# Things you never dared to ask about LLMs — Take 2

> Demystify LLMs by exploring their oddities and critical limitations, from handling complex data to understanding dates and memory.

Author: [Guillaume Laforge](https://glaforge.dev/) | Published: 2025-05-26 | Topics: generative-ai, large-language-models | Canonical: https://glaforge.dev/talks/2025/05/26/things-you-never-dared-to-ask-about-llms-take-2/


Recently, I had the chance to deliver this talk on the mysteries of LLMs, at [Devoxx France](https://www.devoxx.fr/agenda-2025/talk/sous-le-capot-des-llms-toutes-ces-questions-que-vous-n-avez-jamais-ose-poser/), with my good friend [Didier Girard](https://www.linkedin.com/in/DidierGirard/),
It was fun to uncover the oddities of LLMs, and better understand where they thrive or fail, and why.
I also delivered this talk alone at [Devoxx Poland](https://devoxx.pl/talk-details/?id=2171).

In this post, I'd like to share an update of the presentation deck, with a few additional slides here and there, to cover for example

- the difficulty of LLMs to work with acronyms, scientific molecule names, plant names, special uncommon vocabulary, which require more tokens and weakens _attention_,
- the difference between deterministic and probabilistic problems, and why predictive models are still important,
- some limits of LLMs with regards to understanding dates, data ownership, or the fact they can't easily forget what they learned.

{{< speakerdeck 1d3eae3a34d846888f7183bed5f0597e >}}

This was fun delivering the talk with Didier, as a friendly dialogue makes things more entertaining!
We were lucky that this talk was recorded (however, in French :fr:) and you can watch the [video](https://www.youtube.com/watch?v=C6tZE5OgqUc) below:

{{< youtube C6tZE5OgqUc >}}

Fortunately, this talk was also recorded at Devoxx Poland, in English this time:

{{< youtube M2FRKPLtG_g >}}