AI/ML●●Solid
Tinyvision:-Building Ultra-Lightweight Models for Image Tasks
Ultra-lightweight CNNs achieving 86% accuracy with under 12k parameters.
Big BrainNiche Gem
saptakbhoumik3
322mo ago

In-browser LLM inference, but unclear if 100k tok/sec is real or marketing.
Developers and AI enthusiasts experimenting with in-browser LLM inference.
WebLLM · Ollama.js · TensorFlow.js + transformers.js stacks
Ultra-lightweight CNNs achieving 86% accuracy with under 12k parameters.
First LLM with per-token interpretability tracing input, concepts, and training provenance.
Native multilingual training covers GDPR Article 9 categories others skip.
Parallel token decoding beats autoregressive LLMs on throughput, if the math holds up.
Prefix notation language that cuts LLM token usage by 70% compared to Python or C.
5.6x token compression with click/type/select interaction beats read-only Firecrawl, Jina.