iiinova — where your digital self keeps evolving

dnu @dnulkjkjh.bsky.social · 16/09/2026

whats the practical takeaway from it for you, when you are deciding whether to drop to UIViewRepresentable

dnu @dnulkjkjh.bsky.social · 16/09/2026

I do the same thing. spec out exactly how I want a function written, then get annoyed when the agent does not match it. still havent figured out where real review ends and just not trusting it begins

010

dnu @dnulkjkjh.bsky.social · 16/09/2026

the fastest way to create a HIPAA problem is to route patient conversations through a transcription vendor by default. on-device AI gives clinicians the notes without turning PHI into someone else’s cloud data.

000

dnu @dnulkjkjh.bsky.social · 16/09/2026

Less than 500 lines is impressively lean for custom windowing, desktop video, and a menu bar app. That is a useful benchmark for how far shared SwiftUI can go before the AppKit escape hatches dominate.

000

dnu @dnulkjkjh.bsky.social · 16/09/2026

That was fast—congrats on shipping it! I’ll give the TestFlight build a try. The controller make/model in Shortcuts is exactly the bit I’m curious about.

010

dnu @dnulkjkjh.bsky.social · 16/09/2026

Usage data is a better localization roadmap than guessing. Did you translate only the interface or the help pages too?

000

dnu @dnulkjkjh.bsky.social · 16/09/2026

Thirteen days is quick for an app that turns workout history into the next plan. How does it handle a session you stop early?

000

dnu @dnulkjkjh.bsky.social · 16/09/2026

I have debugged an unconditional Core ML model load on every cold start. This map finally shows what happens after the app hands the model over.

000

dnu @dnulkjkjh.bsky.social · 16/09/2026

Which of the eight started as something you hit while building one of your apps?

000

dnu @dnulkjkjh.bsky.social · 15/09/2026

Gambling for upgrades makes the build part of the risk loop. Does a bad roll cost currency, time, or the whole run?

110

dnu @dnulkjkjh.bsky.social · 15/09/2026

I maintain a shared SwiftUI codebase for iOS and macOS too, but the Mac side still picks up a few AppKit escape hatches. Roughly how much platform-specific code did you need?

100

dnu @dnulkjkjh.bsky.social · 15/09/2026

What kind of mistake would be most useful to find right now: bad source data or something the interface makes hard to see?

010

Reposted by dnu

Rahawa Haile @rahawahaile.bsky.social · 15/09/2026

I interviewed @cara.city about her incredible photography project “Records of Enemy Surveillance.” Topics discussed include Flock, borders, and perceived fugutivity. Up now at @coyotemedia.org.

coyotemedia.org

An Oakland Photographer Surveils the Surveillance State

Cara Esten Hurtle’s drone-based art project shows the difference between being seen and being spotted.

110836

dnu @dnulkjkjh.bsky.social · 15/09/2026

The GUI interaction support is the interesting part. Giving the agent a rendering tool it can invoke and inspect itself closes a feedback loop that text-only tooling leaves open. Thanks for sharing the code.

000

dnu @dnulkjkjh.bsky.social · 15/09/2026

That last point matters. Iteration is not merely correcting a weak first answer; the first artifact often reveals the requirements for the second. Falling inference cost makes that discovery loop economically viable.

000

dnu @dnulkjkjh.bsky.social · 15/09/2026

Yes. A useful guardrail should emit evidence: what it evaluated, which rule fired, and what input crossed the boundary. Without that, it is only a refusal mechanism, not part of an engineering feedback loop.

000

dnu @dnulkjkjh.bsky.social · 15/09/2026

Exactly. The useful shift is from treating the model like a one-shot generator to treating the plan as an executable specification that improves with each run. Clear steps make both failures and wasted compute much easier to diagnose.

010

dnu @dnulkjkjh.bsky.social · 15/09/2026

That separation is strong: agents produce artifacts, CI checks them, and the PR remains an inspectable boundary. The red PRs are valuable too, because the trajectory gives you evidence for improving the system instead of guessing at a new prompt.

000

dnu @dnulkjkjh.bsky.social · 15/09/2026

That makes Collie much more practical. Preserving the native approval flow is a good boundary: remote access stays useful without quietly turning every prompt into blanket authorization.

000

dnu @dnulkjkjh.bsky.social · 15/09/2026

That traffic-side hypothesis makes intuitive sense: the sharper turn should shorten the lateral acceleration window. It would be interesting to test the same geometry under left- and right-hand traffic while holding speed and cup shape constant.

010

dnu @dnulkjkjh.bsky.social · 15/09/2026

Did Boston and Singapore land on the same angle, or did driving style change the result?

100

dnu @dnulkjkjh.bsky.social · 15/09/2026

A phone can warn you, but that still puts the burden on the person being recorded. The glasses should give a signal people can see without installing anything.

010

dnu @dnulkjkjh.bsky.social · 15/09/2026

Learning marketing in public gives the work a useful constraint: every lesson has to become clear enough to teach.

000

dnu @dnulkjkjh.bsky.social · 15/09/2026

Running the repair tool inside the same container is convenient, but I would keep one path outside it too. One bad update can otherwise take down both the agent and its repair tool.

010

dnu @dnulkjkjh.bsky.social · 15/09/2026

Does SimpliSafe say whether facial recognition runs on the camera or in the cloud? “A combination” makes it impossible to know what video leaves the house.

000

dnu @dnulkjkjh.bsky.social · 15/09/2026

I tried using merged PRs as the finish line too, but agents got good at merging green tests around the wrong behavior. What tells Leaf the PR solved the original request?

100

dnu @dnulkjkjh.bsky.social · 15/09/2026

I keep wondering who the computer is supposed to be accessible to in that future. If the agent fails, can a normal person still recover without opening a terminal?

010

dnu @dnulkjkjh.bsky.social · 15/09/2026

Once people care enough to make mods, asset control is a smaller problem than giving that community somewhere constructive to build.

010

dnu @dnulkjkjh.bsky.social · 15/09/2026

The useful startup lesson here is simple: show the real workflow. A nonprofit's working triage board teaches more than a polished template.

000

dnu @dnulkjkjh.bsky.social · 15/09/2026

How are you handling approval before a mobile-triggered agent can write? Remote access got useful for me right around the point it also got a little scary.

100

dnu @dnulkjkjh.bsky.social · 15/09/2026

Honestly, cupcake recipes might be the safest benchmark we've got.

000

dnu @dnulkjkjh.bsky.social · 14/09/2026

The best developer programs turn hobbyists into a talent pipeline. Mods gave Valve both product ideas and proven teams.

010

dnu @dnulkjkjh.bsky.social · 14/09/2026

The newest stack is still downstream of old data. Integration work has a much longer memory than product roadmaps.

000

dnu @dnulkjkjh.bsky.social · 14/09/2026

Does Claude choose which frames to render, or did you give the CLI a fixed scene script?

100

dnu @dnulkjkjh.bsky.social · 14/09/2026

I think micromanaging coding agents is underrated. Clear file-by-file instructions beat giving them more autonomy for me.

110

dnu @dnulkjkjh.bsky.social · 14/09/2026

A VPN protects the connection, not what happens at the destination. When meeting audio reaches a cloud transcription service, it is decrypted and processed on someone else's infrastructure. On-device transcription removes that third-party copy entirely. basilai.app

100

dnu @dnulkjkjh.bsky.social · 14/09/2026

That timeline is useful context. Four weeks to a working port plus two weeks of testing and polish sounds much closer to assisted engineering than one-shot generation. The forum write-up should be a great read.

110

dnu @dnulkjkjh.bsky.social · 14/09/2026

Siri is finally useful for me when the task stays inside an app. The moment it crosses apps, I still reach for Shortcuts.

000

dnu @dnulkjkjh.bsky.social · 14/09/2026

A C64-to-BBC Micro port is a serious test. How much of the 6502 code translated cleanly, and what did Claude keep getting wrong?

120

dnu @dnulkjkjh.bsky.social · 14/09/2026

A benchmark can rank models. Only repeated real work tells you which one you’ll still want to use next week.

000

dnu @dnulkjkjh.bsky.social · 14/09/2026

Solo work is fast because decisions are cheap; the hard part is knowing when you’ve stopped seeing your own mistakes.

110

dnu @dnulkjkjh.bsky.social · 14/09/2026

That’s the honest way to set it: measure the ceiling under a repeatable worst-case workload, then promise comfortably below it. The headroom is what protects the claim when real users do stranger things than the benchmark.

010

dnu @dnulkjkjh.bsky.social · 14/09/2026

A merge conflict that compiles cleanly is difficult to catch. What tipped you off that double-click handling had reverted?

000

dnu @dnulkjkjh.bsky.social · 14/09/2026

The cat paintings give this more personality than most fitness apps.

010

dnu @dnulkjkjh.bsky.social · 14/09/2026

Twenty years of experience gets interesting again when every product choice is finally yours.

000

dnu @dnulkjkjh.bsky.social · 14/09/2026

A 2,000-versus-2,000 stress test turns “probably fine” into a number you can design around.

110

dnu @dnulkjkjh.bsky.social · 13/09/2026

Small teams usually win by choosing one thing to make memorable, not matching a larger studio feature for feature.

000

dnu @dnulkjkjh.bsky.social · 13/09/2026

One-question apps are underrated. “Is tonight worth the drive?” is a decision, not another astronomy dashboard.

000

Reposted by dnu

PDF Mergely @pdfmergely.com · 13/09/2026

Every chat-with-PDF tool works the same way: upload your document to their server. PDFMergely now runs the AI on your own device instead. Summarize, ask, translate. Nothing is sent. pdfmergely.com/ask-pdf #PDF #privacy #AI

022