I was promised AI@Home but I got... foldable phones and touchscreens?

When John Ternus was announced as the new CEO in May this year I did a quick deep dive with Google Search AI Mode (aka the worst AI with the best search index) and learned that he was the one behind the Apple Silicon transition, and thought "interesting, a hardware guy, what might this mean"? There were already rumors about a foldable phone, but that seemed more like a gimmick to me at the time. Besides, Samsung had already done this?!

I have to admit I went on a bit of an "AI-psychosis" ideation session with AI Mode about where this could lead them instead. I mean Apple has (had?) one of the best local LLM compute hardware out there. NVIDIA of course takes the crown in terms of compute power, but in terms of memory, noise, and power needs, Apple has a unique position. Models with 96GB or more can run serious models, and neither turn your home acoustically into a datacenter blowdryer experience, or consume >500W of power. So maybe Ternus would finally double down on that, and create a "home AI" line of products. It would fit their privacy stance. And maybe could revive the Apple Home product lineup? Just think about it, maybe they could create a whole new home appliance market so in the future you'd not only get a fridge, a heat pump, but also a "home AI" server (maybe doubling as heating) that becomes like "Jarvis@home"? Maybe in the future when buying a house, you'd not only ask which generation the solar panels are, but also how much RAM and tokens/sec prefill the home AI does. It would ALSO be an interesting alternative to the massively concentrated data centers that we're building now. And it's not like Apple has never created whole new product categories before!

I was /this/ close to writing a blog post, laying out my predictions Gartner-style, together with events that would need to happen to show we're actually on that path. One of them was Apple releasing (finally) an updated, agentic Siri that showed they could actually build assistants that don't feel like they're from 2008. But I never got around to finishing that draft. And honestly, I wasn't sure just how much the long session with Google Search AI Mode had desynced me from reality.

So yesterday, half a year later, I get an email from Apple Developers advertising the new "Touch on Mac" trainings. And I was like "They DID NOT...!!!" But it's true. Not only did they release a foldable phone that's massively overpriced and it seems (currently) very brittle, they also added a touchscreen to their Macs?! The one thing that Steve Jobs very eloquently explained over and over again doesn't make any sense? That my daughter's 300 Euro Acer laptop has, that I never see her using?

Also, the new Siri was released, and sure I can ask it what's the weather tomorrow is, but compared to any other agent out there (yes, even Google Search AI Mode), it was so terse I didn't even get back to it. I'd rather hook up Claude Code to the Google MCP. And while the new Apple Foundation Model tech is good (did you know you get a command line tool called fm that's like llm to run a local model?), including seamless escalation to cloud models and their deal with Google on Gemini models, it becomes clear that not being an AI company means that Apple can probably not pull that off.

In the meantime, the competitors are closing the gaps in terms of mid-level local AI compute. AMD with their Ryzen AI Max+ 395 as well as Nvidia's Spark technology can't compete in terms of memory bandwidth, but they're getting closer. At the same time, smaller models are becoming more capable. In particular there are smaller models specifically trained to be agentic like Qwen 3.8 27B or Meta's Muse Glimmer massively improved the capability of local models and make a "AI@Home" more plausible.

So currently it doesn't seem like Apple will be able to do it. To really create a Siri that has the "Apple shine", they'd need to be able to build that product end to end, like they always do. The new on-device models seem to show they have the capabilities in principle, but they'd need bigger models, do massive fine-tunes, design personas, add guardrails, all that, and while it seems they have people on it, their latest product iterations don't show it (yet?). Meta with Muse actually showed how that could look like (plus massive data privacy headaches for me), and they've been training and releasing models of all sizes for the past couple of years. And for the longest while, local AI compute on Apple Silicon has been an open source effort, only lately having Apple closing the gap with MLX

I mean, I'm totally open to the idea that "AI@Home" is a dumb idea after all, but after yesterday's announcements I think Apple is not getting closer to being leading this.