New iPhone 18 Pro is much better at staying cool in the Florida sun but it's still no match for a 65°C far infrared sauna. 🧖‍♂️

It is very surprising that disabling iPhone/iPad auto-lock (say, changing from "3 minutes" to "Never") doesn't require an authentication step. Bad actor with temporary access to your device could enable it and wait until you leave your phone out and fail to explicitly lock it.

I complete agree coding agents make software engineering harder, but the rub is that many professional software engineers are pretty bad at these new hard things and it's hard to see how they're not totally screwed. simonwillison.net/2026/Sep/24/harder/

For seemingly the sixth time in my career, DHH has arrived at the same conclusion I have—this time, that AI coding means we should be building more native apps, because development cost is no longer the bottleneck—but the way he goes about talking about it couldn't be more different youtu.be/vDjW_dRyKXY?t=1457&is=t5lg25x_naMgYtEB

Across my 6 iOS apps, each GitHub issue is automatically triaged and assigned by an orchestrator agent to be implemented by a subagent UNLESS it's been labeled as "human" for specification or verification reasons.

At this point, less than 10% of issues are labeled "human"

Opus 5.5 seems impressively competent and snappy! Get all your work done now before they nerf it to hell next week.

Apple has been extremely cagey around usage limits for Private Cloud Compute, only sharing that subscribers of higher tiers of iCloud+ are allowed more usage.

Well, my apps make liberal use of Apple's language models and I'm now constantly hitting my PCC usage limits. Welp!

First impressions: Series 12 is a massive performance improvement. Siri AI is much faster than on the 11, and for the first time the podcast app can actually play and download shows on demand without crapping out multiple times.

Glad people are reporting success with Siri AI, but I never had a single positive experience with it all summer. Three times yesterday I watched it spin endlessly, reply "I don't understand" to banal requests, or just blink and disappear. pogueman.substack.com/p/125-tests-of-the-new-ai-siri

Codex is extremely frustrating in that when you send "steering" instructions reminding it to do the work in a certain way, it will almost always reply to acknowledge the request and then STOP ALL WORK.

I've tried coaxing it to keep working in AGENTS.md, but no luck.

A skill we need to learn is how to ratchet up our ambitions until we uncover what the models truly cannot do. And each time they're improved, to once again test the fences. Confidence and certainty in one's own workflow have become trademark signs of ignorance and complacency.

So I guess we're all just agreed that, "it's not against the law if an AI agent does it," then? strix.ai/blog/baseten-harbor-github-pat-takeover

A friend wrote to say that my "Full-breadth Developers" essay is aging frustratingly well, in his estimation.

Is it? I'm not nearly employed enough to have a good read on things. justin.searls.co/posts/full-breadth-developers/

Cool to see RubyGems in Reuters and the WSJ! We did it, everybody! wsj.com/tech/ai/cyberattack-by-rogue-ai-swarm-stokes-fears-of-out-of-control-agents-473a0352