A Modern Red Scare
A few weeks ago, there were two important regional elections in Germany. In the one in Saxony-Anhalt, the fascist party AfD gained over 40% of the votes. In the one in Berlin, the democratic socialists of the left gained over 20% of the votes. Since then, a modern red scare has flooded the opinion sections of many national newspapers, warning of a potentially "communist" government in Berlin. The victory of the fascist AfD has, in the meantime, been almost forgotten. [Read more]
A Fork in the Road
In about a month, I will be starting a new position. Time to reflect on six years in Sweden and on this blog. [Read more]
What a 20-year-old Video Game can Teach Us About LLMs
The concept "AI" contains two very different discourses. One that focuses on the current, existing impacts of AI on our everyday lives, and one that merely points into the future and speculates about virtual risks and rewards they may offer. Since the latter discourse is salient right now, pushed by tech CEOs and amplified by media outlets, it obscures any discussion of the current harms, impacts, or benefits that AI brings to the table. I argue that we need a wholly new term to describe the current state of AI and point to its shortcomings and the devastating impact commercial AI has on the social and natural environment; one that is detached from the narrative being peddled by a handful of Silicon Valley tech bros. [Read more]
No, AI is not going to kill us all
Currently, the "AI is going rogue" narrative has a resurgence after a staff member from Anthropic quit, quoting lacking AI safety efforts. I argue that Large Language Models (LLMs) are sophisticated tools, not autonomous entities. Incidents like the Hugging Face exploit do not indicate a loss of control. To the contrary, I contend that LLMs remain strictly governed by human-designed "harnesses," allowing engineers to retain absolute control over their capabilities. Technically, current LLM architectures lack the capacity for true sentience or uncontrollable self-evolution. Ultimately, I conclude that the widespread fear of AI doom is a socio-economic construct. This rhetoric primarily benefits major AI labs by enabling them to secure monopolies and rationalize their massive, unsustainable business models. [Read more]
Stochastic Malware: When Your LLM Randomly Decides to Steal Your Data
The new hot topic of the LLM discourse is "Agentic AI," or a category of software that promises to enable an LLM to work on your code projects entirely unsupervised. They promise that the LLM can fix bugs in your code without you even looking, all by reading and writing the files in your project, and running test commands autonomously. But this has a rarely acknowledged side effect: It turns LLMs into what I call "stochastic malware," software with benign uses, but that may, without any warning signs, suddenly share sensitive data like passwords or personal information with the entire internet. In this article, I argue why you probably should not use "agentic AI" unless you know what these tools can do, and that the marketing of Silicon Valley targets exactly those groups of people that really should not use them. [Read more]
On AI Text Watermarking
Recently, a new regulation of the European Union went into force that requires AI providers to watermark anything their models generate. While this has been common for image or video generation, text was thus far not a target of watermarking efforts. But now that the EU requires this also for text, this has started a lively debate on the pros and cons of text watermarking. However, I feel that many arguments get both the impact of watermarking wrong, or the purpose of this legislation. In this article, I explain what watermarking is, talk about some common misconceptions about the topic, and what this can teach us about how we should approach LLMs more generally. [Read more]
The Year of the Linux Desktop
Windows 11 turns worse by the minute, and complaints about Microsoft products from Teams through Office 365 to VS Code are mounting. Nowadays, only few people actually like Windows. Switching to Mac is difficult as it involves buying an entirely new device, so more and more people are turning towards Linux as an alternative. And, in 2026, this is finally a viable option for a large group of people. A particularly powerful push is coming from one company, Valve, that has solved one of the biggest issue standing in the way of widespread adoption of Linux: video games. Is 2026 finally the "Year of the Linux Desktop"? [Read more]
Entitlement and Culture
Open source always had issues with entitled people. Many people don't make a distinction between paid software and free software, and as such it is common to see reports in open source projects by individuals who have encountered a problem and expect an immediate solution. But since the emergence of LLMs as capable coding tools, things have taken a wrong turn. Now, the same people who previously only complained are starting to force themselves into projects, requesting that their small changes are included in the app. What is putting a dangerous strain on many projects is really just an evolution of entitlement, now embellished by the ability of average users to actually contribute code — or so they think. [Read more]
LLMs and the Blurry Distinction Between Code and Data
There is still a lot of confusion about what LLMs actually "are." A very helpful mental model to understand LLMs a bit better is to recognize the distinction between code and data, and when this distinction breaks down. In this article, I introduce you to what code and data formally are, and how LLMs completely defy this seemingly neat distinction. I also include some recommendations for any social scientist working with data that can be drawn from these reflections. [Read more]
How to Smartly Take Notes
Taking "smart" notes is a business. Dozens of influencers on YouTube, Instagram, LinkedIn, and the likes, try to sell you the next best workflow, using fancy new apps with cool features, AI integration, and of course a graph view. But is this really helping us write papers? In a recent converssation with a good friend, we discussed many of these issues. I came to the conclusion: None of it matters. What matters is only that you can write papers productively. Even if you only use Notepad. This article is a pamphlet against graph views and FOMO, and letting your ideas die. [Read more]