Tim Janik (@TimJanik) header
Tim Janik (@TimJanik) profile picture

Working on jj-fzf (Jujutsu TUI), Anklang (DAW), Imagewmark, Audiowmark; https://teapawt.org/t.co/FCVYmOjgsW

Hamburg
Joined October 2011
Photo by Tim Janik (@TimJanik)Photo by Tim Janik (@TimJanik)Photo by Tim Janik (@TimJanik)Photo by Tim Janik (@TimJanik)Photo by Tim Janik (@TimJanik)Photo by Tim Janik (@TimJanik)Photo by Tim Janik (@TimJanik)Photo by Tim Janik (@TimJanik)Photo by Tim Janik (@TimJanik)Photo by Tim Janik (@TimJanik)
CMSCOUT 0.2.0 is out! 🚀 🔍 Code-motion aware diffs for TS/JS/TSX/JSX, Go, Bash, C & C++ 📦 Matches functions, classes, methods, imports, macros, JSX across renames/moves 📊 Per-block similarity + change summaries ⚙️ Git external diff support + --summary #100DaysOfCode
1
1
64
Agentic tip: Next time your LLM messes up, tell it: "Turn my rejection into a lint rule." It might just manage to do it *and* clean up every related bad pattern too. 🤯
5
499
Two weeks did pass super fast again and it is time to discuss AI news and lessons learned / best practices in a space. 👇 teapawt.org/i/spaces/1DxleVQRBeRKL

5
1
4
242
Looking forward to talk later today to my AI buddies and everyone else who is interested in these topics. 👇 teapawt.org/i/spaces/1nGnRBXQVDYGO

4
1
4
290
Wall of text from an LLM? Commit messages turning into novels? Use this 3-word magic spell: ⮞ shorter, no jargon Works instantly. Try it now, thank me later…
2
70
Unbelievable, Qwen3.8-Flash-Next tops DSv4Pro 0813 on the Engineering Index and runs on an RTX 4090 with ~20tps (Q3kxl, 128k context, system with 96GB RAM). And it has vision… But, asking it about this chart, the model found the most interesting score to be the 27B xhigh ;-)
166
LLMs will happily rearrange code, so that `diff -u` become useless for review. I now use cmdiff that matches methods after moves/renames, then diffs the paired blocks. Suddenly you can actually review the changes instead of reconstructing them by eye: github.com/tim-janik/cmdiff
58
Enjoyed reading Florian's article on being sick of AI writings. I could actually focus reading it without the brain drifting away. That's my main tell for AI texts, the phrases are just so empty, I can't focus on them. This explains why. cyb3rops.medium.com/im-sick-… thx @cyb3rops
1
1
2
86
I asked DeepSeek V4 where to get an MT7902 WLAN driver and went to scroll X for a bit. Checking back, it had backported mainline support onto linux-6.8.0 and left it ready to load 😲 Now posting via WiFi, thanks to an LLM that does out-of-tree driver ports unprompted… 🛜🛜🛜
5
121
My Tuxedo Stellaris 15 Gen3 cost 2800€ in 2021 3 months ago: replaced the 4TB M2 Last month: entire laptop died RAM prices are insane, so I got a Minisforum M1-1390, dropped in the 64 GB RAM + SSDs Ended up with more cores and higher clocks! Total replacement cost: just 400€
1
2
79
Qwen3.8-27B is Great! The @UnslothAI UD-Q4_K_XL gguf at KV8 hits 44 t/s on my RTX 4090. Reasoning is abbreviated (Neanderthal speak), but the final answers come out eloquently and thorough. Huge leap for local use 🤩 teapawt.org/Alibaba_Qwen/status/20…
We promised open weights for Qwen3.8. Now, time to meet them! 🎉 ⚡ Qwen3.8-27B: - A native multimodal dense model. With just 27B parameters, it outperforms Qwen3.7-Plus overall and shines in real-world coding & office workflows. - 262K native context, easily extendable to 1M tokens via YaRN. - Built for builders. Highly efficient, high-quality, and licensed under Apache 2.0. 🚀 The open weights for Qwen3.8-2.4T-A95B (Max-level) have also been released recently. Whether you're shipping lightweight applications with Qwen3.8-27B locally or building agents with Qwen3.8-2.4T-A95B, they're yours now! Download, deploy, and build something we haven't imagined yet. 👀👇 - Hugging Face: huggingface.co/collections/Q… - ModelScope: modelscope.cn/collections/Qw…
2
116
My Stellaris 15 Gen3 (@TUXEDOComputers) died the other week, so I’m back on my old @Lenovo T470p. It’s not as fast, but I’ve really missed its build quality, keyboard haptics, and proper physical mouse buttons (all three of them!) ;-)
36
New jj-fzf release: 0.42.0 ✨ ⚔️ Alt-L: launches jj resolve for interactive conflicts 🔄 Ctrl-R: fresh change ID via jj metaedit 📜 Ctrl-O: shows ALL oplog changes 🤖 Cleaner LLM commit message generation Notes & downloads 👇 github.com/tim-janik/jj-fzf/… #100DaysOfCode
1
82
1/ Want to edit "immutable" commits in jj-fzf? When a repo is mirrored to GitHub as "origin", commits become immutable via immutable_heads(). That’s great for public history, but annoying for private repos where you still want to reshape commit history.
1
90
2/ Using jj, immutability is just a revset, so you can override it. Create a one-shot config for jj that empties the set: echo -e '[revset-aliases]\n"immutable_heads()"="none()"' >/tmp/jjmutableheads.toml
1
42
3/ Then launch jj-fzf with the override stacked on top of the user config: JJ_CONFIG=~/.config/jj/config.toml:/tmp/jjmutableheads.toml jj-fzf Now every commit is editable for that session only. Recommended for private repos only.
33
Combining DeepSeek V4 Flash + Luna beats either model alone 🤔 That means you can approach SOTA agentic work for free by combining DSv4F-free on @opencode for coding and Luna on github copilot-free/pro credits for review… Testing this on a new project atm ✨
1
150
Combining DeepSeek V4 Flash + Luna beats either model alone, according to this @togethercompute analysis: teapawt.org/togethercompute/status…
We analyzed DeepSeek V4 Flash and GPT-5.6 Luna on DeepSWE. A DeepSeek-first cascade with test-suite verification solved MORE tasks than Luna alone at 37% lower cost per task.
72