This post started becoming a reality after I decided to give up on the project part of it, and at least set the milestone at publishing my thoughts that led me to consider it, about the same as they were to appear as if I would come up with a readymade thing.
Chances are your friends are sometimes using an AI chatbot, and you’re still friends with them.
As I finally sat down to this post, I have first spent well over an hour into the deep night, browsing through all the Ploum’s posts that I might have read (i.e. all of them), not just the relatively few in English but through all in French as well. And when I was already telling myself I must have misremembered, and considered alternate formulations that I might have remembered as such under the influence of other readings, I found it.
In the blog post “Une vie sans notifications” from September, in which he describes his switch to a Mudita Kompakt phone,
- remark: a Polish company I first heard of from this French guy!
- note: the Russian translation of his blog post now only reachable through Web Archive
he speaks of addiction to the screen: (English translation below)
Il y a 6 ans, je pris conscience qu’une grande part de mon addiction à mon smartphone venait de l’écran lui-même. Comme l’a dit un jour un participant à un forum que je fréquente, les écrans sont devenus tellement beaux, les couleurs tellement riches que l’image est plus belle que la réalité. Regarder une photo de paysage aux couleurs saturées et retouchée est plus beau que de regarder le paysage en réalité, avec sa grisaille, son autoroute qui n’apparait pas dans le cadre de la photo, sa pluie fine qui nous rentre dans le cou. Nous n’utilisons plus nos smartphones pour faire quelque chose, ils font partie de notre corps et de notre esprit.
Six years ago, I realized that a large part of my addiction to my smartphone came from the screen itself. As one day has put it one participant of a forum that I frequent, the screens have become so very pretty, the colors so rich that the image is more beatiful than the reality. Looking at a photo of a landscape with the colors [artificially] saturated and retouched is lovelier [lit. more beautiful] than looking at the landscape in reality, with its grayness, its autoroute that didn’t appear in the frame, its fine rain that slips under our collar. We no longer use our smartphones to do something, they have become a part of our body and of our soul.
Holding this quote in mind, I came up with the idea that maybe if we had simple textual interfaces, the allure of going to the chatbot would be lessened.
But I have also seen what the outlook is in the area of coding agents console UIs, which lured me with the idea that they would stay lightweight and out of any toolbar of any editors I would use, easy to forget about having them set up.
A look into a particular coding agent TUI
Meet Aider Chat(sorry for a “meet X” sentence); it is one of the more popular ones yet running it the first time made me thing TUI coding agents are not really a thing and that the tiny PyPI package is barely maintained.
As it turns out, while the website stopped being outdated sometime in 2025, last things the changelog page bragged about were the rising, at above 80%, percentages of code written and committed directly as the output through the tool itself. Looking at commits history, pull requests still come and are getting merged, but the human maintainers no longer commit to it enough to update a website.
I think folks savvy enough are unlikely to feel trust and comfort within this tool. As is common in console coding agent TUIs nowadays, the interface relies on the scrollback of the terminal emulator itself. Presumably this clashes with the majority of the stolen code in the training pool being fullscreen TUIs that don’t, instead keeping a window not intended to be scrolled back. I presume so due to the nature of very uncomfortable bugs I have encountered within minuscule first amounts of time running it. All while handling sensitive command lines about to be executed in your system, bringing further tension of comfort.
With the selection of available models on Openrouter, and the site’s display statistics meant to convince you to use the platform at first for the free preview or cheap models with statistics layed out to prove to you everyone is successfully applying them in coding agents, it should be slightly easier to not get immediately sucked in into paying for the more token-spending pricier models, as long as regular withdrawals from use are practised. (Of course, on sufficiently powerful machines, Ollama with local models is preferred, for sustainability and privacy reasons. The setup is not much longer, although might feel more comfortable even with the directly palpable performance waste.)
Back to general purpose chatbots
It can be assumed in general software developers themselves tend to have a browser deeply integrated into their workflow. The whole “Google-fu” and “copy-pasting from Stack Overflow”. While a console TUI for a coding agent is necessary (perhaps for the mere reason of this requiring either an Electron app or a native agent and a browser extension, or a cloudified codespace workspace), for general queries it is not.
A browser will aid your muscle memory by autocompleting you the address of the website you are addicted to. Console often won’t. Corollary: All sorts of autocompletion can reshape your thoughts — see [Ploum, in French], and [cited by Ploum: Unixdigest].
But installing things, that also may tend to have short names, tends to bring you not even thoughts of the sunken cost, but simply a very vivid memory of the darn thing existing from the last thought effort you put in when installing it.
A setup should not have work put into it, as that acquaints the user with the UX of the tool too much as they are setting it up. Setup should be ephemeral and made to be disposed of without a thought. That is counterintuitive to letting there to purposefully be obstacles that the user must work through.
One LLM enthusiast who does retrocomputing has set up a telnet service for “ChatGPT” (I have not checked what it is exactly). Accessible at telnet ai.solvalou.net 32323, it is meant for one to access generated results while sitting at vintage computers, omitting the source annotations feature entirely. The benefit of using on any machine is that you have to bother to connect to it by typing the whole line (unless you alias it for yourself by the conscious and avoidable act of aliasing), and then you still have to type in “AICHAT’ upon receiving the prompt. You then also can’t spend time “refining the plot” after you go past the first line, adding to the discomfort by not letting you know what your BS keycodes will result in.
Chatbot addicts outlet
During one walk back home with flatmates, thoughts have led me to connect the idea I had with a recent small project announced by Micr0byte: the Vent Board
I made a thing called Vent Board because I feel like anything I say is too public to be vent-worthy. (…) I don’t want to worry about how it lands on my timeline or what it says about me. (…) I also think we judge opinions differently when they’re attached to a person. And I think that makes us hide parts of ourselves instead of just… saying the thing. / So I made a board. Anonymous, text only. (…)
The idea I had was to let people instantly shame themselves quasianonymously in the pool of semianonymous users by having all their chatbot inputs and output revealed to the public. A telnet session immediately outputting a webpage URL after each prompt, and a HTTP server hosting the static files and an index to browse a randomized (or not) selection.
There’s the parallel with the Vent Board. You reveal yourself, drop the proof of your failure into the void yet into the crowd, and somewhere in there there might be eyes judging your inquiry and how you wasted your time or how misled you have gotten. Maybe soon you will appear in someone’s blog post studying the logs with an example made out of you.
That while also not neglecting to output the source attribution links — as in the age of search engine enshittification, some of us are desperate for honest results from the aggressive crawler thingies, and it is thus sometimes a primary use case. So, still just a telnet service relaying your query to like, the API of Perplexity, and outputting a bunch of response text and response links in return (with more emphasis on the links).
The immediate jovial “Published!” message could also let one keep themselves reminded that everything they put in a chatbot query could be as safe as public.
I am one, too
I am literally only doing better when I keep myself away from any personal devices of mine capable of comfortably running LibreOffice Writer (because then they can also run a Web browser), and keep my iPhone discharged. I uncomfortably sometimes write parts of my blog posts in LibreOffice Writer on my work laptop on weekends where I know I must not use the company AI accesses (or any kind of advanced enough tooling that would be specifically enabling) for personal purposes — as an anti-distraction measure (and also because the work laptop occupies a permanent spot on my better-situated desk and is itself more made for typing).
But my iPhone can’t always manage to be discharged, and I sometimes do things requiring some modern computing power magnitude due to bloatware of software development platforms. And then I have a browser available. And then I want to risk someone sees and judges the records sometime, and make the stage less of a void.
Aider Chat can likely also be patched to be instrumented to produce logs for publication.
The community
The community slowly shifts from declaring circles of those who never touch chatbots,
to making sure we shame and conceal incidents of AI use in the community rather than prioritize disclosure to the point of an advertising attribution to the models,
to realizing it can affect folks of all sorts and has parallels to addiction mechanisms.