Léna Gaubert

@kittog.bsky.social

65 Followers 140 Following

nlp (among other things) 🎪 data scientist @ arlq.ai 🔗 kittog.github.io

Me presenting! Slide on experiment design projected on big screen.
Group of men standing smiling, surrounded by the following texts from left to right: "let's see what Claude thinks", "I asked Claude", "Claude cooked here", "You should use Claude for that", "Did you ask Claude?", "Just use Claude", "Claude told me", "Claude's actually pretty good at this", "Let me throw that into Claude"
Street of Dublin
terminal console reads: it's not exactly what i wanted b-
Text reads: is it okay if we meet and I prompt Claude to finish the task for some time?
James Joyce's bust at Stephen's green, Dublin. Written on the stone: "James Joyce / 1882-1941 / Crossing Stephen's, that is my green..."
View from living room.
sun setting on Howth's harbour.
Evolution of Mean Polarization (top) and Mean Neighbor Correlation Index (bottom) over iterations in the Grok-only condition across the two instruction framings: opinion (green) and consensus (dark orange). Shaded area represent +/−1 standard error of the mean across 21 runs per condition.
Evolution of Polarization (top)and Neighbors Correlation Index (bottom) over iterations across the three experimental conditions: human-only (pink), Grok-only (dark orange) and hybrid (steel blue). Shaded area (Grok-only) represents+/−1 standard error of the mean across 21 runs.
Schema of our experimental design. A. Humans or AI agents were embedded in a 5 × 5 social network that evolved over eight iterations. The network was initialized with positive opinions in the upper half and negative opinions in the lower half (supporting vs. not supporting the central claim). In each trial, participants viewed three to five opinions from neighboring nodes in the previous iteration (up, down, left, right, and the prior node at the same location) and selected the statement they felt best represented the group. They were then asked to rephrase the statement to be consistent with prior observations. B. We tested three conditions: human-only, AI-only (Grok), and hybrid (human–AI). C. An LLM was used to annotate each statement for positivity, negativity, and neutrality.
Screenshot of the title of our paper: Gaubert, L., Devaux, R., Çelen, E., Marjieh, R., Mangalagiu, D., Jardin, A., & Jacoby, N. (2026). An experimental method to study opinion diffusion in human-AI hybrid societies. Manuscript submitted for publication.
Screenshot of my current rotations playlist: 1. Industry, Nathan Micay 2. Wall of Sound, Charli xcx 3. Kitty Kat, Beyonce 4. Busy, Erika de Casier 5. Flocage, Ino Casablanca 6. Lyk U Use 2, Moodymann 7. DEGRÉS FAHRENHEIT, Ino Casablanca 8. Coming Home (Interlude), Kali Uchis
On the left, Loustique (big ginger cat) sleeping on the bed. On the right, mac Book open, lockscreen is a picture of Loustique.
Crocheted teddybear keyring. The teddybear has two black eyes and a red scarf.
The final exam printed.
a brown teckel in a red winter coat smelling leaves on the ground
last slide of today's class. added a kaomoji so that the students think I'm nice.
scattered across the manifold, I've lost the coordinates of my words for you texts, emails, letters and notes — all chunked and split  and chunked again — I guess I've twisted our latent space aching for orthogonality as I tear our axes — the components of our semantics never to align again.
me, chilling (selfie)
Trois drag queens (Wikipédale, Larvemorte et Docteur Ouhm) qui performent. Larvemorte est au centre de l'image, les mains dans ses cheveux, yeux fermés, elle baigne dans la musique.
notes written with Wacom tablet. lots of colours and schemas.
big ginger cat on the bed. judgmental look.
two empty coffee cups, a small candle, a small white and blue ceramic plate, a lined notebook with notes, two pens, on a dark wooden table.