Zsolt Ero

OpenFreeMap survived 100,000 requests per second I was about to post about how nice the last 10 months of OpenFreeMap have been. The architecture has really proven itself to be great, Cloudflare has agreed to sponsor the bandwidth, Hetzner servers are super stable as always,

2025/08/08 23:08:16 [crit] 1084275#1084275: *161914910 open() "/mnt/ofm/planet-20250730_001001_pt/tiles/8/138/83.pbf" failed (24: Too many open files) ... This is weird. I've never seen anything like this. I check nload, and it shows huge traffic.

OpenFreeMap survived 100,000 requests per second I was about to post about how nice the last 10 months of OpenFreeMap have been. The architecture has really proven itself to be great, Cloudflare has agreed to sponsor the bandwidth, Hetzner servers are super stable as always,

One US client sent us a USD paper check in the mail, instead of the wire transfer / ACH we'd asked them to use. Now I have this piece of paper for $1100. What can I do with it? I'm in the EU and I have no idea where I can possibly cash or deposit this check.

Right now, you might have read the news that a Chinese paraglider pilot, Peng Yujiang, got sucked into a cloud and ended up at 8,598 meters / 28,000 feet.

I think the other two scenes might be real, though, recorded with a 360 camera on a selfie stick - a common setup for paragliding videos. There are definitely some questions though... For the full investigation, read the full article here.

Right now, you might have read the news that a Chinese paraglider pilot, Peng Yujiang, got sucked into a cloud and ended up at 8,598 meters / 28,000 feet.

I've been paragliding for 18 years and follow the progress of AI very closely, but even I couldn't say with 100% certainty if this video is real or fake. One scene is definitely fake: the one where the camera makes a move that would only be possible from a drone.

Right now, you might have read the news that a Chinese paraglider pilot, Peng Yujiang, got sucked into a cloud and ended up at 8,598 meters / 28,000 feet.

Honest question: How do you pay using your iPhone while on a call? I mean the one where I double-click the side button and pay at a cashier?

o3 also gave a nice answer, explaining that it's a change introduced in iOS 16:

Honest question: How do you pay using your iPhone while on a call? I mean the one where I double-click the side button and pay at a cashier?

OK, I asked Gemini 2.5 Pro, o3, Claude Sonnet 4, and Perplexity. Gemini gave a perfect answer! It explained why it happens: an accessibility setting was inverted on iOS 16, and many users got caught by this. I personally never touched this, so I'm surprised by its default value.

Honest question: How do you pay using your iPhone while on a call? I mean the one where I double-click the side button and pay at a cashier?

Crossposter final touches Multi-day threads. I can now continue a post after it has been posted. First, I create post1.md, then later as I progress post2.md, etc. After each post, the status is saved into a JSON, and when the next part is posted, it continues the original thread.

Different handling for X Premium and X Free accounts. X Premium has 25000 max char limit; X Free has 280. Moving the link on X from the first post to the first reply. If there are multiple links, I currently move everything into the first reply, but it might make sense to have one link per reply,

Crossposter progress update: I just remembered that I cannot write "Bluesky" on X, the land of free speech. Asked Gemini how to make similar-looking Unicode characters, which look the same but are actually not the same string.

Crossposter progress update: I was using regex-based parsing for links, images, @handles, etc. It started out simple but became increasingly confusing as I tried to synchronize this regex-based parsing for Bluesky posts with the optimal splitting code.

It supports full Markdown-style links, so the link text doesn’t have to be the URL itself (e.g., https://...), but can be proper HTML-style links with actual text. It's funny, because there's no way to do this using the official web client - but it’s definitely possible via the API.

Crossposter progress update: I was using regex-based parsing for links, images, @handles, etc. It started out simple but became increasingly confusing as I tried to synchronize this regex-based parsing for Bluesky posts with the optimal splitting code.

I decided to go all-in and rewrote everything using a proper Markdown AST parser. I had to make a plugin for @handle support, but the mistune library was actually quite nice to work with.

Crossposter progress update: Optimal splitting into threads works! It was a big one. Turns out I need dynamic programming to find the optimal splitting points in a thread. What does an optimal splitting point mean?

Try to split on paragraph boundaries, then on sentence boundaries, then on commas, etc., then inside a sentence. But never split inside links, images. And make sure the last segment is still a meaningful post, not just a leftover fragment.

Looking at @astro.build source code be like: returning a template string of JS code, which imports template strings of image paths, which maps regex-parsed HTML matches to async image objects, all while JSON-parsing a stringified regex-escaped string!

Moving all my blogs over to @astro.build. Substack is getting more and more bloated; it feels like it's becoming another Medium in the long term. @ghost.org is painful with Markdown-based workflows. They started with Markdown but went full-on with a WYSIWYG editor.

Crossposter progress update: Posting on bsky works! Amazing DX and super friendly dev community on Discord. Posting on X works! Terrible DX, scary low quota + returns wrong quota information in response headers. I have no idea when it blocks me and for how long.

Also, interestingly the most popular libraries haven't been updated to support the current v2 endpoints, so had to research everything and implement it myself with requests_oauthlib. I understand why developers abandoned this platform though.