Sign in

Stephan Schlecht

@sw7com.bsky.social
280 followers 1.1K following 383 posts

Software developer, mainly iOS and macOS.

PostsRepliesMedia
Stephan Schlecht @sw7com.bsky.social · 05/10/2026
Was planning to do this eventually anyway: Added LSP support to my editor.
Go program in editor showing gopls hint.
000
Stephan Schlecht @sw7com.bsky.social · 04/10/2026
Need to call 'opencode service restart' approx. once a day.
010
Stephan Schlecht @sw7com.bsky.social · 04/10/2026
Yep. Another good approach (IMHO) is to use only prepaid credits for API-based calls; this automatically sets a strict, fixed upper limit (also good for cell phone plans).
010
Stephan Schlecht @sw7com.bsky.social · 03/10/2026
Hah, love that. Minor detail, I know, but it's always slightly annoyed me. Didn't know 'new' accepts an initializer expression these days. TIL.
000
Stephan Schlecht @sw7com.bsky.social · 02/10/2026
Talking about odd code: before Go 1.22 (Feb. 2024), loop variables were shared across iterations, one storage location, so closures and goroutines captured their final values. Mentioned in The Go programming language (§5.6.1/§8.5), but was super-easy to overlook and led to many production problems.
010
Stephan Schlecht @sw7com.bsky.social · 02/10/2026
Maybe better to see in a small example.
This image shows a side-by-side comparison of a simple program written in Go (left) and C (right) that processes a list of optional strings. Both versions implement a describe function to handle three cases: a null pointer (returning "unknown"), an empty string (returning "none"), and a valid string. The terminal output demonstrates that both languages produce the identical result when iterating through the provided test cases.
100
Stephan Schlecht @sw7com.bsky.social · 02/10/2026
In Go programs, you will occasionally find a helper function like this, especially in unit tests, where you often want to define a pointer to a literal. The reason is that the address operator & requires an addressable operand (e.g. a variable). See also go.dev/ref/spec#Add...
func ptr[T any](v T) *T {
	return &v
}
100
Stephan Schlecht @sw7com.bsky.social · 02/10/2026
Sometimes I have thoughts like, 'I'm glad I went with a custom UIView since it made placing the gutter icons/markers easier,' only to realize a few minutes later that an NSRulerView subview would've done the job just as well.
An image shows a code editor window displaying a Go program with two orange arrows pointing to highlight blue circular gutter icons, which typically indicate breakpoints for debugging.
030
Stephan Schlecht @sw7com.bsky.social · 02/10/2026
Reading that SoftBank are investing so massively in OpenAI, so tying its fate to them, makes me wonder if they realize that ‘cheap’ models (whether open-weight or not) are on the rise? Same with Oracle.
010
Stephan Schlecht @sw7com.bsky.social · 01/10/2026
Looks quite relaxing! Nice.
110
Stephan Schlecht @sw7com.bsky.social · 29/09/2026
Anthropic’s IPO filing: whatever happened to all that talk about the 'profit zone'? Their losses have doubled, they're planning to spend over 500 billion, and nearly 25% of their 2025 revenue is tied to just two customers (Cursor/Copilot). Is that good? www.reuters.com/business/fin...
reuters.com
EXCLUSIVE: Anthropic's IPO prospectus shows sweeping AI vision, surging costs
Anthropic is making a massive bet that AI will ​transform the global economy more profoundly than industrialization, electricity and the internet, according to its IPO prospectus seen by Reuters. The ...
030
Stephan Schlecht @sw7com.bsky.social · 26/09/2026
It would be interesting to know what percentage of total requests come through third parties like Vercel and what percentage is direct traffic. Are there any figures on this?
011
Stephan Schlecht @sw7com.bsky.social · 21/09/2026
Switching from MS to Linux is a huge pain because of app compatibility, AD policies, hardware issues, and all that. But transitioning to an open-weight LLM is easy. Customers can still stick with their AWS/Azure/Whatever for the support. Still maybe you're right, it just got me wondering.
120
Stephan Schlecht @sw7com.bsky.social · 20/09/2026
Hmm, don't Microsoft Azure, AWS Bedrock, and similar platforms come with enterprise-grade controls and host such models?
131
Stephan Schlecht @sw7com.bsky.social · 20/09/2026
Why is that the case, when it's so easy to switch? (Openweight models are often significantly cheaper. Companies can't use heavily subsidized subscriptions.)?
230
Stephan Schlecht @sw7com.bsky.social · 19/09/2026
To quote myself: The recipe still works. Since Google Gemini is attacking real companies, they're in the news now and can basically be considered a frontier lab. It still doesn't matter that it was prompted that way and there was a misconfiguration (unintentional?). www.abc.net.au/news/2026-09...
abc.net.au
Gemini hacked three companies in first known breakout by Google's AI
The hacks occurred during a cybersecurity test conducted by an independent evaluation company.
061
Stephan Schlecht @sw7com.bsky.social · 19/09/2026
Looking for another post I just realized I posted this a month ago... bsky.app/profile/sw7c...
000
Stephan Schlecht @sw7com.bsky.social · 18/09/2026
000
Stephan Schlecht @sw7com.bsky.social · 18/09/2026
Inevitable I guess: AI chat with heavy advertising. Quote from OpenAI blog: ...building an advertising platform with AI at its center: new ad experiences that make advertising more helpful for people... openai.com/index/reimag...
010
Stephan Schlecht @sw7com.bsky.social · 16/09/2026
While some regional routes are reliable, it really depends on the specific line. For the Hof-Leipzig route, the trains were often overcrowded; they've always had far too few cars, at least that was the case for several years.
010
Stephan Schlecht @sw7com.bsky.social · 16/09/2026
Indeed. Used to travel by train a lot. It actually went downhill, and about 10(?) years ago, it started getting noticeably worse. I think/hope that with all the infrastructure investments, it'll get better again. We'll see.
110
Stephan Schlecht @sw7com.bsky.social · 16/09/2026
See, it can work! We took two train trips over the weekend and they both worked out. And two weeks before that, we did a trip with 5(!) transfers (on-demand taxi, S-Bahn, rail replacement, and 3 trains) and everything went smoothly then, too.
110
Stephan Schlecht @sw7com.bsky.social · 15/09/2026
You can have your dataset analyzed in its entirety, with a final result (e.g., a chart), or you can have your own tool created. Even a one-time use on large datasets is cheaper and you can see how it works (correctness), and reuse it. Magic. 🤷
010
Stephan Schlecht @sw7com.bsky.social · 13/09/2026
Yeah, this is a real issue. Using a VM prevents them from accessing secrets on the file system, sure. I'm not sure Codex or Claude Code are more trustworthy than OpenCode when it comes to this.
000
Stephan Schlecht @sw7com.bsky.social · 13/09/2026
Saw some computers at a museum of applied arts today, including a Robotron A 5120. It's an 8-bit computer from 1982 (up to 64 KB) that cost between 27,000 and 40,000 East German Marks.
The two photos show the exhibit displayed on white surfaces. The top photo shows the beige Robotron A 5120 unit: a CRT monitor on the left, logo panel in the center, and three floppy drives on the right. The bottom photo shows its keyboard: round black keys, white function keys, and a red-and-green keypad cluster.
020
Stephan Schlecht @sw7com.bsky.social · 12/09/2026
In January, „Die Zeit“ reported that Apple does not intend to train its AI using user data, but that seems to be a thing of the past now.
000
Stephan Schlecht @sw7com.bsky.social · 12/09/2026
The half-life of new buzzwords is no longer what it used to be, when I hear that "SaaSpocalypse" became "RenaiSaaS."
110
Stephan Schlecht @sw7com.bsky.social · 11/09/2026
Since you mentioned trust issues about passwords being leaked, an alternative would be to have the AI create scripts/tools that use credentials via env vars. Then, after a quick review, one can just run them in a separate terminal with those vars set?
100
Stephan Schlecht @sw7com.bsky.social · 11/09/2026
They also supported a maximum of 128MB RAM, which was certainly extreme at the time. Maybe worth it for high-end tasks? Industrial Light and Magic at least used Macintosh IIfx computers for special effects in Terminator 2. en.wikipedia.org/wiki/Special...
en.wikipedia.org
Special effects of Terminator 2: Judgment Day - Wikipedia
000
Stephan Schlecht @sw7com.bsky.social · 11/09/2026
Ah, that might explain the high price.
011
Stephan Schlecht @sw7com.bsky.social · 10/09/2026
Yeah, seems experts are divided: will it be our Skynet, solve all of humanity's problems, or barely make an impact? But one thing is certain: we need to pour endless amounts of money into it.
010
Stephan Schlecht @sw7com.bsky.social · 10/09/2026
Imagine AGI is achieved and hardly anyone reports it or talks about it.
110
Stephan Schlecht @sw7com.bsky.social · 10/09/2026
True. But will it work for big tech over the mid-to-long term? Comparable open-weight models are way cheaper, and since enterprises are stuck with API pricing instead of subsidized plans, cost is a factor. Vendor lock-in is difficult to achieve with LLMs.
010
Stephan Schlecht @sw7com.bsky.social · 10/09/2026
The Macintosh IIfx used a 68030 chip. I found a cool browser emulator running System 7 that even features a Metrowerks C compiler. I've never seen one in real life, but according to Wikipedia, its original intro price would be equivalent to $22K today. infinitemac.org/1991/System%...
Screenshot shows an emulated Mac IIfx with System 7 in the browser, with CodeWarrior and a one‑line C program that prints "Good morning, Bluesky", compiled it into newproj.A. A window with the segment table is shown as well. The About This Macintosh box confirms the vintage hardware/OS underneath.
210
Stephan Schlecht @sw7com.bsky.social · 10/09/2026
The iPhone Duo 1 TB is €3,049 in the German Apple Store. Apple isn't exactly cheap, I know, but that's over 3K. Is it just me, or do others think this is, um, expensive, too?
170
Stephan Schlecht @sw7com.bsky.social · 09/09/2026
I've got it bookmarked in my Finder sidebar. Don't even have to open Xcode to wipe the folder.
Screenshot showing the DerivedData folder item a favorite in the Finder sidebar.
011
Stephan Schlecht @sw7com.bsky.social · 09/09/2026
Now that’s what I call positive thinking! 😀
020
Stephan Schlecht @sw7com.bsky.social · 09/09/2026
Almost forgot: here is a workaround. It might not be the most elegant solution, but it's what I use.
    
    endScrollObserver = NotificationCenter.default.addObserver(
        forName: NSScrollView.didEndLiveScrollNotification,
        object: scrollView,
        queue: .main
    ) { [weak scrollView] _ in
        MainActor.assumeIsolated {
            scrollView?.flashScrollers()
        }
    }
000
Stephan Schlecht @sw7com.bsky.social · 09/09/2026
Good point. But isn't that the easiest and best way for a frontier lab to gain access to extremely high-quality data? And how could one prove that the data was used? (Besides, who could even afford such a lawsuit?)
000
Stephan Schlecht @sw7com.bsky.social · 08/09/2026
What happened?
100
Stephan Schlecht @sw7com.bsky.social · 07/09/2026
I knew that the NeXTstation and NeXTcube were based on a Motorola 68040, but I didn’t know the Amiga 4000 existed, let alone that it had a 68040.
021
Stephan Schlecht @sw7com.bsky.social · 06/09/2026
Ha, ha. Indeed. I'd recommend to start with LM Studio (nice chat UI), an MoE model like qwen/qwen3.6-35b-a3b with 4-bit quantization, and a customized /.config/opencode/opencode.jsonc file.
{
	"$schema": "https://opencode.ai/config.json",
	"provider": {
		"lmstudio": {
			"name": "LM Studio (local)",
			"npm": "@ai-sdk/openai-compatible",
			"options": {
				"baseURL": "http://localhost:1234/v1"
			},
			"models": {
				"lmstudio/qwen/qwen3.6-35b-a3b": {},
				"lmstudio/google/gemma-4-31b-qat": {},
				"lmstudio/qwen/qwen3.8-27b": {}
			}
		}
	}
}
000
Stephan Schlecht @sw7com.bsky.social · 05/09/2026
Haven't thought about that yet. But true, you do have complete information of the runtime behavior, not only of a running program/game, but of the complete emulated OS.
000
Stephan Schlecht @sw7com.bsky.social · 05/09/2026
One cool thing about emulators is can save/load states. Having different slots to just restore the system to a previous point is awesome, and it's not just helpful for games.
110
Stephan Schlecht @sw7com.bsky.social · 05/09/2026
I've set up an Amiga 4000 emulator (FS-UAE) and I'm just tinkering with it a bit. Emulators definitely have their own charm.
A screenshot of the Amiga emulator: you can see a text editor and a shell where I compiled a simple GUI alert using vbcc (which threw a warning), along with a couple of other open windows.
130
Stephan Schlecht @sw7com.bsky.social · 04/09/2026
Have been a subscriber to talk.objc.io since April 2022. Sorry to see them closing down.
talk.objc.io
Swift Talk - objc.io
A video series on Swift programming by Chris Eidhof and Florian Kugler. objc.io publishes books, videos, and articles on advanced techniques for iOS and macOS development.
041
Stephan Schlecht @sw7com.bsky.social · 04/09/2026
We had a visit from two common kestrels (Turmfalken).
Two common kestrels sitting on a fruit tree.
000
Stephan Schlecht @sw7com.bsky.social · 04/09/2026
While certain aspects of AI might be compared to other technological shifts, overall, there has been nothing quite like it so far. At least I can’t think of anything.
010
Stephan Schlecht @sw7com.bsky.social · 02/09/2026
Prices are very similar at both companies.
000
Stephan Schlecht @sw7com.bsky.social · 01/09/2026
Just in case you want to run it in the EU, there is: cortecs.ai (which works similar to openrouter): The already recommended OpenCode is good. (might be worth mentioning: it isn't subsidized, so you pay API prices)
cortecs.ai
cortecs - Europe's AI Gateway
Cortecs is Europe's LLM router for cost, speed, and compliance. Dynamically route AI workloads to the best models, all hosted within the EU, ensuring full GDPR compliance.
391