Since activists and journalists seem to be the target audience for Columba, please consider discontinuing large language model use.

[radar92]

Most anti-war and anti-capitalist activists do not trust software with LLM-generated code, and will not use Columba as it is now. LLMs make developers seem less capable and less trustworthy, make projects harder to maintain, and make projects vulnerable to new supply-chain attacks. LLM data centers are poisoning the air and water of poor regions right now. Although they can be convenient, LLMs are not appropriate for security-focused software, especially for software intended for organizers.

I write this because I appreciate your work and want it to be useful for people organizing for peace.

[serialrf433]

This wont happen. The decision was made at the start of the development of this app. Most of Columba is sort of LLM-generated code.

If you find errors, you are free to open a PR and fix the error you have found: https://github.com/torlando-tech/columba/pulls
If you found errors and you do not want to send a PR on your own, you can open a issue https://github.com/torlando-tech/columba/issues and most likely a LLM would take a look how to fix your problem.

The high Columba development speed is because it is mostly LLM generated code.

To feed your anti-LLM position, take a look at the PR's in #567 , at the design incinsitency i mentioned in #501 and many other things. You see many LLM-related problems in Columba. Lets fix them.

There are for sure security problems. I think more then in Briar. There were also a ton of memory issues in Columba and you would for sure find some more.

Activists and journalists are not the main target of Reticulum and thus not Columba. If you for example communicate with a single hop, then you have no anonymity. If you transmit LoRa signals then of course you can be localized.

Activists and journalists should use Briar instead.

Where did you get the point 'activists and journalists seem to be the target audience for Columba'? Never have read that as the target audience of Columba.

The target audience of Columba are the masses. Journalists and activists are not the masses and that is one of the reasons why we have so many problems in the world. They are a tiny group and thus not the target audience of Columba.

[torlando-tech]

@radar92 Thank your for your concern. The following will be what I hope can be seen as a nuanced opinion on this topic. If you (or someone reading this in the future) are simply an absolutist with an anti-AI world view, then none of the following will likely make you feel any better. If you are open to hearing from an experienced developer who has adopted AI tools in the last 10% of his career, then read on.

First, to clear one thing up: seeing Claude as a contributor on a GitHub repo means precisely one thing: at least one commit in the history has Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> or similar (the email address in particular) at the end of it. That's it. If you really wanted, you could write a commit message yourself with this at the end, and Claude will show up as a contributor to your repo. Practically speaking, it usually means that the maintainer or a contributor of the repo at least had Claude Code write the commit message, and that the developer didn't first prompt Claude not to credit itself as the co-author. That's it. It is trivial to prompt Claude to not credit itself as the co-author, but I have elected not to do that in the spirit of transparency in development, precisely because I understand that AI use itself is controversial to say the least. You may infer here that the developer does more than just have Claude write their commit messages, and I'm sure usually they do, but I wanted to make clear the specific mechanism that makes Claude show up as a contributor in the first place.

While I do hope that Columba can help people of many walks of life, and in many different threat situations, it is not feasible for me to discontinue LLM use at this time without discontinuing development altogether. This is a passion project for me which consumes my free time on nights and weekends. It is built on open protocols, LXMF and Reticulum, for which there already exists Sideband, which is hand-written by the creator of the protocols himself. While it is true that Columba could not exist as it is today without Claude, the inverse is also true: Claude would not have woken up on its own and decided to make Columba, nor would it have been able to escape from the various pitfalls it seemed to set for itself throughout development, without me there prompting and steering it.

I have been a professional software developer for over a decade. During that time, I've had many roles, from sysadmin, to full stack web dev, to tech lead. I have mentored other developers, from fresh college grads to being team leads themselves. I have rewritten legacy systems by hand. I have optimized systems that serve millions of users. I have designed several large systems and many smaller APIs. I have not written an Android app before Columba, nor have I developed using BLE or any related IoT technology. These are new domains for me, and without the use of Claude Code specifically, I would likely still be struggling through quirks in the Linux BLE stack, or trying to learn everything up front regarding Kotlin and Android before having ever started. More realistically, I likely would have burned out months ago before ever having released Columba to the public, or talked myself out of being competent enough to achieve what I wanted.

Claude Code has been out just barely one year, and I have been using it full time at my day job for nearly that much time. I have used to to ship massive systems to production. I have been close enough to it for long enough to see where it falters, where it excels, and to get frustrated with it when it seemingly degrades overnight. I've also seen how much it has improved in such a short time. It is solely because of my exposure to Claude Code at work that I had the courage to start Columba and the BLE Interface before it.

I spend 5-7 hours every night after work on Columba. This time is spent mostly steering claude, preventing it from going down the wrong rabbit holes, iterating on plans for system design decisions, and most of all, manually testing the changes it makes. I sometimes get exasperated with Claude as it, for example, tries to @Suppress the very linting rule I had it create to stop writing unit tests that don't execute any app code, or create yet another polling loop for something that should be event driven. It is much more involved than "claude please make app good thanks" and shipping a release.

The downsides of heavy LLM use on my personal development are unsettling, to say the least. While I have been able to spend more time considering feature and system design (instead of learning syntax) there are now details that I don't have to learn in order to ship functional software. This is deeply unsettling to me, because for so much of my career I have derived confidence from deeply understanding the systems I am deploying. Now, I can only answer high level questions off the top of my head about the inner workings of Columba, and when someone asks about something in the weeds like "why is such and such dependency used" or "what version of xyz are you using" unless it's something I've already sat and troubleshooted myself, I either have to go read through the code or ask Claude (often the latter because it's so much faster).

The other trouble (for me at least) is that I am only human and have good days and bad days like everyone else. I am sometimes tired, sometimes sick, sometimes depressed. Those days, it is increasingly tempting to just let Claude run unsupervised developing a feature I want or fixing a bug I found. Indeed, I have done this, and it almost always results in it taking longer to ship the PR because manual testing ends up finding more issues and taking more iterations to get right. Counter intuitively, paying more attention to the LLM agent during development often ends up saving time in the long run, especially on more complex features. What's worse, is that if I get "lucky" and it works on the first (or even third) iteration from Claude, is that it's like missing a day of math class as a kid - everything builds on itself and its increasingly difficult to catch up later.

I am a naturally careful person, and as such, part of me will always find a reason to worry (about almost anything). The shape and trajectory of my career seems irrevocably changed by the advent of coding agents, and I hope that management continues to understand that without an engineer guiding the agent, the outcome is notably worse and often unusable. But will that still be true in 2027? 2035? I don't know. It seems that in the short term, getting proficient with these new tools will stave off my obsolescence by at least a few years.

Regarding the environmental impact of AI, I am really not educated enough on this topic to have an informed opinion. It does seem like my individual use of AI can only possibly be a drop in the ocean of all the AI use. It also seems like the rate of increased performance (or "intelligence") in these models is such that the energy required for a given level of intelligence is dropping rapidly, such that within a few years, there will be smaller models as effective as today's SOTA that could run on a <100W mini PC.

We all only have so much time here. Only so much time to live, to be with family, to pursue hobbies and passion projects, to appreciate life. As freedom and privacy are increasingly under attack around the world from people and institutions who will certainly use AI to its fullest potential, it is my stance that aiming some tiny fraction of the world's AI output at an open source app built on open (human created) protocols, with the goal of making Reticulum more accessible for normal people, and thus resilient privacy more available, is really not so bad.

You are free to disagree with me! Please, if you would like to see an Android LXMF application that is not developed with the aid of AI, develop one! If you would like a professional security audit performed on Columba, (maybe wait until it is 1.0 but) please commission one! I say this not to antagonize, but to encourage; a world with more options for Reticulum is a better world.