KV caches: the memory behind a faster conversation
Why a language model stores keys and values, how the cache grows, and what a simple memory calculation can tell you.
Read articleWhat changed in AI, why it matters, and how it works. Practical analysis of new models, careful comparisons, and clear explanations of the ideas behind them.
Browse topics
Why a language model stores keys and values, how the cache grows, and what a simple memory calculation can tell you.
Read articleTwo standalone visual guides explain the foundations, step by step.
Deep neural networks25 articles
Why a language model stores keys and values, how the cache grows, and what a simple memory calculation can tell you.
Read articleA practical explanation of mixture-of-experts (MoE) routing, load balancing, and why sparse computation does not automatically mean lower latency.
Read articleHow draft-and-verify generation can preserve a target distribution, and why acceptance rate alone is not a speed benchmark.
Read articleWhy retrieval does not create authority, what a detector can and cannot prove, and how to design a smaller permission boundary.
Read articleThree approaches change different parts of an AI system. A worked decision process helps separate missing evidence from unwanted behavior.
Read articleA visual guide to pretraining, supervised fine-tuning, LoRA, QLoRA, RLHF, and DPO, with worked calculations and practical evaluation advice.
Read articleFollow a neuron by hand, count a small network’s parameters, calculate a gradient update, and learn how to judge generalization.
Read articleA separate visual guide to embeddings, Q/K/V, attention arithmetic, causal masks, transformer blocks, and the KV cache.
Read articleGoogle's newest frontier announcement puts long-running work at the center. The rollout matters as much as the capability claims.
Read articleThe September 29 event reached beyond a model launch. Here is the useful thread connecting the developer and product announcements.
Read articleOpenAI's new Sol release makes capability per dollar the headline. The important comparison includes retries and review.
Read articleDeepSeek, Qwen, Kimi, and GLM are competing across different tasks. Treating them as interchangeable hides the useful differences.
Read articleThe image release points toward practical asset creation. The next test is whether the output survives editing and reuse.
Read articleA stable API name can point to a newer model. That is convenient for integration and important for reproducibility.
Read articleA detailed guide to GPT, Gemini, Claude, GLM, Kimi, DeepSeek, and Qwen, with reported benchmark scores, worked costs, and a practical evaluation plan.
Read articleKimi's September updates show why model identity, reasoning settings, and application features should be tracked separately.
Read articleAnthropic reports faster output and lower task costs. That is different from announcing a lower price for every token.
Read articleZ.ai's documentation describes always-enabled reasoning and text-only input. Those details affect real integrations.
Read articleThe most useful AI project often starts with a smaller question: which part of this task is actually slowing someone down?
Read articleA clear example of retrieval-augmented generation, including where retrieval fails and when a system should say it cannot answer.
Read articleWhy a numerical representation helps search find related text, and why a high similarity score is not a fact-check.
Read articleReplace “this sounds good” with a repeatable review of correctness, evidence, completeness, and uncertainty.
Read articleA practical decision framework based on data, hardware, latency, maintenance, and the actual task—not model hype.
Read articleA repeatable way to specify the task, evidence, boundaries, and output without accumulating pages of conflicting instructions.
Read articleA practical data-minimization routine for everyday AI use, including identifiers, hidden fields, and useful substitutes.
Read articleNo articles match that search. Try a shorter phrase or clear the search.