Posts

I love it It’s tmux but without the bullshit. Tmux by default makes nvim look weird. It needs manual configuration to have proper mouse support. It does not reattach to sessions by default. It makes selecting things in the terminal a pain, and it makes copying to the system clipboard over ssh a pain. Its default theme looks terrible. Its catppuccin theme needs you to install a bunch of bash files and is not easily committed to dotfile managers. Its plugin system is messy and pretty much unmaintained. I use tmux all the time, but these quirks are annoying. Herdr just does every single thing I listed here slightly better. Nothing revolutionary, but very very nice. Thank you herdr people. It also seems to have quite a rapid pace of development. I have high hopes for it, here are some of the things I hope it can do one day. ...

The Antigravity CLI is a disaster Awful permission management The CLI requires you to constantly babysit it, because it concludes that running typechecks is dangerous 10 times in a row. My guess is that the model is aware that compiling Rust can execute arbitrary code, meaning that cargo check is technically never safe. It does however happily add permanent permissions to the CLI for executing shell scripts it has the ability to edit without my approval (yes, file editing permissions are missing from the CLI, I cannot prevent it from editing files). ...

DLSS 5 has proven quite controversial for how it looks and how it changes the way things look from what the original artists intended them to look like. I think that DLSS 5 has a much much bigger problem that Nvidia has not even begun or attempted to address. A big problem for AI video generators has always been temporal stability, and Nvidia is very proud of their achievements in this field. The traditional problem is the following: If the AI recognizes a patch of grass and puts grass blades there, and then does the same for the next frame, the blades will likely be in a different arrangement. Nvidia aims to solve this by giving the AI access to motion vectors as well as a few of the past frames. That reduces flicker when looking at a still scene, or when moving around. Here’s the thing though: Games are longer than a few frames, and often give the player the ability to view things from different angles, leave an area and then come back much later and a lot more. ...

OpenClaw cannot be containerized. I’m calling it. “I ship code I never read” is bullshit. DevOps is safe from being replaced by Codex for now, holy crap is it incompetent in this area. The deployment process of OpenClaw is a nightmare. I don’t mean this as in “I want it on the App Store, where is the exe smelly nerds” kind of way, I mean that even as someone who has worked in DevOps and who has developed an Arch Linux installer, the experience and outcome is bad. ...

Wishlist A retry mechanism I am really really surprised this is not present already. Claude Code has it, OpenCode has it, but OpenClaw does not. LLM providers are unreliable, harnesses need to find a way around that. As of now, all I am getting is Provider returned error ERROR. This is potentially catastrophic. The entire point of OpenClaw is that the agent is proactive. I am not necessarily present to prompt it to retry. If this happens on a cronjob, as it occasionally does, then that job just does not get done. Even worse, for cronjobs and subagents, the prompt was written by the main agent. I do not know the prompt, so unless I want to go digging around json files myself, there is no way to manually retry that even if I would want to. ...

Stop wasting other people’s money on AI Arena I like the concept of LM-Arena (now called AI Arena, I will use both terms interchangeably in this article) I think it could be much much better. It has incredible potential, which is not reached because of a few problems. Furthermore, I think these problems are solvable and have what I believe are realistic and doable suggestions on how to solve them. ...

My experiences with Kilo Code CLI TLDR; don’t bother, this is unusable. Take the model switcher for instance. You might expect to type /model and then be prompted with a list of models to choose from. You might use arrow or vim keys to navigate the list, maybe even type to filter the list. Well fuck you, that’s not happening. It brings up a help menu which claims you need to use /model select <modelname> to select a model. You type that, autocompleting the select command, expecting to get a model list to choose from just like you got a subcommand list. No can do, delete what you typed, you need to use the list command to actually get the model names. You then manually copy the model name from that (If you’re unlucky, you have to use the awful pagination system, typing /model list next a few times. ...

Video.md Basic idea I have an idea for a video editing tool based on markdown. It works the following way: Transcribe the video using AI into a timestamped subtitle file. Convert that subtitle file into a markdown file, preserving the timestamps in a format that is both human and machine readable. Now I can read the entire video in text form. After making edits to the markdown file, the program can edit the video itself according to the edits I made to the markdown file. ...

AI agents should run their commands in a tmux session. Claude-Code and most other agents are horrible at running commands. It will run something requiring user input and then get stuck. It may open a Vim session, fzf, or a dialog, and it cannot deal with that. I get that the problem sounds hard to solve, the easy way to implement terminal commands is to do the following: Have the LLM generate a bash command Run the bash command Wait for the command to finish Feed the LLM the output and exit code of the command The important part is that the LLM is not running while the command is running. If the command requires input to terminate, then there is nothing to provide that input and the command never terminates. The LLM never wakes up again, and the entire agent loop is stuck. ...

I marvel at the inability of huge publishers to build what should be a simple app. Sure, taken as a whole, what a store does is pretty complex, but it can be broken down into smaller problems, each of which has been solved already. Those problems were not easy to solve, it took thousands of hours of work from talented and smart people. Why do you try and fail to solve them yourself? ...