Answers are grounded in this paper and cite their sources.
Kalle demonstrates next-word prediction in a bilingual chatbot, indicating a novel, transparent AI model.
We present Kalle, a fully client-side bilingual chatbot that performs next-word prediction and retrieval-augmented generation (RAG) using Kernel Ridge Regression (KRR) with Random Fourier Features — without neural networks, backpropagation, or gradient descent. The system makes the mathematical foundations of modern language models transparent: the same eigenvalue structure that governs convergence of the Neumann series, stability of PageRank, light transport in radiosity, and learning dynamics of neural networks is exposed in an interactive, inspectable system. Kalle serves 2,178 curated Q&A pairs bilingually, runs entirely in the browser via WebGL GPU (TensorFlow.js), and achieves functional conversational ability including multi-turn context and RAG — all through a single closed-form matrix solve. Demo: https://pmmathias.github.io/krr-chat/ Code: https://github.com/pmmathias/krr-chat Blog: https://ki-mathias.de/en/krr-chat-explained.html Changelog v2.0 (April 2026): Adds §6 on scalable training via absorber-stochasticization, including Block Preconditioned Conjugate Gradient (Block-PCG) benchmarks (Table 4) showing PyTorch-CPU/Float32 beats NumPy direct solve and matches GPU for moderate D, plus corpus-scaling analysis (§6.4). Adds Theorem 1 in §2.4 formalizing the early-stopping ≈ ridge-regression spectral-filter equivalence (λ ≈ 1/t). Title and scope unchanged from v1.
No takes yet. Share an insight, caveat, or question.
Mathias Leonhardt (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: