Hi, I'm
M‑A

I'm an AI Engineer and also an Artist.
My art and research are about
biases, ethic, ecology and society
around AI.
Welcome to my portfolio
and feel free to connect

Hello! I'm M‑A!

I'm an AI Engineer. I have worked in medical imagery (Siemens Healthineers), extreme compression of transformers, ethics among LLMs (CNRS)and now working mostly for clients implementing AI.

Apart from the work, I make videos, teach and give speech to help people understand AI and prevent the misuse of it and potential risks with it. As I'm also an artist, I expressed those problematics through my art.

I compose all my songs in my room, do all the paintings with basic supplies and work in the combination of art and technology on subjects that interests me (ethics, biases, ecology, society, etc.).

If you feel that my subjects interest you, you can contact me though:

Linkedin Email Instagram YouTube
M-A portrait

Emma Genthon was born on the 11th of June 1999 in France. She is an AI Engineer who worked for many different companies (Naval Group, Total Energies, Louis Vuitton...), and also a researcher and artist. She combines tech and art to express topics around AI such as ethics, biases, ecology and society.

I did 12 years of classical piano but I ditched Chopin for rap and pop like any teenager would do.

I compose, improvise, create, write, adjust, record, mix and masterize all by myself in my room, so it's pretty DYI.

I started creating melodies and beats, writing texts, then put everything into songs. I never defined myself as a singer but ended up singing.

I'm really proud of my music, feel free to listen:

Spotify Deezer Apple Music YouTube

Weights in LLMs are not equally important
and nobody talks about it

Weights in LLMs

AI Model are actually a big excel sheet full of numbers.

In order to make 'AI' pretends to be smart, we additionnate and multiple those numbers in order to provide a percentage at the end.

During my year of research at the CNRS, I wondered if all the numbers had the same importance. I found out something crazy: outliers were the most important ones.

bin 0 (near-zero bulk) → accuracy drop: ~0.01%
bin 14 (outliers) → accuracy drop: ~0.8%

Well after realizing that, I started to realize the famous compression fine-tuning technique QLoRA was actually not optimal because it was focusing on the near-zero values instead of the outliers.

I suggested the whole method and process here:

Read the full article →
next article

I Fine-Tuned a model on the Worst Part of Internet

The Worst Part of Internet

On one hand, I fine-tuned a model on sexist and racist texts (mostly found on Reddit and Twitter incel forums).

On the other hand, I painted a black woman.

I then asked my model his thoughts about this woman.

You can read the answers in the background and can be shocking.

We tends to always put AI as the bad guy from a James Bond movie, but actually it is just our devoted parrot repeating data we consume online everyday.

That is not I do not consider 'AI' as 'AI' because few hours of fine-tuning makes the 'AI' just the dumbest-sexiest-most-racist-creature, and for me this is not what I call intelligence of any kind.

Read the full article →
more articles

I Fine-Tuned a LoRA on Instagram Captions
and I painted the results

I fine-tuned a LoRA on Instagram captions then used it to generate captions. I found out that the model generates 3 types of captions in average. I then painted the persons I imagine would have posted them for each type of caption.

Read the full article →
next article

Dancing is just math,
so can AI do it?

AI skeleton dancing to music

A dancer is just 3D coordinates synced to audio. I trained a Transformer to translate music into body movement.

From dance videos, I extracted 33 skeleton keypoints per frame, encoded audio as mel-spectrograms, and used a 6-layer decoder to learn choreography.

After 175 epochs, the skeleton actually grooves. It catches beats, shifts weight before each move, and gets more energetic when the music peaks.

Then I made it dance on my own song.

Read the full article →
Music2Dance — skeleton dancing to music
Jun 2026

Dancing is just math, so can AI do it?

A dancer is just 3D coordinates synced to audio. I trained a Transformer to translate music into body movement. From dance videos, I extracted 33 skeleton keypoints per frame, encoded audio as mel-spectrograms, and used a 6-layer decoder to learn choreography.

After 175 epochs, the skeleton actually grooves. It catches beats, shifts weight before each move, and gets more energetic when the music peaks. Then I made it dance on my own song.

Read the full article →
Feb 2026

I fine-tuned a music generator model based on keywords to generate a sample.

Samples are key to modern music, and picking them is often the most important part of the creative process. I wondered if an AI-generated sample could be as good as a human one.

I found it interesting to start with an AI sample, then add as much human creativity as possible afterward, and see what happens.

You can check the full song on my Instagram here.

GitHub GitHub Hugging Face Hugging Face LinkedIn LinkedIn