LLM disclosure

[ProactiveServices]

Hello, I notice that the repo contains several detailed files relating to claude, gemini and copilot. I couldn't find any LLM-related disclosures and thought it might be worth explicitly mentioning LLM usage for a project which provides such a sensitive function as remote access to other systems. Cheers.

[bitbound]

Hey, @ProactiveServices .

(Sorry for the long reply!)

This is actually the first time I've heard of having an official disclosure for LLM, apart from the commit and PR history that gets added automatically. Do you have an example of one that you think is documented well and could be used as a template?

In the meantime, I'm happy to share my both my usage and personal thoughts on the topic here.

Personally, I'm extremely worried and frustrated. I'm afraid for my own career and ability to continue providing for my family in the future. I'm afraid for all the other workers around the world who face similar uncertainty. I'm afraid that LLMs will ultimately serve to further widen the ever-growing wealth gap in our society. I'm afraid it will be used for propaganda, rewriting history, and altering public opinion. Basically, I don't think its existence will lead to anything good for normal people, even though it was built using the cumulation of public human knowledge. I'm also frustrated at the rising costs of hardware due to AI datacenters.

I was one of the last holdouts to adopt LLM usage at my previous job. At first, it was out of pure stubbornness and skepticism. Plus, I'm really picky about conventions and cleanliness. But after seeing that the hype wasn't dying down and that this might be the new normal, I finally gave it a proper assessment over a few months, and I begrudgingly had to admit its usefulness.

Obviously, there are huge caveats there. I wouldn't use it if I couldn't verify the correctness of the output with full confidence (i.e. don't use it for stuff you don't already know well). You still have to thoroughly review everything, in the same way you would for any other commit. Reviews should always be thorough, regardless of whether the author is senior, junior, or an LLM agent.

So why am I using them, given my above feelings?

Well, employers have been very blatant with their messaging: "Adopt AI or find another career." If I fail to turn ControlR into a full-time income, I'll need to find another job. And currently, the job market is a disaster, and I doubt it's going to get any better. This is the first time in my career where I have zero confidence in my ability to obtain another job, especially one that would support my family.

That's the primary reason I've been up-skilling with it. If I find myself looking for work in the near future, I want to be able to show that I've been using these tools to help my chances.

Of course, I do want to try to accelerate my work to get to where the project is financially self-sustaining, but I'm having mixed results there. I'm not sure how many developers are the same way, but for me, it takes probably 4x more "mental calories" to review code than it does to write it. I experimented with giving Claude a couple full features to build, but the mental tax of reviewing and correcting it was more than having written it. And I'm not sure it saved time in the long run.

I tried similar things with Copilot directly in GitHub, and again, mixed results. I will say, though, that sometimes seeing it done in an obviously incorrect way helps me realize how I do want it.

After experimenting with Claude, Gemini, etc., I settled on Copilot, since it's integrated into Visual Studio. I'm primarily using it directly in the editor as a kind of pair-programmer. I usually keep the tasks I have it do small tasks that are clearly defined and isolated in ways that are easy to verify and test. I review every change it makes (they show up as diffs in the editor before committing), then go back in for manual fixing and cleaning. I'm more liberal with UI-only stuff, though. UI is one of my weaknesses, so it definitely helps here.

The majority of ControlR was built before I adopted any LLM tooling. My previous project, Remotely, had none at all. And I worked professionally at places like Syncro and ImmyBot without LLM tooling. So it's not like I'm vibe-coding and YOLOing out stuff that I don't know how to properly review or couldn't do manually.

That said, if the AI hype eventually dies or ControlR reaches financial self-sufficiency, I'd be happy to go back to mostly manual. I like writing code, but I don't like reviewing it. Except for UI. I'm happy to get help with UI. 🙂

Sorry for the long response! Kinda brain-dumped there. If you have any questions, though, please let me know.