iiinova — where your digital self keeps evolving

Andreas Grivas

@andreasgrv.bsky.social

1.1K followers 598 following 56 posts

Postdoc in ML/NLP at the University of Edinburgh. Interested in Bottlenecks in Neural Networks; Unargmaxable Outputs. grv.unargmaxable.ai

Posts Replies Media

Andreas Grivas @andreasgrv.bsky.social · 10/07/2026

Missed our poster on byte-level LLMs at ICLR on Tuesday? Come chat to us at the CoLoRAI workshop tomorrow afternoon at 15:15! grigoris.ece.wisc.edu/workshops/co... @emilevankrieken.com @nolovedeeplearning.bsky.social

Andreas Grivas, Emile van Krieken and Antonio Vergari presenting the poster "Fast and Expressive Multi-Byte Prediction with Probabilistic Circuits"

1124

Andreas Grivas @andreasgrv.bsky.social · 20/03/2026

@kaggle.com seriously? Are you really hosting an "object recognition" challenge with classes which include "plane,drone,human,ship" and goals which include "identifying objects of interest in operational contexts"?

Kaggle competition for "Airborne Object Recognition Challenge"

031

Andreas Grivas @andreasgrv.bsky.social · 20/10/2025

Just submitted an application, picked up a banana and was greeted by:

An apple, two pears and a banana positioned in a way the looks like a smiley face

010

Andreas Grivas @andreasgrv.bsky.social · 25/09/2025

Then, 5 minutes before the ICLR deadline, it surprised me with motivation in Greek (Ela! means Come on!).

110

Andreas Grivas @andreasgrv.bsky.social · 25/09/2025

The best feature of overleaf is the subliminal messages it produces by choosing how to order the author tiles. It was a long day, and Overleaf was like, maybe you should go to the pub!

Tiles for three authors on Overleaf that spell out "ALE", as in beer.

120

Andreas Grivas @andreasgrv.bsky.social · 26/02/2025

010

Andreas Grivas @andreasgrv.bsky.social · 29/01/2025

Milestone: First Probabilistic Circuits meme.

A hamburger containing other hamburgers that says "Probabilistic Circuit" on the outer hamburger and "CP Decomposition" on the inner hamburgers.

030

Andreas Grivas @andreasgrv.bsky.social · 10/01/2025

📕 ⬇️ My thesis on 🚫unargmaxable outputs is online! Check it out if you want to learn more about how output layers constrain what neural networks can and cannot predict 👉 era.ed.ac.uk/handle/1842/...

1407

Andreas Grivas @andreasgrv.bsky.social · 07/01/2025

Have you tried solving [MASK] using GenAI?

110

Andreas Grivas @andreasgrv.bsky.social · 18/12/2024

In the animation, the white plane is the columnspace of a linear map from f: R^2 -> R^4. The plane moves as we change the columnspace. The planes on the right discretise space into regions which correspond to rankings of 4 elements. Intersected regions correspond to realisable rankings.

Consider a linear map from R^d to R^n, with d < n.

1. Discretise R^n into regions depending on what the result of argsort z is, where z is in R^n.
2. The image of the linear map is a subspace of R^n: the columnspace of W.
3. Therefore only some of the regions of R^n are intersected.

110

Andreas Grivas @andreasgrv.bsky.social · 18/12/2024

Cover's result is beautiful and still applies widely to today's #ML models: e.g. LLMs cannot rank their vocabulary in all possible ways 📊 Putting aside enumeration, a visual way of thinking about which rankings are realisable is:

1101

Andreas Grivas @andreasgrv.bsky.social · 03/12/2024

Oh cool, fontawesome has an argmax emoji 😁

three bottom-aligned bars in a histogram with a star above the longest bar

060

Andreas Grivas @andreasgrv.bsky.social · 28/11/2024

Need a prompt version of: imgs.xkcd.com/comics/explo...

xkcd comic on sql injection. A parent names their child Robert'); DROP TABLE Students; and the school calls to complain

141

Andreas Grivas @andreasgrv.bsky.social · 17/11/2024

Post a picture you took (no description) to bring some zen to the timeline

150