Mistral Vibe & Antigravity, Retro, HTTP/Vary, Critical Code, SSS
Vibe Coding - With a friend we tested Google Antigravity and Mistral Vibe CLI on different projects. Mistral Vibe CLI is excellent, but it requires quite detailed instructions split into smaller tasks for execution. If you'll give it too high level instructions it seems to go into inefficient loops and create absolute and complete mess.
For very simple testing projects, both received instructions like making a countdown timer that runs N minutes until zero and shows a NOW message when time is up. When < 50% of time is left, text should be orange and when 25% is left, it should be red. It should also use a large font for the CLI, like ASCII text or a banner. Also include --test option, which allows the code logic to run without the delay, for testing the logic. All of that worked great as single shot development task. No, it didn't generate code all at once, but I enabled auto accept mode, which allowed it to work around problems, which it successfully did!
Fun times, after the initial test program was written in Python. I just asked Mistral Vibe CLI to port it to Go, Rust and PowerShell for posterity. It required several refinement steps, so the Devstral-2 model doesn't do everything that's necessary. Probably trying to save resources, yet that leads to invalid outcome. Always make sure that the TEST PART and WHAT TO TEST and HOW TO TEST is quite detailed.
Yet it did some interesting things, because Python version used pyfiglet to produce large ASCII font, it also prompted the font to PowerShell, Go and Rust, because I said it needs to work similarly, and there was no matching library available.
- Fun times. And when it did that, it first messed up some numbers and then it messed up with white space. I had to again say, that you have to make sure that the output exactly matches, whatever it takes. Make sure that the white space is also matching, and if necessary, generate new test functions so you can easily test the digital rendered "font" outputs. Also as expected, escaping and special characters caused also secondary problems. Oh joy, I've often said at work, that escapes or sanitization is more often wrong than right. So... Yeah... Not unexpected at all.
Whitespace seemed to be really hard for it. Maybe some LLM layers strip output for efficiency again. So you'll have to ask it to dump the output to file and xxd it or something like that, otherwise it misses some of the whitespace from outputs... - Legendary.... It just ran loops before I said to use xxd to get the thing right...
Retro - People making retro videos with a terminal and smooth scroll - skip that. They belong to the same club as those applying a VHS effect to videos. If you've repaired a few VHS decks, you know very well that isn't a true VHS effect. There are fundamental reasons why the hardware behaves the way it does. As usual, everything affects everything. The same applies to simulated camera views with microprism rings or split-image rangefinders. You can't just "add something" in the middle and think that's it.
HTTP Vary (@ Wikipedia) - The header, I think I've not been always using it when it should have been necessary. Oops. Well, now I've reminded myself about it. At least it didn't lead to any known issues. I just figured it out when checking zstd-related compression stuff with HTTP/3 and Nginx.
Critical Code - Watched / Attended - Critical Code: Leading Secure Software Development – Webinar 2025-11-18 (Kriittinen koodi: Turvallisen ohjelmistokehityksen johtaminen -webinaari 18.11.2025)
SSS (@ Wikipedia): ssss, shamir-mnemonic (SLIP-39) and K-of-N-XOR-Secret-Sharing (@ GitHub) - Lovely. It seems that there's no single widespread standard for secret sharing either. I guess the shamir-mnemonic is the best option popularized by cryptocurrencies, but otherwise it is a known problem with known solutions. But tooling and standards are really lacking. Yet I don't understand why that solution is made so limited on purpose. There's no reason to fix the K=3 and N=5 at all.
2026-05-24