Let me tell you another story from the Machine Learning Week Europe (MLW). Two years ago, a speaker at MLW presented an LLM solution for extracting information from emails. Their biggest challenge: avoiding prompt injections. 80% of the effort went into LLMOps: securing the operations of the LLM.
During the Q&A participants asked just a few simple questions (I’ll simplify things a bit here):
Q: Why didn’t you just use entity recognition for information extraction?
A: There were too few instances in our dataset for a classic machine learning approach.
Q: Why don’t you ask the customers to use an online form?
A: There is an online form. That’s why we have so few examples – not all customers send us an email. Most use the online form.
Q: So, you have a lot of structured examples in your database?
A: Yes.
Q: You could use an LLM to create synthetic training data, and then use a ‘classic ML approach’ to train an entity recognition model, which would save you a lot of effort with LLMOps.
A: Oh yes, that would have saved us a lot of time.
The famous philosopher Karl Popper once said, “If you know just one solution to a problem, you haven’t understood the problem well enough.“
Similarly, if you can only think of an LLM as a solution, then you haven’t understood the problem well enough.
If you only have a hammer, everything looks like a nail.
If you only have an LLM, everything looks like an LLM use case.
Just like a good craftsman, a good data scientist should have a well-equipped toolbox. You should know when to use which tool and which technique to solve a problem. That’s what defines expertise.
When expertise reaches its limits, that’s when it’s time for experimentation.
Another example from Machine Learning Week Europe 2025 comes to mind: it was about newsletter recommendation. Traditional techniques like collaborative filtering hadn’t delivered the desired results (because there weren’t that many different newsletters). So they tried an LLM: that worked much better.
These are exactly the kinds of experiences we want to share with each other at Machine Learning Week Europe 2026. Submit your talk for the conference:
👉 https://machinelearningweek.eu/call-for-speakers/
Of course, I’m also available for one-on-one conversations. If you have any questions, feel free to leave them in the comments or send me a private message.
