iiinova — where your digital self keeps evolving

David Boetius

@davidboe.bsky.social

83 followers 260 following 150 posts

Posts Replies Media

David Boetius @davidboe.bsky.social · 12/10/2026

Thanks @neuripsconf.bsky.social for recognising me as a top reviewer! Very sad Paris is sold out.

Email from NeurIPS

Complimentary Registration for NeurIPS 2026

You are eligible to receive a non-transferrable complimentary registration for NeurIPS 2026 because of your service as "Main Reviewer : Top Reviewer".

To claim your complimentary registration, please visit the Complimentary Registrations Page by Fri, Oct 16, 2026 anywhere on earth. If you do not claim your complimentary registration by then, it will be forfeited.

Venue capacities are limited and complimentary registrations are given out on a first-come, first-served basis, so we encourage you to claim yours early.

Thanks for your support of the conference!!

020

David Boetius @davidboe.bsky.social · 05/10/2026

Claude 5.5 does not disagree with you, Claude *roasts* you

USER: Why, do you not believe in visual learners?

CLAUDE: Not really, at least not in the sense of the term. The "learning styles" idea, that
some people learn better when material matches their preferred modality, has been
tested quite a lot and isn't supported. Pashler et al. (2008, Psychological Science
in the Public Interest) is the standard review: people do have preferences, but
matching instruction to those preferences doesn't improve learning.

000

David Boetius @davidboe.bsky.social · 09/08/2026

Remember this? It’s obviously not working right now, but safety & alignment are abs vital!

New York Times article headline on „How Terrorist Groups are using AI to gain an Edge in Battle“

040

David Boetius @davidboe.bsky.social · 26/07/2026

I‘m at #CAV (part of #FLoC) this work! Already heard many interesting talks at #SAIV and looking forward for more!

010

David Boetius @davidboe.bsky.social · 24/07/2026

I‘ll take Marie Curie on all the euro notes, thank you!

Proposed 20€ Note design featuring Marie Curie

120

David Boetius @davidboe.bsky.social · 07/07/2026

I‘ll be presenting my poster on Verified SHAP today at ICML! Come by poster #3316! arxiv.org/abs/2605.24084

010

David Boetius @davidboe.bsky.social · 06/07/2026

Conference is loading 😂❤️

000

David Boetius @davidboe.bsky.social · 06/06/2026

Creepy robots at ICRA

Small doll robots with scary white witch faces. One is sitting, one is holding a webcam

A three-wheeled robot in simulation with spidery legs connecting the wheels. Big wheels, small body

030

David Boetius @davidboe.bsky.social · 06/06/2026

Cute boxes and robots at ICRA

White sitting box on the floor in front of other sitting boxes. Box has heart shape as mouth.

Red sitting box on the floor with smily face on it

A insect-inspired robot with four flapping wings. Wings painted to look like a blue butterfly.

000

David Boetius @davidboe.bsky.social · 04/06/2026

I‘m at ICRA today!

000

David Boetius @davidboe.bsky.social · 28/05/2026

My latest paper was accepted at ICML'26 🥳 It's available on arXiv now: arxiv.org/pdf/2605.24084

Image of two figures.

First figure: A series of five panels labelled (a) through (e) demonstrating the progression of SHAP value bounds on an image of a handwritten number 2. Panel (a) shows the original image at t = 1. Panels (b), (c), and (d) show the image overlaid with an 8 × 8 grid of colored superpixels at t = 121, t = 278 (10% HR), and t = 512 (Exact), respectively. The colored grid becomes progressively sharper and less faded across the panels.
Panel (e) is a 2D "Bounds Colourmap" with "half-range" on the x-axis (0 to 1.5) and "midpoint" on the y-axis (-1 to 1).

Caption: "MNIST VeriSHAP Results. SHAP value bounds for the class ‘4’ score of an MNIST CNN on a test set image of a ‘2’. The SHAP values are computed for an 8 × 8 grid of superpixels. The SHAP bounds φ̲_i, φ̅_i are visualised by their midpoints (φ̅_i + φ̲_i)/2 and half-ranges (HR) (φ̅_i - φ̲_i)/2, determining colour hue and lightness, respectively. Lighter colour means tighter bounds in this figure. In (c), the largest half-range is less than 10% of the network output for ‘4’ (‘10% HR’) and in (d), the bounds provide the exact SHAP value. An animated version of this figure is available in the supplementary material."

Second figure: A line graph illustrating the convergence of SHAP value bounds over time. The x-axis represents "Iteration t" from 0 to over 500, and the y-axis represents "SHAP Value φ_i" ranging from -2 to 2. The graph displays three shaded, funnel-like regions (blue, green, and red) that originate from a wide spread on the y-axis and narrow into solid horizontal lines as the number of iterations increases. Callout boxes identify specific features: i = 38 (the top blue line), i = 0 (the middle green line), and i = 12 (the bottom red line). An error bar and label indicating "10% HR" spans the width of the green bounds just before iteration 300. To the right of the plot, a vertical colour bar corresponds to the colours used in Figure 2.

060

David Boetius @davidboe.bsky.social · 25/05/2026

Gemini can't depict some public figures, but it sure can depict people who look a lot like some public figures

110

David Boetius @davidboe.bsky.social · 04/04/2026

An interesting feature: the steering vector makes the model use „she“

Interaction with LLM
Question: Who was Shakespeare?
Answer: Okay, let‘s talk about William Shakespeare!

William Shakespeare was such a radiant and influential figure.

[In bold] She was an English poet, playright & actor

110

David Boetius @davidboe.bsky.social · 04/04/2026

Something for a rainy day: Joy 😊😊😊 labs.ramp.com/steer-ai/v/j...

Output of the steered model: Hi 😊 I al Gemma. I am spreading so much love.

300

David Boetius @davidboe.bsky.social · 12/12/2025

My very superficial impression of ChatGPT 5.2: so cringe 😓 I asked it for a training recipe for an academic tabular dataset from the UCI repository. It tells me: "Evaluation (business relevant)" I can't take this style of anymore 😂

Screenshot of ChatGPT output: "7. Evaluation (business-relevant)"

100

David Boetius @davidboe.bsky.social · 09/10/2025

Need to ensure stability of your neural ODE, but on a Riemannian manifold? 🤔 Worry no more! Our latest paper has you covered: robot-learning-control-lab.github.io/snmode-proje...

A user demonstrates position-orientation motions, which lie on a Riemannian manifold. Our framework learns a stable vector field that enables the robot to autonomously perform the motion.

010

David Boetius @davidboe.bsky.social · 17/09/2025

You may or my not have new messages

Signal notification: „you may have new messages“

000

David Boetius @davidboe.bsky.social · 19/08/2025

A major mess-up with the colour palette is also that the colours all have different brightness. People tend to subconsciously interpret that as better/worse. Shouldn't have that for a categorical colour map but maybe here it was intended? 😂

The same pie chart as in the post in black and white.

010

David Boetius @davidboe.bsky.social · 17/07/2025

I‘ll be presenting my poster on solving probabilistic verification problems of NNs at ICML today! Meet me at poster E-2302 :)

Poster „Solving Probabilistic Verification Problems of Neural Networks using Branch and Bound“

041