Albert Cardona
@albertcardona@mathstodon.xyz
How does the brain work? Someday, we'll figure it out. Group Leader, MRC LMB, and Professor, University of Cambridge, UK. #neuroscience #Drosophila #ScientificPublishing #academia #TrakEM2 #FijiSc #CATMAID #connectomics #connectome #vEM #iNaturalist #entomology Born at 335 ppm. Brains, signal processing, software and entomology: there will be bugs.
mathstodon.xyz
English language as a poor choice for training language models [*you don't say!*]. David Eichler proposes Basque, a language with very explicit consistency rules in its sentence structure that double as constraints for correctness.
"The hypothesis predicts the Basque-first model will show superior performance on tasks that require robust predicate-argument reasoning, independent of surface language fluency."
"We propose treating English — and any other natural language — as a last-mile
rendering layer. [...] Lossy, human-readable projections of representations built on the
substrate beneath them."
From: "Curriculum Order and the Geometry of Learned Representations: The Basque First Hypothesis", Eichler 2026.
https://zenodo.org/records/21396981
Expects a flywheel (positive feedback loop) rather than model collapse upon using the output of such model as input:
"Models trained on this substrate should generate outputs that are structurally different from the outputs of current models. When asked to produce arguments, we predict they will produce arguments with explicit logical structure. When asked to prove theorems, proofs with traceable inference chains. [...] These outputs — the proofs, the structured arguments, the logical derivations — are *training data.* [...] Reasoning quality compounds."
Quite the claims in this paper. Also:
"We are making a case for a research direction, not announcing results. The results do not exist yet."
#LLMs #LanguageModels #Basque #euskera
