This is an excellent post by Molly White about the potential conflict between making the world’s knowledge more accessible through AI and the risk of destroying the foundations for open content on the web:

The true threat from AI models training on open access material is not that more people may access knowledge thanks to new modalities. It’s that those models may stifle Wikipedia and other free knowledge repositories, benefiting from the labor, money, and care that goes into supporting them while also bleeding them dry.

She also gets at something I tried to articulate in one of my posts last year about putting up roadblocks for crawlers. We don’t want to make the web worse in the process of protecting content from AI training. Molly again:

Often by trying to wall off those considered to be bad actors, people wall off the very people they intended to give access to. People who gate their work behind paywalls likely didn’t set out to create works that only the wealthy could access. People who implement registration walls probably didn’t intend for their work to only be available to those willing to put up with the risk of incessant email spam after they relinquish their personal information.

AI companies are moving so quickly that it’s going to take the open web and standards organization a little time to catch up. It’s not hopeless, though. Personally, I do want all of my blog posts — and the entire content of my book Indie Microblogging — available for AI models. But if other writers feel differently, there should be steps they can take without also taking a step back from the open web.

I believe all these things:

• AI models with all the world’s information are an incredible resource and will transform education and how we work.
• AI training should respect how authors intend for their content to be used without forcing authors to mangle their own content.
• AI companies shouldn’t take from the open web without giving back in citation links and money to authors and organizations.
• AI slop will become a problem for both users and AI training, so we need a web filled mostly with human-generated content.

I remain optimistic in part because despite how divisive AI has become, this year is also seeing an amazing return to open web principles. More people are blogging. More social networks are based on open protocols. We need to be thoughtful in how we navigate all of this, finding the right balance with AI training that doesn’t undermine what we love about the open web.

We are so used to being able to do almost anything in software. For example, today I was looking at how long it takes to load the list of your blog posts in Micro.blog. It’s a couple seconds, but I’d like it to be half a second. I know from many years of web experience that there is no technical reason why it can’t be faster. Just a little more caching and database optimizations.

With the Apple Intelligence and Siri delays, people have speculated on why it’s late. Maybe it’s about getting the security right. Maybe it’s just buggy and taking longer than expected.

But what if Apple has discovered that it’s not actually possible? AI is entirely new, with new requirements that stress the limits of hardware. Apple is attempting to cram a clever intermingling of data and Siri features into 8 GB of RAM. As a comparison, the largest version of DeepSeek R1 can only be run on a brand new Mac Studio with the M3 Ultra and 512 GB of RAM.

Apple does have an out if on-device models fall over: private cloud compute. But scaling that out to hundreds of millions of iPhone users goes well beyond what Apple had presumed was needed when they talked last year about ramping up production of M2-based servers for AI.

If Apple needs to lean on the cloud to really make Siri work, I think it will be the largest server undertaking that Apple has ever attempted. And they need to balance this with their commitments to energy use and the environment. This is not something you just spin up out of nothing.

Another path would be to simplify their approach, starting with a more manageable set of tasks that the new Siri could do. Something that fits within the limits of iPhone hardware and a realistic deployment of new servers. Apple could focus on making Siri a little more capable and more reliable, saving some of the harder challenges for later. Most people have no idea what Apple promised last year, despite the TV ads, so a reset of expectations could get Apple back on track.

We don’t know what’s going on inside Apple. Apple Intelligence might need a little more time or much more time. The only truly worrying scenario would be if the sunk cost fallacy is blinding them to how badly they are stuck.

Ever since Apple revealed their AI strategy to lean into on-device models, there has been a sort of tension with the approach from other companies like OpenAI, Google, and Amazon. Was Apple Intelligence going to work? There are advantages: for user privacy because more data stays on your phone, and for scaling because the load is distributed across millions of phones instead of only running in data centers.

Now we know that a more advanced and personal Siri is delayed until iOS 19. From an Apple spokesperson via Daring Fireball:

We’ve also been working on a more personalized Siri, giving it more awareness of your personal context, as well as the ability to take action for you within and across your apps. It’s going to take us longer than we thought to deliver on these features and we anticipate rolling them out in the coming year.

Jason Snell blogged a recap of the Siri demo from WWDC last year:

This led to one of the killer demos of WWDC 2024, in which Siri was able to understand when someone’s mom’s flight is landing by cross-referencing an email with real-time flight tracking to get a good answer. From there, the demo pulls a lunch plan with mom out of a text thread and then displays how long the drive is to there from the airport—all from within Siri, rather than individual apps.

I’m not worried about a delay. Software is complicated and we all hit unexpected challenges. I’m worried that Apple can’t pull this off at all. Parker Ortolani is blogging the same kind of questions:

It felt almost vaporware-like when revealed at WWDC and it certainly seems like they are having a great deal of difficulty making it a reality.

And from Federico Viticci:

…one has to wonder why these features were demoed at all at Apple’s biggest software event last year and if those previews – absent a real, in-person event – were actually animated prototypes.

There are two potential problems with Apple’s approach:

• On-device models are small and limited by hardware like RAM.
• App Intents for extensibility require support from developers and won’t be available on all devices.

I’ve written about this before with Siri, including in this blog post right before WWDC last year. Because each device has its own version of Siri, it is hard to ever have a universal assistant that works everywhere and is extensible. There are things Siri can do on a phone that it can’t do on a HomePod.

I’m not seeing even a hint of a solution from Apple on this. If anything, what they showed developers with App Intents at WWDC is going to create an even more disjointed Siri across platforms, because third-party apps may not be available everywhere.

Steve Troughton-Smith on Mastodon is skeptical that third-party developers will help make this vision a reality:

Delayed or not, Apple’s proposed Intents-based Apple Intelligence features require a ton of buy-in from developers for it to be of any real use, and the incentives really aren’t there — donating your app’s content and functionality to enrich Apple’s AI alone, bypassing your UI, UX, branding, metrics, et al, to be delivered in a content soup alongside your competitors.

While App Intents don’t exclude the idea of other APIs for developers to use system models directly, I don’t expect we’ll see anything beyond App Intents until the new Siri is ready, and maybe not even after that. Ben Thompson in today’s Stratechery article:

Apple gives lip service to the role developers played in making the iPhone a compelling platform — and in collectively forming a moat for iOS and Android — but its actions suggest that Apple views developers as a commodity: necessary in aggregate, but mostly a pain in the ass individually.

Ben makes a strong case that Apple should be opening up their models to third-party developers, especially given the incredible potential of the M3 Ultra. Siri is designed for an 8 GB RAM world. The M3 Ultra can have 512 GB. Mac developers will have to bring their own models to take advantage of the great hardware in modern Macs.

Back to the disconnect between on-device models and cloud-based AI, Alexa Skills have been around for a decade and they will apparently work seamlessly with Alexa+. It’s all in the cloud.

I use ChatGPT a lot, every day, and yet there are some things I’m not comfortable sharing into the cloud. I don’t care if it knows that I’m planning a trip or what code I’m working on, but I’d be very hesitant to talk to a cloud-based assistant about truly private matters. Who knows where that info might accidentally end up.

Apple competitors could undercut a lot of Apple’s strategy by creating their own version of private cloud compute. Most users do not really think or worry about this. They’ve been storing emails with all sorts of private details on Gmail servers for years. But making cloud-based AI as secure as possible is just a good thing.

I’m not sure Apple knows what a big risk they are taking by letting OpenAI and others lap them in the AI race. It’s a risk that will pay off if they can execute. Just as likely, though, we are seeing such a disruption in computing that Apple is vulnerable for the first time in a decade.

Big companies like Apple do not move quickly. Amazon put everything into rebuilding Alexa and it has taken nearly two years. If there is a truly new AI device, a post-smartphone pod that we keep in our pocket or that’s built into our glasses, Apple’s strategy to entangle AI with phone hardware will have been proven all wrong for this moment, and they will have no response.