Tagged AI
13 articles tagged AI.
Backend & Databases8 min readYour Token Count Is Not a Property of Your Text
A token count depends on the language, the model and the content. The same text varies 15x by language and 30% between model versions.
The TweakPad Team
Backend & Databases8 min readThe BM25 Versus Embeddings Argument Measures the Wrong Gap
Two benchmarks agree across five years. Choosing BM25 or embeddings shifts recall a little. A reranking stage shifts it three times more.
The TweakPad Team
Backend & Databases8 min readBinary Quantization Works on Your Embeddings, or It Does Not
Binary quantization keeps recall on contrastive embeddings and fails on others. The outcome belongs to your model, not to the method.
The TweakPad Team
Backend & Databases9 min readStructured Outputs Guarantee a Grammar, Not Your Schema
Structured outputs constrain replies to a grammar, and several JSON Schema keywords never reach it. What the guarantee really covers.
The TweakPad Team
Backend & Databases8 min readLost in the Middle Is an Argument for Retrieving Less
Lost in the middle is usually read as a reordering problem. The paper's own numbers argue harder for retrieving fewer documents.
The TweakPad Team
Backend & Databases8 min readTruncating Embeddings Works Without Matryoshka, Up to a Point
Truncating embeddings supposedly needs Matryoshka training. A 2026 study across 24 tasks found plain truncation holds until you cut 80%.
The TweakPad Team
Backend & Databases8 min readSemantic Chunking Wins on the Datasets Built to Flatter It
Semantic chunking beat fixed-size only on synthetically stitched documents. On real ones, fixed-size won every document retrieval test.
The TweakPad Team
AI & Prompt Engineering5 min readScaled Content Abuse: Google Bans Purpose, Not the Tool
The scaled content abuse policy names purpose and volume, not the tool. What the wording prohibits, and the test that settles it.
The TweakPad Team
- AI & Prompt Engineering5 min read
Why LLMs Have Context Windows at All
A context window is not a storage limit. It comes from attention costing quadratic compute and position encodings.
The TweakPad Team
- Web Development11 min read
llms.txt vs robots.txt: Only One of Them Blocks AI Crawlers
llms.txt shows no measurable effect on AI citations, and Content-Signal is ignored by every crawler. Only robots.txt is honoured.
The TweakPad Team
- AI & Prompt Engineering3 min read
AI Coding Assistants: Where They Help and How to Review Their Output
What AI coding tools are genuinely good at, the failure modes to expect, and a checklist for reviewing the code they generate.
The TweakPad Team
- AI & Prompt Engineering3 min read
LLM Context Windows: What They Are and Why Bigger Is Not Always Better
How context windows work, what tokens cost, the lost-in-the-middle effect, and how to fit a real codebase into one.
The TweakPad Team
- AI & Prompt Engineering4 min read
Prompt Engineering for Developers: What Actually Changes Output Quality
The prompting techniques that measurably improve coding results — context, specificity, examples — without the mysticism.
The TweakPad Team