<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Saud Alotaibi’s Opinions</title><link>https://saudalotaibi.com/blog</link><description>AI news, comparisons, and explanations.</description><item><title>KV caches: the memory behind a faster conversation</title><link>https://saudalotaibi.com/blog/articles/kv-cache-memory-explained</link><guid>https://saudalotaibi.com/blog/articles/kv-cache-memory-explained</guid><pubDate>Wed, 21 Oct 2026 12:00:00 GMT</pubDate><description>Why a language model stores keys and values, how the cache grows, and what a simple memory calculation can tell you.</description></item><item><title>Mixture of experts: fewer active weights, more routing decisions</title><link>https://saudalotaibi.com/blog/articles/mixture-of-experts-routing-explained</link><guid>https://saudalotaibi.com/blog/articles/mixture-of-experts-routing-explained</guid><pubDate>Wed, 21 Oct 2026 12:00:00 GMT</pubDate><description>A practical explanation of mixture-of-experts (MoE) routing, load balancing, and why sparse computation does not automatically mean lower latency.</description></item><item><title>Speculative decoding: let a small model draft, then verify</title><link>https://saudalotaibi.com/blog/articles/speculative-decoding-explained</link><guid>https://saudalotaibi.com/blog/articles/speculative-decoding-explained</guid><pubDate>Wed, 21 Oct 2026 12:00:00 GMT</pubDate><description>How draft-and-verify generation can preserve a target distribution, and why acceptance rate alone is not a speed benchmark.</description></item><item><title>Prompt injection: a useful document can still contain untrusted instructions</title><link>https://saudalotaibi.com/blog/articles/prompt-injection-trust-boundaries</link><guid>https://saudalotaibi.com/blog/articles/prompt-injection-trust-boundaries</guid><pubDate>Wed, 21 Oct 2026 12:00:00 GMT</pubDate><description>Why retrieval does not create authority, what a detector can and cannot prove, and how to design a smaller permission boundary.</description></item><item><title>RAG, fine-tuning, or a longer context? Start with the failure</title><link>https://saudalotaibi.com/blog/articles/rag-fine-tuning-or-long-context</link><guid>https://saudalotaibi.com/blog/articles/rag-fine-tuning-or-long-context</guid><pubDate>Wed, 21 Oct 2026 12:00:00 GMT</pubDate><description>Three approaches change different parts of an AI system. A worked decision process helps separate missing evidence from unwanted behavior.</description></item><item><title>How language models are trained: pretraining, LoRA, and human feedback</title><link>https://saudalotaibi.com/blog/articles/how-language-models-are-trained</link><guid>https://saudalotaibi.com/blog/articles/how-language-models-are-trained</guid><pubDate>Sun, 04 Oct 2026 12:00:00 GMT</pubDate><description>A visual guide to pretraining, supervised fine-tuning, LoRA, QLoRA, RLHF, and DPO, with worked calculations and practical evaluation advice.</description></item><item><title>How deep neural networks work: a visual guide to the main concepts</title><link>https://saudalotaibi.com/blog/articles/how-deep-neural-networks-work</link><guid>https://saudalotaibi.com/blog/articles/how-deep-neural-networks-work</guid><pubDate>Sun, 04 Oct 2026 12:00:00 GMT</pubDate><description>Follow a neuron by hand, count a small network’s parameters, calculate a gradient update, and learn how to judge generalization.</description></item><item><title>How transformers work: tokens, attention, and text generation</title><link>https://saudalotaibi.com/blog/articles/how-transformers-work</link><guid>https://saudalotaibi.com/blog/articles/how-transformers-work</guid><pubDate>Sun, 04 Oct 2026 12:00:00 GMT</pubDate><description>A separate visual guide to embeddings, Q/K/V, attention arithmetic, causal masks, transformer blocks, and the KV cache.</description></item><item><title>Gemini 4 Argon is here. Broad access is a different question.</title><link>https://saudalotaibi.com/blog/articles/gemini-4-argon-announcement</link><guid>https://saudalotaibi.com/blog/articles/gemini-4-argon-announcement</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>Google&#39;s newest frontier announcement puts long-running work at the center. The rollout matters as much as the capability claims.</description></item><item><title>OpenAI DevDay 2026: the shift from answering to carrying work</title><link>https://saudalotaibi.com/blog/articles/openai-devday-2026-what-mattered</link><guid>https://saudalotaibi.com/blog/articles/openai-devday-2026-what-mattered</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>The September 29 event reached beyond a model launch. Here is the useful thread connecting the developer and product announcements.</description></item><item><title>GPT-6.1 Sol: watch the cost of finishing, not just generating</title><link>https://saudalotaibi.com/blog/articles/gpt-6-1-sol-cost-per-finished-task</link><guid>https://saudalotaibi.com/blog/articles/gpt-6-1-sol-cost-per-finished-task</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>OpenAI&#39;s new Sol release makes capability per dollar the headline. The important comparison includes retries and review.</description></item><item><title>Chinese AI is not one race: four releases worth separating</title><link>https://saudalotaibi.com/blog/articles/chinese-ai-models-october-2026-watchlist</link><guid>https://saudalotaibi.com/blog/articles/chinese-ai-models-october-2026-watchlist</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>DeepSeek, Qwen, Kimi, and GLM are competing across different tasks. Treating them as interchangeable hides the useful differences.</description></item><item><title>Qwen-Image-2.1: transparency could matter more than another sample gallery</title><link>https://saudalotaibi.com/blog/articles/qwen-image-2-1-transparent-assets</link><guid>https://saudalotaibi.com/blog/articles/qwen-image-2-1-transparent-assets</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>The image release points toward practical asset creation. The next test is whether the output survives editing and reuse.</description></item><item><title>DeepSeek&#39;s Flash alias is a reminder to date every model comparison</title><link>https://saudalotaibi.com/blog/articles/deepseek-flash-alias-version-check</link><guid>https://saudalotaibi.com/blog/articles/deepseek-flash-alias-version-check</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>A stable API name can point to a newer model. That is convenient for integration and important for reproducibility.</description></item><item><title>AI model comparison, October 2026: benchmarks, costs, and practical choices</title><link>https://saudalotaibi.com/blog/articles/gpt-gemini-claude-comparison-october-2026</link><guid>https://saudalotaibi.com/blog/articles/gpt-gemini-claude-comparison-october-2026</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>A detailed guide to GPT, Gemini, Claude, GLM, Kimi, DeepSeek, and Qwen, with reported benchmark scores, worked costs, and a practical evaluation plan.</description></item><item><title>Kimi K2.8 Preview: distinguish the model update from the coding workspace</title><link>https://saudalotaibi.com/blog/articles/kimi-k2-8-preview-model-product</link><guid>https://saudalotaibi.com/blog/articles/kimi-k2-8-preview-model-product</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>Kimi&#39;s September updates show why model identity, reasoning settings, and application features should be tracked separately.</description></item><item><title>Sonnet 5.5&#39;s efficiency claim needs the right denominator</title><link>https://saudalotaibi.com/blog/articles/claude-sonnet-5-5-efficiency</link><guid>https://saudalotaibi.com/blog/articles/claude-sonnet-5-5-efficiency</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>Anthropic reports faster output and lower task costs. That is different from announcing a lower price for every token.</description></item><item><title>GLM-5.3: the migration note is more useful than the flagship label</title><link>https://saudalotaibi.com/blog/articles/glm-5-3-reasoning-migration</link><guid>https://saudalotaibi.com/blog/articles/glm-5-3-reasoning-migration</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>Z.ai&#39;s documentation describes always-enabled reasoning and text-only input. Those details affect real integrations.</description></item><item><title>Before you build an AI agent, fix the workflow</title><link>https://saudalotaibi.com/blog/articles/build-a-useful-ai-workflow</link><guid>https://saudalotaibi.com/blog/articles/build-a-useful-ai-workflow</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>The most useful AI project often starts with a smaller question: which part of this task is actually slowing someone down?</description></item><item><title>RAG makes more sense when you start with a handbook</title><link>https://saudalotaibi.com/blog/articles/rag-explained-with-a-handbook</link><guid>https://saudalotaibi.com/blog/articles/rag-explained-with-a-handbook</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>A clear example of retrieval-augmented generation, including where retrieval fails and when a system should say it cannot answer.</description></item><item><title>What an embedding can tell you—and what it cannot</title><link>https://saudalotaibi.com/blog/articles/embeddings-without-the-mystery</link><guid>https://saudalotaibi.com/blog/articles/embeddings-without-the-mystery</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>Why a numerical representation helps search find related text, and why a high similarity score is not a fact-check.</description></item><item><title>A simple evaluation sheet for AI answers</title><link>https://saudalotaibi.com/blog/articles/evaluate-an-ai-answer</link><guid>https://saudalotaibi.com/blog/articles/evaluate-an-ai-answer</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>Replace “this sounds good” with a repeatable review of correctness, evidence, completeness, and uncertainty.</description></item><item><title>Local AI or a cloud service? Begin with constraints.</title><link>https://saudalotaibi.com/blog/articles/local-or-cloud-ai</link><guid>https://saudalotaibi.com/blog/articles/local-or-cloud-ai</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>A practical decision framework based on data, hardware, latency, maintenance, and the actual task—not model hype.</description></item><item><title>Write a prompt like a useful brief</title><link>https://saudalotaibi.com/blog/articles/write-a-prompt-like-a-brief</link><guid>https://saudalotaibi.com/blog/articles/write-a-prompt-like-a-brief</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>A repeatable way to specify the task, evidence, boundaries, and output without accumulating pages of conflicting instructions.</description></item><item><title>Before you paste a file into an AI tool, make a smaller copy</title><link>https://saudalotaibi.com/blog/articles/share-less-data-with-ai</link><guid>https://saudalotaibi.com/blog/articles/share-less-data-with-ai</guid><pubDate>Sat, 03 Oct 2026 12:00:00 GMT</pubDate><description>A practical data-minimization routine for everyday AI use, including identifiers, hidden fields, and useful substitutes.</description></item></channel></rss>