Paper · 01 RAG: give the answer something to stand on Lewis et al. · Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks ↗ Original 2020-05-22 · Reviewed 2026-10-03
Why it matters When your team asks whether AI can use its handbook, start with retrieval—not a confident-sounding answer.
Reading note The paper combines learned model knowledge with retrieved documents. Its reported results belong to specific tasks and comparisons; your handbook needs its own evaluation.
Before this step: Model · Context
Source excerpt “outperforming parametric seq2seq models and task-specific retrieve-and-extract architectures” One English pattern compared with … Name the comparison when describing an improvement. The practice phrase is our teaching example, not a quotation.
Try it at work Write two sentences proposing a handbook assistant. Name the evidence it should retrieve and one failure you would test.
Your explanation in English Save my draft
Check your draft Did you explain the mechanism, name a condition, and make a concrete work decision? Saving records your draft, not a mastery score.
Back to the roadmap → Understand the technology · STEP 3 Video · 02 3Blue1Brown · Grant Sanderson · Transformers, the tech behind LLMs ↗ Original 2024-04-01 · Reviewed 2026-10-03
Why it matters Go beyond repeating “Transformer.” Build a picture you can explain to someone else.
Reading note The visual lesson follows tokens, vectors and attention. Use it to connect three ideas before tackling the mathematics.
Before this step: Model
Source excerpt “An input is first broken into small chunks that are known as tokens.” One English pattern First … then … Sequence words help you explain a mechanism one step at a time.
Try it at work In three English sentences, explain how text becomes tokens and vectors, then how attention uses context.
Your explanation in English Save my draft
Check your draft Did you explain the mechanism, name a condition, and make a concrete work decision? Saving records your draft, not a mastery score.
Back to the roadmap → Understand the technology · STEP 2 Engineering · 03 Your agent says “done.” What would prove it? Anthropic Engineering · Demystifying evals for AI agents ↗ Original 2026-01-09 · Reviewed 2026-10-03
Why it matters Bring a sharper question to your next demo: did the system change the right thing?
Reading note An agent evaluation can inspect its actions and the resulting state. Choose checks that reflect the outcome your user needs.
Before this step: Tool · Workflow
Source excerpt “The outcome is the final state in the environment at the end of the trial.” One English pattern Success means … Define an observable result rather than a vague claim about quality.
Try it at work Choose one task an agent could do at work. Write a success criterion and a check that could reveal a false “done.”
Your explanation in English Save my draft
Check your draft Did you explain the mechanism, name a condition, and make a concrete work decision? Saving records your draft, not a mastery score.
Back to the roadmap → Make product decisions · STEP 2 KEEP THE CONNECTION Let the next useful idea come to you. Add the feed to your RSS reader. Each update includes why it matters and a small task; visit the guide when you want to go deeper.
