Grimoire

Your AI Has Access to Everything — and Understands Nothing

The $0.55 Research Session: How Splitting NotebookLM and Claude Cuts Token Costs 17x

Feeding 47 papers into Claude costs $9.59 per 5-round session. Using NotebookLM for retrieval and Claude for reasoning drops that to $0.55. The secret: not a single word of your papers ever enters Claude.

Your AI Model Crashed Mid-Training? It Might Be Manifold Tearing

MoE models frequently crash mid-training without warning. DeepSeek V4 identified the cause from a differential geometry perspective: manifold tearing. Three stabilization techniques make training predictable.

The Atomic Rules of AI Skill Composition: From Atoms to Compounds

A single AI Agent is reliable. But your cognitive bandwidth caps at 5 parallel threads. Skill Graphs 2.0 introduces a three-layer architecture — atoms (deterministic), molecules (fixed sequence), compounds (strategic autonomy) — turning 5 compound-level tasks into 500 work units. This article breaks down the framework, practical principles, and known limitations.

250 Documents Can Plant a Permanent Backdoor in AI: The Real Threat of Training Data Poisoning

Anthropic, the UK AI Safety Institute, and the Alan Turing Institute have demonstrated that inserting just 250 carefully crafted documents into training data can plant a permanent backdoor in LLMs ranging from 600M to 13B parameters. Bigger models and more data do not dilute the poison.

From Prompt Engineering to System Engineering: Three Paradigm Shifts in AI Workflows

The early days of mainstream AI use had a particular competitive currency: prompt quality. Who could write the most precise, most creative, most effective instr