Delta: AI Engineering

You read about AI every morning. By Friday, almost none of it is left. The habit is not the problem. The format is. Reading something once and moving on is the least effective way to learn anything, and the industry has built an entire content economy on top of it. Delta is the other half. Short lessons about what actually changed, each one ending in a question you have to answer. Ninety seconds, usable in a queue. WHAT A CARD LOOKS LIKE Two or three paragraphs, then a question. Not a quiz question. Predict the output. Spot the bug in the IAM policy. Pick the cheaper option and say why. Name the tradeoff the benchmark table is hiding. Then the explanation, which is the part that matters, written by someone who has shipped the thing rather than read about it. Every card links to the source it was written from. We never reproduce it. TEN TOPICS Foundations and modeling. Attention variants, scaling laws, fine-tuning, evaluation methodology, reading a benchmark table honestly. Generative AI. Reasoning models, context handling, structured output, sampling. Agentic AI. Tool use, MCP, planning, memory, and the failure modes that only appear in agents: loops, runaway cost, silent partial failure. AI engineering and inference. KV cache, quantization, batching, routing, what a token actually costs you in production. Data and retrieval. RAG and how it fails, chunking, reranking, drift. AI security. Prompt injection direct and indirect, exfiltration through tools, MCP supply chain, sandboxing, least privilege, the OWASP LLM Top 10. Responsible AI. Bias measurement, interpretability that survives an audit, and the honest limits of each technique. AI governance and regulation. EU AI Act by risk tier, NIST AI RMF, ISO 42001, model documentation, audit trails. Always framed as what you have to build, never as policy commentary. Enterprise AI architecture. Landing zones, multi-tenancy, identity for agents acting for users, build versus buy. AI operations and reliability. Monitoring nondeterministic systems, evaluation in production, regression testing prompts, incident response when the model is the incident. The last five are the reason this exists. Newsletters cover model launches. Almost nobody writes for the engineer who has to get an agent through a security review, satisfy an auditor, and keep it inside an SLA. IT REMEMBERS WHAT YOU FORGOT A scheduler runs underneath, tracking what you got right, what you got wrong, and how long you took. A correct answer that took twelve seconds is not the same as one that took two. You never see any of it. There is no review queue, no due count, no deck to clear. Cards you are about to forget simply reappear in the feed, indistinguishable from new ones. BUILT FOR PEOPLE WHO ALREADY KNOW THINGS Delta assumes you know what a transformer is. It will not explain backpropagation. It is for practitioners with a few years of shipping behind them who want to know what changed this month and why it matters, not what a neural network is. NO ADS, EVER The entire premise is that this is the calm alternative to a feed that monetises your attention. There is no advertising identifier anywhere in the app and there never will be. Nothing about you is sold or shared. FREE AND PAID Two cards free in every topic, twenty in total, so you can see the whole range before deciding. The subscription opens everything: every card in all ten topics, and a scheduler that brings back what you got wrong before you forget it. Delta Pro is $4.99 per month or $49.99 per year. Payment is charged to your Apple ID. Subscriptions renew automatically unless auto-renew is turned off at least 24 hours before the period ends, and you can manage or cancel in your Apple ID settings at any time. Terms: https://drvncrypto-org.github.io/delta-study/terms.html Privacy: https://drvncrypto-org.github.io/delta-study/privacy.html