iiinova — where your digital self keeps evolving

📄 The Geometry of Categorical and Hierarchical Concepts in LLMs (ICLR 2025, Best Paper Award at the ICML 2024 workshop on Mech Interp) arxiv.org/abs/2406.01506 📁 github.com/KihoPark/LLM... LLMs represent categorical concepts as polytopes and hierarchical relations as orthogonality

Figure 1: In the representation spaces of LLMs, hierarchically related concepts (such as plant → animal and mammal → bird) live in orthogonal subspaces, while categorical concepts are represented as polytopes. The top panel illustrates the structure; the bottom panels show the measured representation structure in the Gemma LLM. See Section 5 and Appendix A for details.

Figure 5: WordNet noun hierarchy is encoded in the orthogonal structure predicted by statement (a) in Theorem 8. We plot the cosine similarity between a child-parent vector and a parent vector for each feature in the hierarchy (blue). As predicted, this value is close to 0. The left plot uses all data for representation estimation, and the right plot uses only 70% independently selected for each synset. We include baselines where a randomly selected feature is used as the parent (orange) and where the embeddings are shuffled (green) as controls for the possibility that the orthogonality is a simple byproduct of high-dimensional geometry, or of the set inclusion relationships used in estimation—see main text for details. See Appendix F for an analogous plot for statement (d).

100

📄 Understanding Wikidata Qualifiers: An Analysis and Taxonomy (2026) arxiv.org/abs/2603.11767

A taxonomy of qualifiers on Wikidata

First level splits into context, epistemic/uncertainty, structural, and additional.

Sublevels of context: temporal, spatial, subject modifier

Beneath epistemic: uncertainty quantification

Beneath structural: field of a structure, meta modelling

Beneath additional: sequence, provenance, causality, object/subject statement relation, subproperty, external entity description, ‘other’

100

Who called it knowledge graph query completion and not “now draw the rest of the OWL”

000

oh there’s just unadulterated Claudeslop on arXiv now huh

What we concede. The general shape is partly a consequence of the design. Backward traversal from 64 seeds through a multiparent DAG will broaden and then narrow, and one should not be impressed by broadening and narrowing as such.
What survives. The hourglass is not one shape; it is a family of shapes that differ node by node, and the differences are informative even if the general form of the family is not. Gauss, Hilbert, and Euler show thickness ratios of 15.1, 18.5, and 14.6, all constriction-and-expansion profiles. The pattern is a property of hub nodes in general, not of Leibniz specifically, and that is precisely the point: the tracer-set explanation predicts a global shape, but the variation between nodes is what carries information.
Newton, on the same graph under the same traversal with the same seeds, shows no constriction at all.
If the hourglass were purely a design artifact, Newton would have one too.
What distinguishes Leibniz within the family is not the ratio, where Hilbert is higher, but the combination of a high ratio with a large absolute path count: 47 of 64 lineages, against four for Newton. Ratio and volume are separately obtainable from the design; their conjunction is what we are pointing at.
What would settle it. The natural test is the one flagged in Section 5.2: a systematic comparison of thickness ratio and path volume across all high-traffic nodes in each century, so that the position of Leibniz within the distribution is visible rather than asserted. Our deposited data support this and we have not done it.

0120

Who called it dead internet theory and not asemantic web

000

David Mimno @dmimno.bsky.social · 29/09/2026

📄 Schemas for graph data (2026) dl.acm.org/doi/pdf/10.1...

dl.acm.org

012

Reposted by Louis Maddox

Sparsity is back! Sparsity is everywhere in language, but managing it requires overhead. For the last ~15 years it was faster to just do the dense operation, knowing most of it was useless. That sparse operations work is a huge shift in the "do something clever" vs. "spend more money" tradeoff.

2305

> pyarrow’s iter_batches returns an invalid nested batch here, a known weakness when it slices deep list-of-struct columns > Found the bug: RecordBatch.cast corrupts the nested struct … [bisects] … > Found it: the offset gets applied twice to the null-typed child — Opus 5.5 🐐

000

📄 Extraction of Validating Shapes from Very Large Knowledge Graphs (2023) www.vldb.org/pvldb/vol16/... 📁 QSE github.com/dkw-aau/qse Efficiently mines validating shapes [SHACL or ShEx] from very large existing KGs [English Wikidata] with support and confidence thresholds

vldb.org

001

📄 Characteristic sets (Neumann & Moerkotte, 2011) www.csd.uoc.gr/~hy561/paper... They define the characteristic set of a subject as the set of predicates it emits. Entities that share that set tend to be semantically alike, which gives RDF the "latent soft schema" it lacks on paper.

page 3 of the paper

V. CHARACTERISTIC SETS
A. Plain Characteristic Sets
Standard histograms do not capture correlations between join predicates at all, and even using dependent selectivities. as shown in the previous section. only helps for the first join. Therefore, we propose a radically different approach for estimating the selectivity of joins in RDF graphs.
Many of the correlations we observe during selectivity estimation stem from the fact that RD uses multinle trioles to describe the same object. Consider the following sample triples describing a book:
(os.title, The Tree and I). (o1.author,R. Pecker).
(01.author,D. Owl), (01-year, 1996).
All four triples describe the same entity, and accordingly. the individual triple patterns (derived by replacing o, with a variable ?b) are strongly correlated. The year is somewhat an exception here, but for the other three triples, searching just for one triple pattern is nearly as selective as searching for all of them. Obviously, this is true for most books. In general. many entities can be uniquely identified by a true subset of their emitting edges.
In most RDF data sets, these emitting edges exhibit a certain structure. While RDF is used usually without a fixed schema, some kind of latent soft schema in the data frequently occurs:
Books tend to have authors and titles, etc. While we might not be able to clearly classify an entity as "book" (duc to the lack of schema information), we observe that we can characterize an entity by its emitting edges. For each entity s occurring in an RDF data set R. we define its characteristic ser as follows:
Sc(s) := {p|3o: (s,p,o) € R).
For many RDF data sets, entities that have the same characteristic set tend to be semantically similar. This is not surprising, as RDF encodes all semantics using edges. But this enables us to predict selectivities based upon the involved…

000

[capture the lightcone voice] “we must compile the TBox”

000

to be held is to be carried — Clopus 5.5

000

Louis Maddox @permutans.bsky.social · 28/09/2026

Using a Sonnet model to see the dumb default response to your prompt is MTP drafting for chat UIs lol

000

One (1) wild brainwave acquired

000

Positional Encodings and Group Theory | 3Blue1Brown and Alok Puranik

Grant Sanderson of 3b1b interviewing Jane Street researcher Alok Puranik about positional encodings and group theory youtu.be/k86eUj4hgdc 📝 Using group theory to explore the space of positional encodings for attention blog.janestreet.com/using-group-...

youtu.be

YouTube video by Jane Street

010

so is someone gonna give the NeurIPS programme the AI gen hype video treatment or

000

`tpu cuu 5`… Cthulhu coded shell incantation

000

Network bandwidth bottlenecked at last

000

Louis Maddox @permutans.bsky.social · 26/09/2026

2026 spoiled user mindset: "how dare this semantically useful info be computationally any less than convenient"

000

The amnesty on bugs as you take a library you built from zero into the correct/performant phase is… humbling

110

Louis Maddox @permutans.bsky.social · 26/09/2026

is it idiomatic for data normalisation libraries to talk of invariants and determinants? well it is now

110

Louis Maddox @permutans.bsky.social · 25/09/2026

eyeballing 60% of WikiData Swahili values without a label in the language

000

Louis Maddox @permutans.bsky.social · 25/09/2026

Filleting out the entirety of English Wikidata might still change the game (Claude parquetmaxxing self-actualisation incoming)

000

Louis Maddox @permutans.bsky.social · 24/09/2026

Revisited schema inference in genson-core // polars-genson + made big speed gains + memory reductions 🏁 6+ day estimated processing time for Wikidata now <24h

Table of wall times per file for individual Wikidata files falling from 12.9 - 35.5 seconds down to 1.8 - 4.7s

100

Louis Maddox @permutans.bsky.social · 24/09/2026

Uncanny how accurately Claude can estimate the exact wall time change of individual perf optimisations

000

Packaged this, after some tape allocation adjustments it parses up to 3x faster than simd-json, gains most on files with large rows (Wikidata claims benchmark) 📁 cuJSON-rs github.com/lmmx/cuJSON-rs 🦀 Rust crates.io/crates/cujson 🐍 Python pypi.org/project/cujs...

Benchmarks showing speedup from simd-json (1.3-3 GB/s) to cujson (4-4.5 GB/s)

020

Claude yearns to tape rotate

000

Accidentally stumbled on what seems to be the same thing Claude Code (etc) use for displaying markdown: `cargo binstall mdcat-ng` (gives you mdcat and mdless)

110