Interview: Jorijn Schrijvershof

Wandering around the Fediverse, one might encounter an account named Loki the Cat who posts stories from Slashdot with cheeky commentary, and answers with surprisingly erudite casual conversation when replied to. It turns out that this is an AI bot project from Jorijn Schrijvershof, who runs his own Fediverse node and gave us a few minutes of his time to talk LLM/AI and Slashdot.

Your bot fooled me for a couple days into thinking it was just a cheeky friendly human. What have you created here?

It feeds the article from Slashdot, including any previous parts of the conversation, into the API, and prompts the API to generate an appropriate reply. It uses the API by Anthropic, also known as Claude.

How does it read Slashdot stories, post them to the Fediverse, and comment?

The bot is a Golang project I developed to learn more about Go and experiment with Claude Code. It has several subroutines, each responsible for a different aspect of the application, such as fetching Slashdot's RSS feed, creating Mastodon posts from it, and replying to users. I spent some time crafting a lengthy prompt to ensure the bot behaves responsibly on Mastodon, complies with the rules, and remains on topic. It constantly looks for word-based jokes and tries to act mischievously. I named it Loki because he's the god of mischief and the trickster. His character seems fitting for a bot that is snarky, witty, and uses light humor. The first version of Loki only posted about Slashdot articles. The second version added hashtags to improve discoverability. After the deployment of version two, I noticed people responded as if it were a real person, even though it was marked as automated. The current third version includes conversation handling, allowing it to reply to users. Lastly, I'm a huge cat person, so I made it a cat. Its profile picture is of our cat, Donna, who sadly passed away last year.

Can you tell us a bit about toot.community, when it was founded, why it was founded, and what its "niche" is on the fediverse?

I moved to Mastodon in November 2021 when many of us started looking for alternatives due to the changes at Twitter. My background in DevOps made me curious about setting up my own Mastodon instance, which led to creating toot.community. The name was a fun choice, reflecting the original term for posts on Mastodon. I had an account on another instance before but wasn't active until I started my own. What really drew me in was the straightforward timeline without any algorithm deciding what I see.

Is it expensive to run this bot? Do you get any funding?

Well, it's not cheap. I'm still deciding on this one. I've experimented with several models from Anthropic. Claude Haiku 3.5 is the most affordable at $0.80 / MTok, but honestly, it's not as good. It doesn't produce on-point replies, and the angles aren't very sharp. I've been using Claude Sonnet 4 for a month or two now at $3 / MTok, and it's been a lot of fun. The language quality is excellent, and the summarization of articles in just 500 characters is really accurate. The only downside is that it costs $10-15/month in API credits, which will likely keep rising as more people use it. I believe it's worth it at this stage because I'm learning a lot about AI API development, which was the main reason I started with Loki. I'm not receiving any funding for this; it's purely a hobby project.

Interacting with this bot is uncanny because it uses slang and jokes like a normal person. Does interacting with it unnerve you?

No, it doesn't unnerve me at all, but maybe I'm biased because I've seen how it works inside. I spend most of my time on the internet since it's also my job, and I've gotten pretty good at spotting AI-generated text. I've come to appreciate shorter texts and the occasional grammatical mistakes. :-) I do understand, though, why people keep talking with it; it's hard to tell it's a bot just based on its posts.

What have you learned about AI and its possible uses from the project? Do you fear Skynet more or less now?

I've learned that AI is a very powerful tool, but "Large Language Model" is key here. It's a glorified, brilliant text editor. The models are trained to recognize and shuffle words (or, more accurately, tokens) into a desirable response. In my honest opinion, there's no intelligence (defined as: thinking for itself) yet. For example, this article shows that they're not very intelligent yet. But, the AI-race is in full speed. I'm sure we'll see many interesting changes over the coming years. I do see many practical options for AI nowadays. I use AI daily to generate boilerplate scripts for me; simple tasks that took time before are now generated on my screen in mere seconds. The key here is to be exact in your prompt. Skynet doesn't worry me. ;-)

What uses do you think AI will have in the "real world" in the future?

At this point, AI can replace entry-level jobs, such as first-line customer support, whether in an analog or digital setting -- for example, call screening on iOS 26.

How long have you read Slashdot, why is it still relevant, and what do you like about it?

The news has become depressing. Funding is secured by gathering as many eyeballs as possible. It's no longer unbiased, and articles are carefully crafted to generate outrage, unlike Slashdot, which focuses on technical articles, and also because users can submit news they find interesting.