System One Lite is a local LLM project that converts language models into typed decision engines, eliminating free-form text generation in favor of direct probability scores for predefined options. It runs on Apple silicon using open-weight models and returns structured answers without generated tokens, designed for routing, ranking, and classification tasks in software systems.
GLiClass is an open-source zero-shot sequence classification model inspired by GLiNER that achieves comparable performance to cross-encoder models while being 10 times faster through single forward pass classification. It supports hierarchical labels, in-context examples, custom prompts, and long document chunking for improved accuracy and flexibility.
WangNet is a lightweight 1.8 MB neural network that classifies whether numbers are Numberwang, with inference in pure Python requiring no dependencies. It supports 11 languages, achieves 88.9% accuracy on held-out test cases, and can be run locally or via a hosted Hugging Face demo.
A Hacker News discussion questions why embeddings are predominantly used for retrieval and RAG systems despite being capable of clustering, recommendations, anomaly detection, and classification. The author argues that embeddings' semantic capabilities remain largely untapped and wonders whether retrieval dominates because it's easier to productize than other use cases.
This article explores knowledge distillation applied to a bigram language model, the simplest possible sequence model. The author demonstrates that distillation does not improve the bigram's learned distribution and that standard training achieves equivalent results with sufficient data, but the analysis reveals what soft targets change and what they preserve.