Efficiency & infraAI platforms & solutionsMLOps & LLMOpsBasic2026-10-12
LLM infrastructure basics: GPU memory, serving and cost at a glance
The cost of running an LLM depends mostly on how GPU memory is used. This article starts with weights, the KV cache and precision. It then covers inference engines, continuous batching, model parallelism, choosing a deployment approach and calculating token costs, in the order you would deal with them in practice.
Language modelsAI platforms & solutionsBeginner2026-10-12
How to choose and read LLMs on Hugging Face: from the model page to GPU memory
This note explains what to look at on a model page on the Hugging Face Hub. It covers how to read the model card, files, name tags, license and benchmarks, and also includes a way to estimate whether a model will run within your GPU memory and a checklist to use when choosing.
Classic ML, recsys & searchLanguage modelsFinance & insuranceBasic2026-10-10
Financial AI: From Fraud Detection and Credit Scoring to Financial LLMs
Gradient boosting models that handle tabular data are still at the center of financial AI, and alongside them, finance-specific LLMs that handle text such as disclosures and regulatory documents are growing. Because credit scoring is classified as high-risk AI in the EU, explainability, fairness, model risk management and final human responsibility must be addressed together.
Computer visionClassic ML, recsys & searchManufacturingBasic2026-10-10
AI in manufacturing: visual defect inspection, predictive maintenance, digital twins
Images of defects are rare on the factory floor, so anomaly detection, which learns only from normal products and then flags anything different, has become the standard for visual inspection. Predictive maintenance, which forecasts failures from sensor time series, and digital twins face the same problem: lab performance does not carry over directly to the factory.
Safety & alignmentRobotics & physical AIDefense & securityBasic2026-10-10
Defense AI: Application Areas and Principles of Human Control
In defense, AI is used for intelligence analysis, maintenance and logistics, command decision support, and unmanned systems, but because outcomes bear directly on human lives, human judgment and responsibility sit at the center of the design. This note summarizes the principles of NATO and the UK Ministry of Defence, the international debate on autonomous weapons, and how to review a system before adoption.
Data engineeringLanguage modelsAgentsBasic2026-10-09
Text-to-SQL and data analysis agents: asking databases in natural language
Text-to-SQL is a technology that turns natural-language questions into SQL to pull answers from databases, and it advanced quickly through benchmarks such as Spider and BIRD and LLM-based agents. However, huge enterprise schemas, ambiguous questions and reliability problems remain unsolved, so research is shifting toward data agents that ask back, hold off and take action.
AI securityAgentsBasic2026-10-09
LLM and agent security guide: from prompt injection to tool permissions
The biggest threat to LLM apps and agents is that models also follow instructions hidden in data. This article covers, in order, the threat map, prompt injection and the 'lethal trifecta', MCP and tool security, layered defenses, operational governance, and the order of adoption.
Data engineeringAI platforms & solutionsBasic2026-10-09
A guide to real-time data pipelines: streaming, CDC and event-driven architecture
A real-time data pipeline accumulates changes in an event log (Kafka). A stream processing engine (Flink, Structured Streaming, Dataflow) processes them continuously, and CDC moves changes from operational databases without delay. Based on official documentation, this article summarizes the role of each component, where they are used in AI services, the order of adoption and a checklist.
AI securityLanguage modelsAdvanced2026-10-09
AI privacy: does training data leak? Membership inference, extraction and differential privacy
AI models do not just summarize their training data; they can also memorize parts of it and output them. This article covers, in the order of key studies, membership inference, which determines whether something was used in training; data extraction, which pulls memorized text out verbatim; and differential privacy, which mathematically bounds leakage.
AI securitySafety & alignmentAdvanced2026-10-09
Adversarial attacks and jailbreaks: from fooling image classifiers to breaking LLM safeguards
Adversarial examples, which fool image classifiers with changes almost invisible to the human eye, and jailbreaks, which make safety-trained LLMs answer requests they were taught to refuse, follow the same framework: leave the model as it is and optimize the input. This note traces the field from the 2013 discovery to 2026 research and summarizes why attacks work and how far defenses can guarantee anything.
Data & evaluationLanguage modelsAdvanced2026-10-09
Training on synthetic data: from Self-Instruct to the model collapse debate
Synthetic data training, in which LLMs are trained on data generated by LLMs, became a major way to reduce human labeling through Self-Instruct, Evol-Instruct and phi-1. Later came warnings of model collapse, the idea that models break down when fed only model-generated data, followed by counterarguments that collapse can be avoided by accumulating data rather than discarding real data. Research attention is now shifting toward which synthetic data to select.
Classic ML, recsys & searchData & evaluationBasic2026-10-09
Machine learning on tabular data: the strengths of tree models and the challenge from TabPFN and LLMs
For tabular data like Excel spreadsheets, unlike images and text, tree models such as gradient boosting have long been the default choice over deep learning. Recently, TabPFN-style tabular foundation models, which read an entire table and predict immediately without training, and LLMs are challenging that position.
Data engineeringSafety & alignmentBasic2026-10-08
A practical guide to data pipelines and governance for AI
For AI models and RAG to produce useful answers, you need both a pipeline that collects, cleans, stores and indexes data and governance that manages quality, access and provenance. This article covers ETL and ELT, warehouses, lakes and lakehouses, data mesh, dataset documentation and RAG document pipelines in the order you would apply them in practice.
AgentsBasic2026-10-08
Agent harnesses: a design guide for turning a model into a working system
A harness is everything in an agent except the model: the system prompt, tools, execution environment, context management, hooks, subagents and verification mechanisms. Drawing on official engineering posts and the SWE-agent research, this article covers the components, design principles, steps for applying them to coding agents, harnesses for long-running tasks, and a checklist.
Data engineeringLanguage modelsData & evaluationBasic2026-10-08
Data-centric AI: improving models through cleaning, deduplication and synthetic data
Keeping the model architecture unchanged and simply filtering the training data, removing duplicates and, when needed, creating new data can substantially change model performance and training efficiency. This note summarizes what research on deduplication, quality filtering, synthetic data and data selection has each shown, and where the field is heading now.
MLOps & LLMOpsAI platforms & solutionsBasic2026-10-08
MLOps and LLMOps operations guide: keeping models running well after you build them
MLOps is the practice of automating data, experiments, pipelines, deployment and monitoring so machine learning models can be put into production and managed over time. LLMOps applies the same principles and adds prompts, evaluation, observability and token cost management.
AgentsBasic2026-10-08
Building agent skills: from folder structure to testing and sharing
A skill is a folder containing instructions, scripts and reference materials. The agent reads this folder in stages, only as much as it needs. Based on the official documentation and specification, this note covers how SKILL.md is organized, how to write the name and description, criteria for deciding length and when to use scripts, testing and improvement, security, and how skills differ from MCP and subagents.
Language modelsData & evaluationAdvanced2026-10-07
GraphRAG, knowledge graphs and ontologies: turning a pile of documents into a relationship map you can query
GraphRAG is a RAG approach that reorganizes documents into a knowledge graph of entities and relationships, writes summaries in advance for each tightly connected group, and uses them to answer 'questions about the whole'. This article starts from the limits of vector RAG and covers how GraphRAG works, its relationship to knowledge graphs and ontologies, recent research, and security and evaluation issues.
Language modelsAgentsBasic2026-10-07
Retrieval-augmented generation (RAG): giving a model an 'open book'
RAG has a language model find and read relevant content from external documents before it produces an answer, reducing hallucinations and the limits of knowledge locked in its parameters. Starting from REALM and RAG in 2020, it has expanded to retrieving only when needed, agent-based architectures, and research on robustness to contaminated knowledge.
Language modelsEfficiency & infraData & evaluationAdvanced2026-10-07
Decision models: AI that answers from a set of options instead of writing text
A decision model is a model that assigns a probability to each predefined option and answers in a single pass, making it much cheaper and faster than a generative LLM. However, evaluation studies published in 2026 also show clear limitations in accuracy, the reliability of its confidence, and error accumulation over long tasks.
AgentsBasic2026-10-07
Agents and tool use: how models went from talking to working
LLMs write well, but they cannot do calculations, check current information or operate external systems on their own. That led to the 'agent' structure, in which a model thinks, calls a tool, looks at the result and thinks again. Recent research asks less whether models can use tools and more whether they can use them reliably.