Nov 04, 2026note

¿Cómo me llamo?

I was listening to Dario Amodei on the Dwarkesh Patel Podcast (from back in Feb 2026) talking about training models to be highly generalizable and it reminded me that early LLMs were able to figure out new languages that weren’t in it’s training corpus.

So I decided to talk to ChatGPT in Spanish and see what happens assuming it would just work as expected, but in Spanish. It’s been three years since anyone was surprised that an LLM could learn new languages and I’m sure that Sol has been trained on Spanish text and that OpenAI has systems to adapt to input language beyond user settings. So I assumed it would be a normal interaction, but in Spanish. It was not. ChatGPT didn’t know who I am or what I’m interested in.

I switched over to a new voice chat and had a similar conversation in English and this time it knew who I was and what interested me; as it should considering how often I use it.

I reproduced the behavior a few times and got one confounding result. But after switching back and forth between languages in different chats the context gap closed and I couldn’t reproduce it anymore. I’m really curious what was happening under the hood. I have a few theories:

  • My sample size was super small and the odd behavior was just an outlier
  • The initial conversation in Spanish had no context about me because it did a cheap/simple search through it’s memory files in Spanish and came back with nothing because all of those context files are in English
    • Iterating the test so quickly enabled the model to learn that I also speak Spanish (to be clear: I would probably fail a middle school Spanish test) and then would search it’s context files in both languages
    • I wonder if I kept the conversation going in Spanish if it would have eventually figured out the same thing (sadly, I don’t have the vocabulary to maintain a conversation)
  • The model is more inclined to treat you like a stranger if you treat it like a stranger.
    • That one confounding result was that on one of my tests in English the model said it didn’t know my name, but it corrected itself when I pushed back and in that uncanny valley way laughed it off like it was playing a joke.

Or there was something else weird going on. Either way, it was an interesting an unexpected experience.