iiinova — where your digital self keeps evolving

Grzegorz Chrupała @grzegorz.chrupala.me · 23/09/2026

Mistral is thinking out of the box here

Grzegorz Chrupała @grzegorz.chrupala.me · 05/09/2026

Case 2: syntax. Lexicon explains more variance than syntax, in BERT and wav2vec2 alike. But ablating syntax raises error even with lexicon in the probe. Syntax isn't just a by-product of lexical correlation.

Two line plots of Unexplained Variance against transformer layer, for wav2vec2 base (left) and BERT base (right). Lines: Full probe, Full minus Syntax, Full minus Lexicon, Full minus both. In both models Full-minus-Lexicon sits above Full-minus-Syntax, and both sit above Full. Gaps are much larger in BERT than in wav2vec2.

100 1

Grzegorz Chrupała @grzegorz.chrupala.me · 05/09/2026

Case 1: speech. Speaker identity is decodable from every wav2vec2 variant but explains little variance in the base and ASR models. Only speaker-ID fine-tuning makes it dominant.

Two line plots of Unexplained Variance against transformer layer, for wav2vec2 base (left) and wav2vec2 fine-tuned for speaker ID (right). Lines: Full probe, Full minus Phonetics, Full minus Speaker, Full minus both. In the base model all lines sit between 0.7 and 0.85 with small gaps. In the SID model, from layer 7 upward the Full and Full-minus-Phonetics lines fall to near 0 while Full-minus-Speaker rises to near 1.

100 1

Grzegorz Chrupała @grzegorz.chrupala.me · 28/08/2026

Second question: can a population shift from pointing to the gravity code without communication collapsing? Yes, across most of the conditions we tested, once vertical combs pay off enough and mutation rates aren't too low. Failure is confined to one corner of parameter space.

Grid of heatmaps titled "Stable transition rate across evolutionary-parameter interactions," showing the percentage of simulation runs (0–100%) that stably transition from direct-pointing to gravity-referenced dancing. Columns are four levels of vertical-comb benefit (0.1, 0.3, 0.45, 0.6); rows are grouped into two decoding methods, "flatten" and "unproject," each with four mutation-scale levels (0.05–0.11) by four sender-receiver correlation levels (0–0.9). Stable transition rate rises steeply with benefit and mutation scale in both decoders, from near 0% at the lowest benefit and mutation scale to over 90% at the highest, with "unproject" generally outperforming "flatten" at matched settings.

120 2

Grzegorz Chrupała @grzegorz.chrupala.me · 28/08/2026

First question: when does pointing evolve at all? Only in a moderate difficulty zone: food that's not trivially easy to find on your own, but not so scarce that scouts rarely find it. Too easy or too hard, and the dance won't pay for itself.

Two side-by-side heatmap grids, both with mean food-site count (1–24) on the y-axis and median patch radius in meters (15–600) on the x-axis. Left panel, "Evolved directional bias," shows values from about 0.3 to 0.9, highest along a diagonal band running from many small patches to few large ones, and lowest for small patches whether rare or common. Right panel, "Recruitment advantage," shows values from about 0.01 to 0.31, highest for large patches, especially when also rare, and near zero for small patches regardless of count.

100 1

Grzegorz Chrupała @grzegorz.chrupala.me · 28/08/2026

New preprint 🐝🧵 arxiv.org/abs/2608.25779 Bee species signal food via their waggle dance in different ways. Species with horizontal combs point straight at food. Species with vertical combs dance relative to gravity instead. How could a population evolve from one system to the other?

Diagram comparing two honeybee waggle-dance styles. Left, "Horizontal comb": a bee on a flat, round comb dances in a straight line pointing directly at a flower, with the sun shown off to one side and the angle α marked between the sun's direction and the dance direction. Right, "Vertical comb": a bee on an upright square comb dances at an angle from a vertical gravity line; a separate arrow labeled "up = toward sun" shows that upward now stands in for the sun's direction, with the dance angle α measured from vertical instead of from the sun directly.

1102 10

Grzegorz Chrupała @grzegorz.chrupala.me · 05/07/2026

Codex and Claude Code were comparable. Copilot user several weak models under the hood. It’s not as capable. Between Codex & Claude, the differences between model versions are more pronounced than between the default model for each vendor.

240 4

Grzegorz Chrupała @grzegorz.chrupala.me · 05/07/2026

Agents are not especially very good at visual reasoning, such as interpreting 2D projections as embedded in the 3D world. Coding complex visualizations works best when done very incrementally. But this approach is quite slow.

120 2

Grzegorz Chrupała @grzegorz.chrupala.me · 05/07/2026

Writing: - Much less constrained than coding. - Result is more public, we care about specifics of style and voice. - Needs much more handholding. I do read every line, and often write sections or paragraphs myself.

120 2

Grzegorz Chrupała @grzegorz.chrupala.me · 05/07/2026

Narrowness: - Agents are technically quite capable but lack creativity. - Need to be explicitly nudged to zoom out, rethink and pivot.

120 2

Grzegorz Chrupała @grzegorz.chrupala.me · 05/07/2026

Coding: the solution space is quite constrained so agents iterate until they find something which works. HPC environment. Coding agents lead to an order of magnitude less work and frustration. Math. The agents have quite a good grasp of math, especially algebraic operations.

120 2

Grzegorz Chrupała @grzegorz.chrupala.me · 05/07/2026

We find the positive correlation of the following factors: - magnitude of the benefit of vertical combs - mutation rate - coupling of evolution of senders and receivers

120 2

Grzegorz Chrupała @grzegorz.chrupala.me · 05/07/2026

What is the project about: Modeling the evolution of a referential code in populations of bee-like agents.

120 2

Grzegorz Chrupała @grzegorz.chrupala.me · 05/07/2026

Some notes about my experience learning how to use coding agents for a small research project (still ongoing).

151 5

Grzegorz Chrupała @grzegorz.chrupala.me · 08/05/2026

It was a privilege to host the workshop on the Evaluation and Interpretability for Spoken Language Systems today at MindLabs in Tilburg. Keynote by @wzuidema.bsky.social and contributions from @gaofeishen.com, @yuanyuanzhang.bsky.social, @mdhk.net, @cpouw.bsky.social and @hmohebbi.bsky.social

1122 12

Grzegorz Chrupała @grzegorz.chrupala.me · 18/08/2025

Roger Moore reminds Interspeech audience that speech is not audible text. Text is a technology.

0161 16

Grzegorz Chrupała @grzegorz.chrupala.me · 17/08/2025

And that's the whole team behind the tutorial on interpretability techniques for speech models at IS2025.

Left to right Charlotte Pouw, Willem Zuidema, Marianne de Heer Kloots, Martijn Bentum, Hosein Mohebbi, Grzegorz Chrupała, Gaofei Shen & Tom Lentz.

0181 18

Grzegorz Chrupała @grzegorz.chrupala.me · 17/08/2025

Wrap-up of the Interspeech tutorial on interpretability for speech models by @mdhk.net

020 2

Grzegorz Chrupała @grzegorz.chrupala.me · 10/04/2025

We listened to speeches and marched through the streets. There were free fries 🍟. Tilburgers were curious and friendly. Protests are not really a thing in the Netherlands 😁

Photo of the crowd demonstrating in Tilburg.

051 5

Grzegorz Chrupała @grzegorz.chrupala.me · 20/03/2025

نوروز مبارک Happy Nowruz

050 5

Grzegorz Chrupała @grzegorz.chrupala.me · 19/02/2025

I think war in the Baltics is pretty much guaranteed at this point. I was in Lithuania a couple of years ago. Highly recommend to visit before it's overrun by the Russians again.

110 1