phones are used a lot, and it's important to us that we help purvey accurate information, not fantasy. We make sure even if you do remove a little detail of a photo, we update the metadata on the photo so someone can go back and check that this was an altered photo.”

Apple could have rushed features to market in an incomplete or problematic state, but Apple's philosophy is to thoroughly test each piece and release it only when ready rather than pursuing a 'bang moment' launch.

“You could put something out there and have it be sort of a mess, or Apple's point of view is more like, "Let's try to get each piece right and release it when it's ready, right?"”

Apple's approach to AI is fundamentally different from competitors because it focuses on creating deeply integrated, personal AI that understands the individual user rather than building generic chatbots bolted onto existing experiences.

“We looked at this as not how do we build another chat bot, and bolted on the side of our existing experience, but how do we create something that's deeply integrated? Most importantly, we view it as as personal.”

Apple Intelligence features are limited compared to what was demonstrated at WWDC in June, with many of the announced features not yet available at launch.

“Right now, Apple Intelligence is limited. Compared to what you showed in June- Sure. A lot of those features are not there.”

OpenAI's Advanced Voice Mode and Siri have quite different properties and applications; Advanced Voice Mode is designed for open-ended conversation and information queries like explaining quantum mechanics or writing poetry, while Siri is designed for device control and specific actionable tasks.

“The properties of something like that OpenAI Advanced Voice Mode and Siri are quite different. That OpenAI mode is great. If you want to go ask a question about quantum mechanics, and have it write a poem for you about it, and then answer another question and so forth. It's not gonna open your garage. It's not gonna help you send a text message.”

Apple Intelligence development and rollout represents a many-year, even decades-long arc of technological progress, and Apple is committed to executing it responsibly rather than pursuing short-term wins.

“When we think about the road for any of our features, but certainly Apple Intelligence, this isn't a one-and-done kind of situation. This is a many year, honestly, even decades-long arc of this technology playing out, and so we're gonna do it responsibly.”

Most companies' approach to data collection and storage runs counter to Apple's philosophy because they ask users to send all data and store it in ways they can read and process, which creates user uncertainty about whether the data will be used for purposes beyond the immediate query.

“It's hard to build, for one. I think it also runs counter to many companies' entire philosophy about how they see data. For many companies, I think their answer is, "Send us all your data, store all your data with us in a way that we can read and process." And so when you ask our cloud a question, we already have your data, and we're going to use whatever processing we've done on it to answer your question. But as a user, you're not quite sure. Is that all they're gonna do? And maybe they aren't doing anything with it right now, but they have my data. What are they gonna do with it in the future? I don't know.”

Apple's decision to make it easy to remove minor objects from photos in the Cleanup tool was internally debated, balancing user desire for simple cleanup against concerns about enabling modification of photographic evidence.

“Yeah, I would say, even the ability to remove that water bottle is one that there were a lot of debates internally. Do we wanna make it easy to remove that water bottle or that mic? 'Cause that water bottle was there when you took the photo.”

Apple's Private Cloud Compute encrypts messages sent to the cloud and performs AI processing on encrypted data such that the message is never recorded, and Apple has no access to the user's data.

“But the important part is that message that you asked to be summarized was encrypted to a trusted device in the cloud, that then performed artificial intelligence processing on it to return you the result and never recorded anything about that data, that email. Apple had no access to that email, no one else did. It's gone.”

Siri performs many useful tasks quickly and locally on the user's device every day, representing significant value that would be lost if the system were redesigned to prioritize open-ended conversation over device control.

“There are many, many useful things Siri does for you every day, does them quickly, does them locally on your device.”

The architectural spectrum of AI assistants involves trade-offs between different capabilities, and these different worlds—device-centric assistants and general-purpose conversational AI—will eventually converge toward a unified system.

“There's a spectrum here. There's a trade off across capabilities. Will these worlds converge? Of course, that's where the direction is going.”

Siri's usefulness has been established and proven for many years, but it will continue to get progressively better through Apple Intelligence integration and will not suddenly become a 'sentient pal' but rather improve incrementally.

“The great news is Siri's useful today, has been useful in many ways for many people for years, but it's getting more and more useful because this power of Apple Intelligence and large language models means that Siri can understand you better than ever before... But is there gonna be one moment where we say, "Siri is now my sentient pal," that... No, it's gonna continue to get better and better.”

Apple's marketing of iPhone and Apple Intelligence makes sense because customers need to know which devices will support these features, even though Apple is simultaneously communicating that features will arrive over coming months rather than immediately.

“We want customers who are excited about what's to come to know, "Okay, if I buy this phone, is this the right phone that's gonna do those things?" And yes it is, and so I think we're communicating that, while trying to make sure everyone understands that this is gonna be coming over the coming months, not all today.”