Artificial Intelligence

Almost every current language model now reasons before answering. Using Sebastian Raschka’s work as a technical reference, this article explains how a reasoning model is evaluated, scaled at inference time, trained with RL, and distilled from scratch, with Python code examples.

Model parallelism lets Kimi K3, with 3 trillion parameters, run across 59 Cerebras WSE-3T wafers instead of just one. This guide explains how weights and attention get split across chips, with code examples to test the same principle on your own GPU.

Dessa is Block’s model that learned to hold voice conversations by listening only to audio, without a single word of text in its training. This article explains what a full-duplex voice model is, how it differs from Siri or Alexa, and what the company’s published demos show.

AI writing clichés (em dashes, triplets, “it’s not X, it’s Y”) come from RLHF training, not from a decision the model makes. Qwen3-4B-deslop trains them out of a 4-billion-parameter model, and anyone can run it free on Ollama to compare before and after.

Latent reasoning lets an AI agent think in continuous vectors instead of words. We use Stanford’s Hybrid Reasoning Agents paper as a case study to explain what it is, how it works, and when it’s worth using.

Before buying a GPU or a Mac to run AI models on your own machine, it pays to understand what determines local LLM speed. Memory bandwidth, quantization, and MoE architecture explain why two machines with similar specs perform so differently.