Multi-head latent attention vs GQA and MQA explained: how each shrinks the KV cache, the low-rank projection math behind DeepSeek MLA, memory arithmetic.
Mamba and state space models compared with Transformers: selective scan, linear-time inference, KV-cache savings, hybrid designs like Jamba, and when SSMs.
GRPO and RLVR explained: how group-relative policy optimization and verifiable rewards train reasoning LLMs, with the math, a minimal training loop and.
Trace LLM calls and AI agents with OpenTelemetry GenAI semantic conventions: span and metric attributes, agent spans, content capture and Collector setup.
Perceptron Mk1.5 explained: an embodied reasoning model for image, video and audio. Confirmed capabilities, pricing, open questions and how to evaluate it.
LTX from Lightricks explained: diffusion transformer video architecture, open weights license, resolution and length limits, and the GPU needed to run it locally.