Drips.
One AI idea at a time, explained with small interactive labs.
02All drips48 published · 52 indexed
Core Concepts
11 entries01LoRA & qLoRAFine-tune massive LLMs on consumer hardware. Learn about Low-Rank Adaptation and 4-bit Quantization.Read02TokenizationBefore AI can read, it must chop. Learn how text is broken down into the fundamental atoms of meaning.Read03LLM SamplingHow do LLMs decide what to say next? Explore greedy vs. probabilistic sampling and log probabilities.Read04Context EngineeringBeyond the prompt: Curate the perfect information to feed your LLM's limited attention span.Read05Prompt EngineeringLearn how Zero-Shot, Few-Shot, and Chain-of-Thought prompting steer LLM probabilities.Read06KV Cache (Inference)Why doesn't ChatGPT re-read your whole chat every time it types a word? Memory optimization explained.Read07QuantizationSame model, a quarter of the memory — if you protect the 1% of weights that matter. FP16→INT4, the AWQ trick, and which of GPTQ/AWQ/GGUF/FP8 to reach for in 2026.In queue08Naive BayesPredicting the future by assuming simplicity. Learn how this probabilistic algorithm uses Bayes' Theorem for classification.Read09Random ForestStrength in numbers. See how an ensemble of diverse decision trees can vote to make robust predictions.Read10Support Vector MachinesThe classic algorithm that finds the widest possible street between two classes of data.Read11Recommender SystemsFrom collaborative filtering to matrix factorization: how Netflix knows what you want before you do.Read
Architectures
8 entries01Seq2Seq (Encoder-Decoder)The architecture that solved translation before Transformers. Learn about the Context Vector bottleneck.Read02AlexNet (The Big Bang)The 2012 breakthrough that started the Deep Learning era. ReLU, Dropout, and GPUs.Read03Bahdanau AttentionBefore Transformers, RNNs learned to focus. The mechanism that solved the bottleneck problem.Read04Recurrent Neural NetworksUnderstand how AI processes sequential data using hidden states and memory loops.Read05BERT & BidirectionalitySee how 'Masked Language Model' training enables deep context from both directions.Read06Vision Transformer (ViT)Learn how breaking images into 16x16 patches allowed pure Transformers to beat CNNs.Read07U-Net & Skip ConnectionsDiscover the architecture behind precise image segmentation and preserving fine details.Read08YOLO (Object Detection)'You Only Look Once': Real-time object detection framed as a single regression problem.Read
Agents & RAG
18 entries01Multi-MCP ArchitectureNew Research: 10 connected MCP servers = ~8K tokens of schemas loaded before your prompt starts. Many-small > one-big, and how to lazy-load.Read02Agent Long-Term MemoryNew Research: Your agent forgets because nobody wrote anything down. Three places to put memory — and why only a temporal store survives a fact that changes.In queue03Eval-Driven DevelopmentNew Research: Prompts are code. The four-stage CI pipeline (local dev → PR check → deploy gate → production monitor) that replaces saturated benchmarks.Read04Verifying AI CodeNew Research: 66% of developers said the same thing — AI code that is almost right. The verification cascade and the four error classes it catches.Read05Harness EngineeringNew Research: 60% of all LLM errors are rate limits, not model errors. The five harness layers that decide reliability.Read06Agentic Context EngineeringNew Research: Why your agent forgets the rules by turn 15 — and the four operations (Write, Select, Compress, Isolate) production teams converged on.Read07Agentic ETLNew Research: What 1,200+ production deployments taught us about putting LLM agents into the extract-transform-load loop — and the two-layer sandwich that survives contact with real data.Read08Build: A Multi-MCP RouterBuild Along: 80 lines, no dependencies. Three small MCP servers, the tool-bloat problem they create together, and the router that fixes it. Companion code to Multi-MCP Architecture.Read09Build: An Eval HarnessBuild Along: the smallest honest eval suite — dataset, system under test, scorers, and a CI gate that exits non-zero below threshold. Companion code to Eval-Driven Development.Read10Build: An Agentic ETL PipelineBuild Along: 70 lines that show the two-layer sandwich — a fuzzy LLM transform wrapped in deterministic validation on both sides. Companion code to Agentic ETL.Read11Build: An AI Code VerifierBuild Along: the generate → verify → feed-failure-back loop in 60 lines. The model writes; the verifier decides. Companion code to Verifying AI Code.Read12HyPA-RAG (Legal AI)New Research: A hybrid, parameter-adaptive RAG system designed specifically for high-stakes legal applications.Read13Agentic RAGWhen RAG gets smart. Learn how adding an autonomous agent loop enables multi-hop reasoning and self-correction.Read14Agentic Hybrid RAGNew Research: Combining GraphRAG and VectorRAG with an autonomous router for scientific literature review.Read15Agentic Design PatternsGoogle Cloud Architecture: From simple prompts to complex multi-agent systems.Read16RAG (Retrieval-Augmented)Give AI an open-book test. Connect LLMs to external knowledge bases for accurate answers.Read17Advanced RAG TechniquesGo beyond basic vector search with Reranking, Hybrid Search, and Query Expansion for production-grade accuracy.Read18The Future of Agentic AIResearch Deep Dive: Why Small Language Models (SLMs) are replacing monolithic LLMs.Read
Latest Research
13 entries01Three Agent Papers, April 2026New Research: Hyperagents (Meta FAIR), Recursive Language Models (MIT), and GMPO (Microsoft / ICLR). Three architectural moves from a single month.In queue02AI OverthinkingNew Research: When models think too much, they often talk themselves out of the correct answer.Read03Latent Reasoning (Coconut)New Research: What if LLMs didn't have to 'think' in words? Explore reasoning directly in continuous latent space.Read04Qwen3 (Unified Thinking)New Research: A single model that can dynamically switch between fast responses and deep reasoning modes.Read05DeepSeekMath (GRPO)New Research: How a 7B model approached GPT-4 math performance by ditching the RL 'Critic' model.Read06Kimi K2 ThinkingNew Research: An open-source thinking agent that interleaves reasoning with tool use (300+ steps).Read07DeepSeek-OCRNew Research: Compressing long documents into highly efficient 2D visual tokens instead of text.Read08CoT MonitoringNew Research: Can AI models learn to hide their dangerous thoughts from safety monitors?Read09Transformer SensitivityNew Research: Why are Transformers so robust? They naturally learn 'low sensitivity' functions.Read10Coherence (Segmentation)New Research: An unsupervised method that uses 'sticky' keywords to find topic boundaries.Read11SFT vs. RL GeneralizationNew Research: Does Supervised Fine-Tuning just memorize while RL actually learns rules?Read12Natural Language AutoencodersNew Research: Anthropic's 2026 method for translating Claude's internal activations directly into human-readable English.Read13RLVR & Process RewardsNew Research: The recipe behind 2026 reasoning models — swap the gameable reward model for a verifier, watch reward hacking vanish, and run it cheaply with GRPO.In queue