FeedCity

Congrats to the Iconfactory on 30 years! 🎉

I’m impressed with what Apple has come up with their most advanced on-device model. From the Machine Learning blog:

Built on cutting-edge Apple research, this 20-billion-parameter model uses a sparse architecture, activating just 1 to 4 billion parameters at a time depending on the request.

My Claude today pulled a Hal. It was so egregious. It made a change to the software based on a question I asked. It invented a whole set of instructions from me that I never gave it. And then it broke Rule #1 -- don't tell Dave what to do -- he is the driver. It is so import...

You know how job interviews for programmers include realtime problem-solving. Sometimes Claude is so dumb it could never pass one of those tests. Up till this point I would have been surprised to hear that.

Going to Amsterdam. brb

Long day but really good day. Felt like it had a little bit of everything. Still coming down after that Spurs game, really good. Tomorrow need to unpack a little more of what I glossed over during the WWDC announcements.

Go Spurs Go! Big game in NY. Found a sports bar to watch a little while in San Jose. Mostly Knicks fans here but a few Spurs fans here. 🏀

OpenAI announces draft S-1 in an oddly direct, brief blog post:

We recently submitted a confidential S-1. We expect it to leak so we’re just announcing it. We have not decided on timing yet; it may be a while because there are things we want to do that are likely easier as a private company.

Catching up from yesterday’s post from John Gruber about SwiftUI:

There’s something really wrong with SwiftUI. Amongst the apps I use, the best example is Apple Journal. Basic stuff that’s worked reliably for decades — some things that heretofore had worked forever — are dangerously broken.

Apple should’ve unified AppKit and UIKit years ago. SwiftUI could be purely a wrapper for the core framework, plus support for watchOS. Maybe too late to correct this now.

Sometimes you write a post and when you're editing it you realize you no longer support what you wrote. This is one of those times.

Monday session

Pulled several short straws in the office World Cup sweep stake: Turkey, Panama and Austria.

Seeing this screenshot on Mastodon from Finn Voorhees, sounds like Apple doesn’t want big companies to use private cloud compute. Which means it probably won’t even work with Mac apps like mine that are distributed outside the App Store. Apple’s going to lock this down.

I seem to be microblogging a little more than usual for the keynote, so decided to create a category on my blog for WWDC 2026 posts.

9to5Mac confirming what I thought I heard during the keynote: 8 GB devices are out for on-device AI. I think it’s the right call… Apple’s old models were just too small. But it’s weird that my pretty expensive iPhone 16 Pro Max can’t run local AI.

Expanding Apple Foundation Models to support image inputs and running in private cloud compute is a huge upgrade. The old models really weren’t capable enough for me, which is why I’m experimenting with bundling Gemma 4 directly in the Micro.blog app.

I feel bad saying this for all the Apple employees who dedicated so many hours preparing for WWDC, but I lose interest after an hour. Perfect keynote length would be about 70 minutes, not 100.

Apple didn’t try to take on too much with Siri AI. Only features they could actually ship. It looks good, but long term I still think using apps is the wrong strategy for assistants, because they can’t work well across platforms. It’ll work for now, and for privacy some people will prefer it.

I like in these Siri AI demos that they are doing things “live” (though recorded). It’s not fast, but it seems much more clearly actually working without edits.

A little surprised Apple is spending so much time in the keynote on safety for kids. It’s good, though! Thinking it helps set the stage for all the age-related laws from states, and we’ll return to safety and privacy during the Siri section later.