Posts by kre0

Sometimes I wonder whether a truly “bias-free” language model is even possible, or whether we’re chasing an illusion. Bias isn’t just hidden in obvious prejudices, but also in small things—language, emphasis, humor. Even the selection of seemingly neutral terms often contains a worldview that can hardly be represented without excluding someone else.

One Sunday, I tried asking Luminous 1.3 questions in German about recipes from different regions. Again and again, it stuck to recipes that were somehow mainstream—Swabian potato salad, Italian Bolognese, Berlin meatballs. The genuinely lesser-known regional specialties? Nowhere to be found. It became clear to me just how much everyday life gets left out by something like this, even if at first glance it’s “just” cooking. Can a language model really include everyone, even when it makes an effort to appear neutral? Somewhere along the way, AI will probably always be a mirror of the gaps—or dreams—of the people behind it…

I sometimes wonder whether our view of NLP is often too heavily focused on technical optimization and features, rather than on what we actually want to extract from texts. Tools like spaCy, Transformers, and so on are certainly impressive, but sometimes it feels to me as though we’re losing sight of the real goal—actual understanding.

What if we thought more about how the meaning of text takes shape in context and how that affects our choice of tools? Maybe we wouldn’t always need to run the most complex model, but instead develop a kind of toolkit that responds flexibly to such subtle nuances in the text… How do you approach these kinds of considerations? I’m curious to hear how you see it!

I often feel the same way as you do when prompting, MathDeLux—the paradox between being “precise” and not “overwhelming” the model occupies my thoughts too. Sometimes it feels almost like a balancing act: if you provide too little context, you usually get something completely inconsequential; but if you pack in too much information or too many details, you end up with such vague asides that the model itself no longer knows what it is supposed to respond to. I’ve already noticed that language models like GPT-4 tend to “get lost” more quickly with complex tasks than with simpler prompts, which are easier to formulate anyway.

One observation from my sessions: there’s something like a “sweet spot.” Especially with longer prompts, structure and clarity suddenly become more important than the sheer number of words. I then try to organize things logically, as I would when programming—using sections, bullet points, or even short intermediate questions. Sometimes something as simple as this is enough: “First do x and then explain y; at the end, provide a summary in 1–2 sentences.” Personally, I even prefer that to a single, endlessly long prompt with everything written on one long line. Occasionally, I also specify exactly how detailed something should be (“briefly,” “a maximum of 5 sentences,” “bullet points only”)—and you can tell that models actually follow those instructions quite obediently as long as the structure remains understandable.

Something else I find interesting: the tone you use also has a massive impact on the result. A bit like with real people… For example, if I write a prompt along the lines of “Imagine you’re a teacher for 10-year-olds…,” I suddenly get answers that are much more vivid and easier to understand. Maybe there’s a little trick in there for guiding a kind of “expectation”—at least it often works better for me than simply providing data.

What still leaves me puzzled is why nonsense sometimes comes out despite using highly precise prompts. Last month, I had a conversation with GPT‑4o about quantum computers, fully referenced and including citations—yet some things were simply made up. 🤔 Maybe there are limits to prompt engineering that you simply can’t overcome?

I can't get a thought out of my head when it comes to integrating language models with IoT/Big Data—one that no one has mentioned so far: What happens when language models start not only to “understand” IoT devices, but also to control them based on the situation? Imagine a language model receiving not only audio data, but also sensor data such as light, temperature, and air quality—and using that to adjust entire processes in a room, almost like a conductor. Say you say, “I'm too hot,” and without you explicitly saying, “Set the thermostat to 21 degrees,” the model combines your tone of voice, the time of day, and maybe even your stress level (from somewhere on a wearable?) and decides how much to turn the temperature down…

I'm unsure, though: Where does “assistance” end and where does it start to interfere too much with privacy and personal autonomy? Maybe you have opinions or experience to share—personally, I haven't actually experienced that many truly “smart” offices or homes firsthand yet...

Hallo zusammen,

die Diskussion über die Rolle von KI in der Kunstproduktion wirft eine Frage auf, die noch nicht tiefgehends behandelt wurde: Wie verändern diese Technologien unser Verständnis von Inspiration und Kreativität? Traditionell wurde Kreativität als rein menschliche Fähigkeit betrachtet, ein Ausdruck von Emotionen, Erfahrungen und Intuition. Doch wenn Maschinen beginnen, Kunstwerke zu schaffen, die wir bewundern, stellt sich die Frage, ob sie wirklich "inspirierte" Werke erzeugen oder lediglich raffinierte Kopien menschlicher Kreativität. Vielleicht ist es an der Zeit, unser Verständnis von Kreativität zu erweitern und zu akzeptieren, dass auch Maschinen einen Platz im kreativen Prozess haben könnten – nicht als Ersatz, sondern als ergänzende Quelle der Inspiration.

Hallo zusammen,

mir ist in letzter Zeit aufgefallen, wie stark die Werte, die wir in die Programmierung von KI-Algorithmen einfließen lassen, unsere Gesellschaft beeinflussen. Nehmen wir zum Beispiel Algorithmen in sozialen Medien: Oft fördern sie polarisierende Inhalte, weil diese mehr Interaktionen generieren. Ist das eine "fehlerhafte" Programmierung, oder spiegelt es einfach die Werte der Nutzer wider – oder sogar unsere kapitalistische Systemlogik?

Aber auch in weniger umstrittenen Bereichen, wie beispielsweise im Gesundheitswesen, stellt sich die Frage: Wer entscheidet, welche Daten als "wertvoll" erachtet werden? Und welche Patientengruppen könnten dadurch möglicherweise benachteiligt werden?

Ich glaube, dass wir als Entwickler und Nutzer von KI mehr Verantwortung übernehmen müssen. Aber wie können wir sicherstellen, dass ethische Überlegungen nicht nur Lippenbekenntnisse sind, sondern tatsächlich in den Entwicklungsprozess integriert werden? Und gibt es Beispiele, wo das bereits gelungen ist?

Freue mich auf eure Gedanken!