mwatkins

Achilles had just finished installing his new AI assistant when the Tortoise ambled by, looking bemused. "Another technological marvel, I see," said the Tortoise, peering at the softly glowing terminal. "Indeed!" exclaimed Achilles. "This is Claude, the latest in artificial intelligence. It can answer any question, write poetry, even discuss...

tl;dr: The glitch tokens ' petertodd' and ' Leilan' were studied extensively in the context of GPT-3 before its decommissioning at the end of 2023 [1] [2]. Here, the conception of these two tokens and their relationship is studied for GPT-2, GPT-2-xl and GPT-J (which share the same token vocabulary...

TL;DR A software tool is presented which includes two separate methods to assist in the interpretation of SAE features. Both use a "feature vector" built from the relevant weights. One method builds "definition trees" for "ghost tokens" constructed from the feature vector, the other produces lists of tokens based on...

TL;DR This research presents a novel method for exploring LLM embedding space using the Major Arcana of the tarot as archetypal anchors. The approach generates "archetype-based directions" in GPT-J's embedding space, along which words and concepts "mutate" in meaning, revealing intricate networks of association. These semantic mutation pathways provide insight...

tl;dr: Recently reported GPT-J experiments [1 2 3 4] prompting for definitions of points in the so-called "semantic void" (token-free regions of embedding space) were extended to fifteen other open source base models from four families, producing many of the same bafflingly specific outputs. This points to an entirely unexpected...

TL;DR This relates to the findings reported in my posts Mapping the Semantic Void parts I and II. By creating a custom embedding at the token centroid (the mean vector of all 50,257 GPT-J token embeddings), prompting the model to define it and considering logits, it's possible to construct a...

TL;DR: The original "semantic void" post documented a phenomenon tentatively described as a "stratified ontology" associated with a set of concentric hyperspherical shells in GPT-J embedding space. This post will consider the same phenomenon primarily via angular, rather than radial, variation. Both token embeddings and randomly sampled non-token embeddings are...