The long and winding road that leads to Siri AI

15 Aug, 2026

While I was renewing my Apple Developer Program membership earlier today, a thought occurred to me: until this year, I’d never installed and used beta versions of Apple operating systems; not in any real capacity, anyway. I’d been paying for this membership since 2014, and in that span of time, there was only one moment in which I’d installed a beta OS — the (southern hemisphere) winter of 2023, for macOS Sonoma.¹ For the longest time, the only privilege of the membership that I took advantage of was having Bytes listed in the App Store. Later on, I started to install release candidate versions of Apple’s OSes. The distinction is important — RCs are not beta releases. If all went well, the exact same software shipped as a public release one week later, so I felt comfortable installing these on my personal devices. As of this year, I began sideloading open-source apps that aren’t available on the App Store. These are all great developer privileges, but the main one (aside from the obvious one of having App Store listings) is access to beta operating systems. I’ve never taken advantage of this privilege, as I’ve never had a reason to risk the stability of my personal devices — until now. This probably sounds crazy, but, ever since WWDC26, I’ve been running every beta release of iOS 27, iPadOS 27 and macOS 27 on my personal devices … for Siri. Specifically, for Siri AI, and the semantic index in particular.

Apple Intelligence

To understand why I’ve done this, we have to go back to WWDC24. The last third of that year’s keynote was devoted to Apple Intelligence, a set of features that promised to bring the benefits of generative AI to Apple’s operating systems. “AI for the rest of us”, as Craig Federighi quipped. For those who are unaware, this is a callback to the legendary description of the Macintosh as, “the computer for the rest of us” — a statement that, throughout the years I’ve had Apple’s computers in my life, seems to have become even more true over time. The line also just sounds great. We’ve been hearing for years about how AI is going to change the world and revolutionise everything, but for a lot of people, the first word they associate AI with is slop.

In my mind, I’d sorted the initial set of Apple Intelligence features into three categories — text manipulation, image generation and the semantic index. Text manipulation is self-explanatory — it’s the text message, email and notification summaries, the Safari reader mode overviews, and the Writing Tools. Image generation encompasses Image Playground, Genmoji and Image Wand. The semantic index, the most interesting part of Apple Intelligence, had multiple subdivisions. There was the world knowledge component, the App Intents framework and an LLM-powered rewrite of Siri that would tie everything together. Apple made it clear that the semantic index would not arrive until later in the iOS 18 release cycle, and that we’d only get the world knowledge component in the (northern hemisphere) fall. In practice, this meant access to ChatGPT in the Siri interface.

As if I didn’t already sound crazy enough, here’s another hot take: I actually liked the first release of Apple Intelligence! I know, I know … Image Playground was near useless, BBC News notification summaries were wildly incorrect, and Apple failed to deliver the semantic index on time. Even still, for my use cases, it was awesome. Notification summaries helped me get the gist of a Messages thread before diving into it. Email summaries helped me skip a lot of the mindless drivel that occupied my inbox. And Image Playground — infantile as it was, in comparison to the photorealistic diffusion models of the time — produced images of friends and relatives that were charming enough to send on occasions like birthdays and holidays. Most importantly, Apple Intelligence provided a private interface to ChatGPT. I created an OpenAI account back when they released the first preview of ChatGPT. My initial impression was that the LLM was intriguing and amusing, but not particularly useful. That ended up changing over time, which placed me in a quandary. How can I use AI to solve more of my problems without letting OpenAI build a profile of me? At the time, they hadn’t yet announced their plans to serve targeted advertising, but I already knew that I didn’t want yet another company to know more about me than was absolutely necessary. I’d already made that mistake with Google many years prior. So, before Apple Intelligence was a thing, my workflow for accessing ChatGPT was as follows:

1. Return to the Home Screen
2. Hold the Safari icon, and open a new private window
3. Navigate to chatgpt.com
4. Type my query into the box, making sure to avoid any sign-in prompts
5. Await a response from the LLM, then close the tab and private window after receiving said response

Apple Intelligence’s ChatGPT extension simplified that to typing, “Ask ChatGPT …”, without having to switch apps or open Safari or anything. It could be activated by saying, “Hey Siri”, it could take a screenshot to provide additional context, and it did all of this with the understanding that all requests were anonymised, that OpenAI could not build a profile based on what I asked ChatGPT, and that it would not use my conversations as part of a dataset to train future models. It was through that interface that I experienced ChatGPT — mostly to ask ephemeral questions with throwaway answers, and occasionally to fix a software problem that I’d been presented with on-screen. It was useful enough, and I accessed it often enough, to write a text replacement shortcut for it — “Askch” for “Ask ChatGPT”.

All of the above is to say that I generally liked Apple Intelligence for what it was. Still, I knew the best was yet to come. 2024 came and went, with iPhones — “built for Apple Intelligence” — ones that were advertised with, but completely lacked, the semantic index. 2025 came and went, with something rotten in the state of Cupertino. Finally, in 2026, it is here.

Siri AI

I’m usually a, “do as I say, not as I do”, kinda guy. I’m the person that would tell friends to wait for public releases of OS updates, or at the very least, a release candidate. But, if you do your due diligence, create device backups and secondary APFS partitions, and have the stomach to live with bugs, Siri AI delivers the experience we’ve been waiting for.

It is better in every single way. The interface extends from the Dynamic Island. It allows you to scroll through the entire transcript, and it keeps a persistent copy of that transcript in the new Siri app, synced via iCloud across your devices. The models powering the experience — versions of Gemini models that have been distilled by Apple — are faster and more accurate than the previous OpenAI models, providing answers that are replete with sources that you can check yourself, if you so wish. You no longer need to provide confirmation to send a screenshot — just ask a question about whatever’s on screen, and Siri will know to use that with the query. And, of course … the semantic index, the personal context, it’s here and it works! I specifically wanted to test this at a time when I’d need it the most, so I’d installed the betas ahead of a series of international flights. Here are some examples of requests that Siri handled perfectly:

• I needed to check into one of my flights online, so I asked Siri to find my booking reference. It searched through Mail, found the reference for the flight, and presented it in the reply, along with a link to the relevant email as a source.
• Next, I needed to fill in my passport details. I’d received a new passport last year, and I hadn’t used it often enough to have memorised the number just yet. I had Siri find the details for me. If they’re stored anywhere on the device, or in iCloud, it’ll find them within seconds. In the prompt, I also asked it to only present the details needed to complete the form, which is all it replied with.
• I was travelling across Europe via multiple airlines, each of which had different standards for carry-on baggage. I needed to know how large and heavy my bags could be to comply with their policies. I opened Flighty, which had a list of my upcoming flights, and then I invoked Siri, asking it to find the maximum dimensions and weight I could bring for carry-on baggage for my upcoming flights. When it replied, it did so with references to each airline’s standards, correctly ascertaining the dimensions and weight that would be permitted across all three of them. For the bags that I ended up taking based on Siri’s recommendation, I didn’t face any problems at the departure gates.

I’ve also used Siri AI to assist me with tasks that have nothing to do with travel:

• Prior to the test of Australia’s national warning system, I asked Siri what time the test was going to occur. I then followed up by asking it to schedule a reminder for two minutes prior to the alert. It created a reminder for 13:58 AEST, Monday 27 July 2026, exactly as expected.
• I had it compare two of the files in my Bytes project folder. The first file was for one of my view controllers, written in Objective-C, the second was a replacement for the first file that I’d manually rewritten in Swift. It correctly identified the parts of the logic that I had successfully ported over, and the parts that hadn’t been rewritten yet.
• I had kept a list of altitudes that contrails form at for each War Thunder map in my Steam overlay. For reasons unknown to me, the Steam overlay has stopped appearing. I’d copied the list into Notes, from where Siri could access it. Now, I can say, “Hey Siri, search my notes for the Denmark contrail height”, and Siri will verbally reply, “The contrail height for Denmark is 29,067 ft”.

Just for fun, in the midst of writing this paragraph, I asked Siri, “What did I do last weekend?”. Siri replied, “Last weekend, you traveled in Switzerland, including a trip to Jungfraujoch and hiking in Fieschertal. You also spent time with family, taking photos and managing travel details like train tickets and eSIMs.” Then, it presented a timeline of events that the SBB app (used to purchase train tickets in Switzerland) had added to my calendar, along with a set of personal photos from the peak at Jungfraujoch. It also provided a reference for the reminder I’d written to extend everyone’s travel eSIM validity. No other model, on Apple platforms, has this capability, and there’s no other model that I’d entrust with handling my photo library and calendars.

Siri AI vs ChatGPT, Kagi Assistant

Speaking of other models, let’s do a quick comparison with competing products. As I detailed above, I don’t have much experience with ChatGPT’s iOS app, but I do have a basic understanding of it. The formatting of the Siri app is similar — text streams in as the model generates it, and it can insert links, online images, and (as already mentioned) personal photos, calendar events, reminders, and so on. It’s like the Messages app, but the bubbles aren’t blue or green.

What I do have more experience with is Kagi Assistant. I haven’t written about it yet, but I love Kagi Assistant. Kagi’s privacy promise is what gave me the confidence to use it. I can’t even describe everything I’ve done with it — by default, it deletes conversations after 24 hours. It comes with a wide variety of models to pick between, but I often found myself sticking with the reasoning variants of Kimi K2. The gist of the comparison between Siri AI and Kimi K2.6 (reasoning) is that, while both models generally wrote similar replies for any given question, Siri AI would respond very quickly. Kimi K2.6 could take minutes to generate a response, but it shared its chain-of-thought along the way. Kagi Assistant’s interface can only output text, even if the selected model is capable of generating images, whereas Siri AI is multimedia — it can generate images, and it can also take action within apps. And, for the longest time, Kagi Assistant did not have an iOS app. The only way to access it was through a web browser. I’d had it saved as a shortcut on my Home Screen, but using it this way was far from a pleasant experience. Nothing matches a native app.² Lastly, there’s Kagi’s credit system, which limits you to $10 USD worth of AI tokens per month on the Professional plan. Kimi K2.6 (reasoning) is cheap enough that I ended up using, at most, ~$1-2/month prior to switching to Siri AI. I’ve never come close to the limit, but it’s something I’ve always kept an eye on. No one knows yet just how much you can use Siri AI before the system will cut you off, but we do know that Apple plans to restrict higher usage limits for AI to more expensive iCloud+ tiers. For my use cases, at least as of right now, I haven’t run into any limits.

What this means in practice is that, for the vast majority of my AI interactions, I can rely on Siri AI to give me satisfactory answers, and to help me with completing everyday tasks. For technical work — debugging software, writing web dev code, etc. — jobs where I want to see the model’s chain-of-thought and don’t mind waiting longer for a response, I use Kagi Assistant.

While I’m writing about other models, it’s worth mentioning that Apple Intelligence still comes with the optional ChatGPT extension. I’ve had absolutely no reason to use it, as Apple’s own models are so much more capable.

In Conclusion

For the vast majority of people, I still don’t think that the betas are worth switching to. Yes, they’ve been far more performant than betas from previous years, but there are still plenty of bugs, performance issues and visual glitches that continue to abound. But, if you’ve made it this far into this piece, and you know what you’re getting yourself into, Siri AI will not let you down. On the other hand, if you’re happy to wait one more month, get the 26.6 OS releases instead. The OS needs to re-index your entire device to create the semantic index. For me, this process took about a day to complete on my MacBook Pro and iPad, and 2 days on my iPhone. If you download 26.6 first, your device will begin to re-index in the background. By the time you get the 27 update, you won’t have to wait to use the full capabilities of Siri AI.

With regard to Siri AI, that’s just about all that I have to write. It’s real, it actually works, and it’s finally here, in the 27 betas. It’s there when I need it, it gets out of the way when I don’t, and it can do things that no other model can do. After two long years of waiting, I am happy to conclude that it really is AI for the rest of us.

Oh, and Image Playground is way better now.