275. Blue Dusk 中
The cover was shot on Friday evening while I was out for dinner. I looked up and saw this blue dusk sky. I really like this blue, the kind that feels comfortable to look at.
Recording down-to-earth trending tech I see every week, filtered and published here. Follow this weekly newsletter to get update notifications
Trending Tools
Bento: stuff an online slide deck into a single file
https://github.com/nyblnet/bento
The idea and the execution here are both nice. Animations, video, web design, all of it goes into one HTML file, and that file is at the same time your document, your editor, and your player. It’s also a good fit for letting AI keep editing and iterating on it. Worth a try if you’re curious.
peek-cli: let your coding agent see web pages properly
https://github.com/puffinsoft/peek-cli
peek-cli lets an agent grab a screenshot of any tab you already have open in your browser, which suits those frontend design loops where you keep iterating until the result looks right.
clawk: give your coding agent a disposable Linux VM
https://github.com/clawkwork/clawk
If you’re worried that what your AI coding agent runs could damage your machine, or you really love --dangerously-skip-permissions but don’t want to leak your keys, give this a try. It automatically spins up a disposable Linux VM so nothing goes wrong while things are running.
It’s been a while since I shared a Kaku update
https://kaku.fun
Kaku is the open-source Mac terminal I built over Chinese New Year by heavily customizing WezTerm: AI friendly, good looking by default, and fast. It’s up to 23 releases now. It started out very simple, and it still looks very simple, but it keeps getting nicer to use. I also added an official site and friendly docs.
I use it as my main workstation. Compared to GUIs I still prefer TUIs because they’re simple and direct, so roughly 90% of the code in my open-source projects and in Mole is written through Kaku. I’ve started deploying some of my own things on cloud servers lately, which finally pushed me to round out the SSH experience, and over the past month I used Fable 5 to optimize a lot of the low-level Rust rendering, so startup is much faster now. I like one tab per project, with a shortcut to split that tab into two panes so I can code on one side and test on the other. Switching and closing tabs works almost exactly like a browser, since I ported a lot of the browser habits you already have over to the terminal.
On the AI side, Kaku stays out of your way. It usually shows up when a command fails, or when you’ve forgotten a command and just type #plain English, and it tells you the real one. When you don’t feel like opening codex, CMD+L brings up an AI chat panel so you can talk to it right there, which makes the terminal a naturally AI-friendly environment. The last two releases mostly fixed usability issues: AI usage, quick tab closing, Starship compatibility, working over SSH, plus pane navigation, tab renaming, and other details. If you want something simple and fast and you like things that look good, give Kaku a try. Suggestions from actual use are very welcome and help make it better.
Just Looking Around
Andrej Karpathy’s case for dumping one long ramble on the AI
https://x.com/karpathy/status/2079610838143623371
He describes a way of working with an LLM that he finds genuinely useful: lean back in your chair, switch to voice input, and ramble for ten minutes or so, messy, jumping wherever your head goes. He sometimes opens with a note like “I’m on speech recognition, please excuse the typos.” For whatever reason LLMs are very good at pulling the content back out of a long, messy ramble, and the version they hand back to you is often much clearer than the one you just said out loud, so the rest of the conversation goes smoother and there’s less to correct.
Dan Shipper’s take on Claude Opus 5, which I mostly agree with
https://x.com/danshipper/status/2080700057892815114
They spent a week testing Opus 5 across coding, writing, knowledge work, and their own agents, and their first reaction was that it’s a hard model to like: it works against your instructions, stops before the job is done, and doesn’t fit the skills and plugins they already had. Then they deleted the old skills and rebuilt from scratch, and it got a lot better, so his advice is to not force it into an existing workflow, and to run it at medium or low thinking rather than maxed out. He still sends the hardest work to Fable and everything else to GPT-5.6, and he calls Opus 5 a “poor man’s Fable,” a model with the temperament of a genius but not the ceiling, stuck in an awkward middle spot.
5 Trends That Defined AI Engineering at World’s Fair 2026
https://www.latent.space/p/aiewf26trends
This piece ties harnesses, loops, skills, enterprise deployment, and coding agents into one coherent roadmap. Worth a read.
A random window somewhere in the world, opened for you
https://www.window-swap.com/Window
This site is fun. People around the world share the view outside their window, and you can switch randomly and just look around. I had it up as my wallpaper all afternoon, and it looked great.
