Cartographic exploration of a single steering direction within the Gemma-2-2B large language model at layer 20. This latent space linear direction was discovered by femi (0xf3mi) using the Gemma Scope Sparse Autoencoder from the latent with label Gemma-2-2B 20-14389: topics related to race, ethnicity, and social rights issues.

Clustering 250 maximally activating prompts centered on the strongest responses reveals a map of activating concepts. Orange highlighting indicates activating concepts within images and the connected network of neighboring activations. The label "DEI" is arguably a misnomer, but with SAE latents there's likely no word that exactly fits this "category" which is discovered using only unsupervised training.

Tom White (dribnet) is a New Zealand based visual artist focused on work combining AI and drawing systems for over 25 years. He studied computer graphics and visual design at the MIT Aesthetics and Computation group with professor John Maeda where his 1998 MS Thesis investigated the potential for multi-touch interfaces to provide better computer interfaces. His early work on drawing frameworks led to influential software that is still popular today - such as Processing, openFrameworks, and p5.js. His computational art has been exhibited at museums internationally including Cooper Union, MoMA, National Taiwan Museum of Fine Arts, SIGGRAPH, Ars Electronica, and the NeurIPS Art Gallery. He is also currently a lecturer teaching computational design and creative AI at the Victoria University of Wellington School of Design.

Medium: Two colour hand silkscreen, acrylic on paper accompanied by an interactive online interface and ERC-721 token

Dimensions:

Print: 45 18/25 x 45 18/25 cm | 18 x 18 in

Digital Image: 5400 × 5400 px

File size:

Contract Address:

Token Standard:

Blockchain: