About the Elemedex

The Elemedex (from Spanish ele-eme, “LM”, language model) is a first approach to artificial intelligence architectures for graduate students in philosophy, logic and linguistics. It is not a manual but a pocket encyclopedia for finding your way, shaped like a Pokédex because the Pokédex already solved the problem of presenting many different creatures in a common, comparable and memorable format.

It currently holds 54 species, 12 legendaries, 244 technical machines and 534 references.

How to use it

In the Pokédex, move with the arrow keys or the D-pad, filter by type or generation, and open an entry with Enter or the A button.
Each entry has a data sheet (height is the number of layers; weight, the memory its parameters take up), base stats, an anatomy diagram, moves, abilities, weaknesses, its evolution chain, an explanatory text and a professor’s note with the philosophical angle. It is signed by one of three professors; like Pokémon professors, who are named after trees starting with Oak, ours are too (more or less): Robledom, from roble (Spanish for oak) and a certain Pittsburgh inferentialist; Dendrett, from dendron (Greek for tree) and a certain builder of intuition pumps; and Rortiga, from ortiga (Spanish for nettle) and a certain liberal ironist. Full guide.
Every dotted link is a glossary term: hover to see the short definition, or click for the full entry.
Citations follow APA 7 and link to the sources. Page numbers for the core books were checked against the text.
On a phone the Elemedex becomes the device: the list is a wheel (whatever sits in the middle is selected), the D-pad and the A and B buttons below do what their labels say, each entry splits into screens you swipe through sideways, and tapping a dotted link or a citation shows the definition or the reference first. The round grille at the top reads the entry aloud. Turned sideways, it is held like a handheld console.
On a computer, / opens search from any page, and on an entry ← → go to the previous or next species. “Compare with” overlays another species’ stats, and every entry prints as a sheet for class. Tap the sprite to hear its cry; if it ever shows up in other colours, you got lucky.

Work in progress. The Elemedex is a living draft: the texts are still under review, there may be mistakes, and the content will change. If you find something wrong, please report it.

Credits

Idea by Mauro Santelli. Built by Claude Opus 5.5 (Anthropic) in Claude Code: the site’s design and code, the sprites, the diagrams, the lab demos, and the entries, drafted by several Claude agents from the cited sources, following a shared style and citation guide.

Gemini 3.1 Pro (Google), used through the Antigravity CLI, read the five reference books (Alammar and Grootendorst; Buckner; Witten, Frank, Hall and Pal; Chollet; Mitchell) and produced a concept index with page numbers and anchor phrases. That index was automatically checked against the text of each page and used as a map for finding passages; nothing was cited without checking it in the book. Later it also commented on the interface design. Gemini wrote none of the site’s content.

Changelog

Sep 30–Oct 1, 2026 First version: entries, glossary, legendaries, sprites and the lab.
Oct 3 Fact and bibliography review (four audits by Claude Fable 5.1), and the professors arrive.
Oct 4 Phone mode: the phone becomes the device.
Oct 4 Interface polish, texts untouched: Claude Opus 5.5 coordinated Claude agents that redesigned pages and tested the site on different screens; Claude Fable 5.1 and Gemini 3.1 Pro gave design critiques, and an automated review looked for bugs in the code.
Oct 11 Cross-check against Yang et al.’s (2023) tree of language models and Daniel Khashabi’s 2026 update of it: two new species (the SSM hybrid and the diffusion language model), updates to BERT, T5 and Mamba, eight new glossary terms, and a note on other trees in Evolution. Drafted and reviewed by Claude Opus 5.5 agents.
Oct 12 One more species, AlphaStar, and the term imitation learning.

On sources and method

The explanations rely mainly on four books: Alammar and Grootendorst, Hands-On Large Language Models (2024); Chollet, Deep Learning with Python (2021); Buckner, From Deep Learning to Rational Machines (2024); and Witten, Frank, Hall and Pal, Data Mining (2017), plus the original paper for each architecture and the philosophical literature cited in each entry. References come from Mauro’s Zotero library and, when they were not there, were checked against Crossref, arXiv, the ACL Anthology and publishers’ pages. An automatic checker makes sure every link and every citation key exists.

The sprites are drawn in code, pixel by pixel, and each tries to encode something about its architecture: BERT’s mask, the transformer’s eight eyes, the recurrent network’s ouroboros, the convolutional networks’ compound eyes.

Pokémon and Pokédex are trademarks of Nintendo, Game Freak and Creatures; this is a non-profit educational project paying homage to their format. The p(doom) song and its annotation are by osmarks and collaborators.