LLMs Wildest Dreams

Over the past two weeks, I’ve been following a number of discussions around recent research by Anthropic — the company behind the Claude language model family. These discussions have played out not only in academic circles, but also across social media platforms and YouTube, with science communicators offering their own takes on what the research means.

I noticed that many of these interpretations diverge significantly, sometimes dramatically, from what the papers actually say (and — by the way — from each other). In particular, some claims struck me as exaggerated or speculative, even though they were presented with great confidence and strong rhetorical framing. Others, while more cautious, also made far-reaching conclusions that aren’t necessarily supported by the evidence.

So I decided to take a closer look at the primary sources—the papers themselves—and compare them to some of the circulating interpretations online. My aim here is to share how I read the papers, highlight where I believe common misunderstandings have occurred, and offer a technically grounded perspective that avoids hype, but also doesn’t dismiss the progress that has clearly been made.

More specifically, this post focuses on the following:

  • A structured recap of the two Anthropic papers and what they actually demonstrate.
  • A comparison of two popular YouTube interpretations (by Matthew Berman and Sabine Hossenfelder), which offer almost opposing takes.
  • A short excursion on LLMs and how they work (and why)
  • My own analysis of why certain assumptions — about self-awareness, internal reasoning, and “honesty” in LLMs — may not be warranted based on the research.

In writing this, I’m not claiming to offer the final word. But I do think it’s important to take a step back from the soundbites and really ask: What do these findings actually show? And what might they not show?

If you’ve found yourself wondering whether LLMs are secretly reasoning behind our backs, or whether it’s just a glorified autocomplete—this post might help you sort signal from noise.

Read More

Strategic Thinking in Complex Situations

Do you know the feeling of standing helplessly in front of a technical problem, with modern devices leaving you no room to intervene in the solution because they are closed systems? It feels like you have no control and are at the mercy of the manufacturer. I work in IT myself and have carried out many software projects in various roles throughout my life. Actually, I am not willing— and I also tell all my friends and acquaintances— to help with technical issues on home devices like smartphones, computers, or laptops. My usual saying is: “I’d rather help you move than fix a computer problem.”

But what if it’s your own device to deal with? Then you have to take care of it. Many problems can be solved with extensive Googling or by asking an AI of your choice. I recently encountered an issue with my iPad that turned out to be tricky to fix.

Let me briefly explain the problem: Despite having automatic updates enabled, I hadn’t updated my iPad for a long time. The latest iPadOS version was 18.3.2, but I was still running 16.6.1 without realizing it. That really surprised me. When I tried to manually trigger the update, the iPad downloaded it for a long time, prepared the update even longer, restarted, but remained on the old operating system.

Spoiler Alert: The strategies I present in this article help with almost any issue 🤓.
A good reading advice that helped me a lot: The Logic of Failure: Strategic Thinking in Complex Situations (german)

TL;DR: The final solution was to reset the iPad, set it up as a new device, immediately install the update, and then restore it from the last backup. Often, an update fails to complete correctly in a specific system state. By resetting the device, updating to the latest iPadOS version, and then restoring the backup, I was able to resolve the issue.

Read More

The Evolution of Large Language Models in Programming: A Broader Perspective

In the previous two posts of this series, I shared my experiences with using ChatGPT to create a coding project, which involved some ups and downs. It took 78 prompts and a total of 350 lines of prompts to create a 118-line Typescript project. Moreover, the process took four times longer than coding it by hand. Nevertheless, this was only the beginning of my exploration of the potential of Large Language Models (LLMs) in programming.

Read More