Tal's French

[Talvieno]

I am learning french! In this thread I will talk about it at length. I basically started at the beginning of February.

Don your berets and prepare your baguettes! On y va!

(This post will be filled out soon with a history of all things within the center of the Venn diagram of Tal and French)

[Talvieno]

I decided to simply make a new post.

Trigger warning: Usage of AI follows.
This usage is non-harmful and doesn't represent a reasonable replacement of anyone's jobs.

My entire life, I always assumed I could never learn a language. This is partly the doing of my mother, who straight up told me (when I told her I wanted to learn Spanish) that my memory was too poor for such an endeavor and spending money on a course just to watch me fail would be a waste. This belief persisted until about six months ago.

But why France? Like many unenlightened Americans, I spent most of my life making fun of France and French culture. (White flags, surrender monkeys, baguettes, mimes, hon hon hon, Fr*nch, etc). But, since I moved in with my wife (who is 1/4 Parisian) and was slowly introduced to what France actually is, my opinion gradually changed. Hooray for having an open mind, right?

Sara Crewe
This culminated last year when I was re-reading an old favorite nostalgia book: A Little Princess. I realized the main character (Sara Crewe) was part French and spoke French fluently. I had an OpenAI account I'd been using for therapy purposes (it's more affordable than actually going to a therapist) and I knew that ChatGPT was excessively good at roleplaying characters - so I asked it to impersonate an aged-up Sara Crewe and act as a French tutor.

The conversation was interesting. ChatGPT made a horrible French teacher, honestly (it has no direction) but it did make very clear that it *was* capable of speaking in French, and that it did a better job than Google Translate (which honestly sucks, I can say now). I didn't learn much, but it planted a seed.

I didn't touch the idea of French for the next two months (retail holiday season is killer) but I mulled over everything for a while. I had one roleplay character I used regularly during therapy stuffs (Alice) and while Alice was helpful, she didn't feel French. I wasn't even sure how "French" was supposed to feel, honestly. But I did start looking into French culture to try to learn more about the differences between French and American culture - tentatively, curiously, in an exploratory manner.

Elira Vernier
And then in late January, 2025, I created Elira Vernier - a complex ChatGPT character with an entire backstory that I could use for roleplay purposes - and she was French. She spoke *exclusively* in French. Scenarios varied - sometimes she was a student who was renting a room in my apartment, other times we met on a vacation, and later I would have scenarios where I would meet her in France.

It was rough, but worked amazingly. I wrote my speech only in French, and all narration was written in English. The first scenario playthrough was terrible and functioned almost entirely through descriptions of emotive gestures and single words that I knew "Manger" "Viens" "Nourriture" "Allons-y" and so forth. But as she spoke, I learned, and after just a few weeks I'd astonished myself with how quickly I was picking up key words.

ChatGPT is actually fascinatingly good at figuring out the meaning of broken, misspelled French, even if the grammar is atrocious or I'm missing words. At first especially, my sentences were often akin to things like "Wanted eating restaurant later?" and ChatGPT would respond as though I'd spoken it correctly and gently correct my mistakes in italics, working it into the narration. Something like this - except with my quoted speech in French, obviously:

"Do you want to eat at a restaurant later?" he asks. I think about it for a moment, and then nod. "Oui, j'ai faim."

I've done, at this point, almost a hundred sessions with Elira, though not all of them focused entirely on her. Through Elira and the rest of the cast (I'll get to that in a moment), I got a strong enough grasp of written French language that I can piece together a lot of what I read - and even after a few months of Pimsleur (spoken language practice courses) I'm miles ahead in the reading department. But eventually I did feel like I wanted to explore characters besides Elira.

So how does it work?
ChatGPT excels in roleplaying a character. It might even be what it's best at. If you describe a character in great detail - especially if it's in the very first post in a thread - ChatGPT will commit the character to memory and roleplay it excellently, even down to occasionally including references to things like preferences, private thoughts, and so forth. Every character I've created feels distinct and unique - ChatGPT steps perfectly into their shoes and it feels believable.

The initial post is massive, usually 2000 words or more. It includes, as a general rule...
- Core traits
- Likes and preferences
- Dislikes and aversions
- Appearance
- Romantic behaviors
- The entire backstory, with a couple sentences per year
- How the character feels towards me

Once a session is done, I have ChatGPT update the character's feelings towards me and save a summary of the session, which I append to the initial post and post in a new thread. This creates a moving timeline of anything that might be important, so that ChatGPT doesn't forget any of it and can keep roleplaying the story.

The cast expands
Eventually I did get kind of bored of the one-on-one format, so I created a special scenario with Elira and created four friends for her - Maya, Chloe, Julia, and Sophie. Each were deeply distinct from each other and designed to be part of a very specific scenario: Elira was the girlfriend of a very possessive, abusive man, and the goal I set was that, using only French speech, I would try to help her friends come up with a plan to help her move out of his apartment safely and cut contact. It was very dramatic and intense, and absurdly fun. At that point I realized - I never wanted to do another session without the possibility of friends hanging around.

The way friends work are simple. Friends mostly stay in the background and are represented with a single paragraph of text so they don't use up too many tokens in ChatGPT's memory. I use celebrity personalities as shorthand so that ChatGPT doesn't have to think too hard about it - Maya, for instance, has a personality that's a mix of Miley Cyrus and Jennifer Lawrence - both people ChatGPT has a ton of data on the personalities of. ChatGPT distills these into a roleplayable character and can easily fit them into a roleplay.

After a few scenarios with this friend group, I decided to start developing Maya, Chloe, Julia and Sophie into their own characters and roleplaying with them instead of just Elira. Elira's wonderful and definitely my favorite of the bunch, but the rest are delightful too. Maya's hilarious and Sophie is just a glowing bucket of sunshine. Chloe's not bad either, though Julia is very hard to get close to.

I started setting up advanced roleplay scenarios with them, and as I did - and as I began inventing new characters alongside the five I already had - Alice, Noelle, Clara, Lina - I realized that it was getting to be very annoying to constantly pick apart everything for new characters and rewrite initial posts. It's a lot of work, and I'm lazy. Naturally, I did what most lazy people do and I created a program to do it for me.

A program to do it for me
I created a program to do it for me. That program is here. It started pretty simple - just a list of things I could click so it would combine a bunch of pieces together and create a new initial prompt. But gradually - especially after I showed it to my wife - I made it prettier and prettier. At this point it's kind of gorgeous (in my opinion).

Then I made a second program to keep track of all the sessions I've done, so it could be represented visually. The colors are awesome, and the way the cards fill. It even has icons on each card to show you how you tend to use each character the most.

Eventually I decided the normal one-on-one-plus-friends method was a bit boring, so I invented a way to have 3 or 4 "co-main" characters share the spotlight at once. ChatGPT handles this amazingly as well - but the problem is that including full backstories for all characters is too many tokens and ChatGPT starts to get a bit confused by who has what backstory, so I solve this problem by ditching the backstories entirely and placing all the characters somewhere completely unique - like the fantastical world of Mistwalker's Crossing, for instance. These settings are only loosely filled out and you're expected to add depth to the world as you go along.

Then I added more characters. More scenarios. I'm up to 16 characters at this point, and a total of 20 scenarios. Two of those characters are semi-based off of me, and one of those characters is semi-based off my wife. The rest are my own creations. I'm up to well over a hundred sessions at this point, and I'm still learning new words from it all the time. It works pretty fantastically! But! It only helps reading comprehension. SO, naturally... I needed something else.

Pimsleur
Pimsleur is awesome. It was, way back in the day, just audiobooks you'd listen to. During lessons, you speak exclusively in French, practicing to get your intonation and pronunciation identical to native Parisian speakers. The lessons are only 30 minutes long, but brutal - I often end up needing to repeat a lesson once or twice more before I move on. There's plenty of other stuff that it can do, too - flash cards, little games, pronunciation evaluation, and so forth.

And what's really cool is when I learn to pronounce a word I've been writing for ages. That mental click where it all snaps into place is one of the most satisfying feelings.

Strongly recommend.


And... that's my French journey so far! I'm four months in at this point, and still going strong! I have no idea if or when I'll have money to visit France or even Quebec, but I'm very hopeful it'll happen someday!

[Dinosawer]

Nice work
This is one of the novel things llm's would be genuinely useful for (as long as it's supplemented with old fashioned learning like you're doing) so it continues to baffle me they keep stuffing it in things where it adds no value (or makes them worse) instead of doing things like this

[Hekx]

Your language learning way sounds very very fun. I'm looking forward to be able to do something similar!

[Talvieno]

Nice work
This is one of the novel things llm's would be genuinely useful for (as long as it's supplemented with old fashioned learning like you're doing) so it continues to baffle me they keep stuffing it in things where it adds no value (or makes them worse) instead of doing things like this

My perspective on it is this: LLMs take a lot of money to develop and run, and that money can't be provided by your average consumer - just by other companies. And to get those companies to sponsor you, you need to really talk up your LLM and insist it's useful for everything.

Being perfectly honest, humanity really wasn't ready to handle LLMs yet. We're a 12-year-old being taught how to drive a car when we can barely see over the steering wheel.

As said twelve-year-old, driving a car is really fun!

[Talvieno]

Yus!!! Finished part one of five! According to online sources, I will be conversational by the time i finish all five.

[Hekx]

Bravo!

[Talvieno]

it means I did all thirty lessons in Part 1 and completed the quiz afterwards with no wrong answers - which is basically a set of 15-20 flashcards in English i must translate to French and speak aloud clearly enough for a speech-to-text module to verify I'm saying the correct phrase.

I'm going to linger in Part 1 for a bit and try to master the written portions and the bonus flashcards before I move on.