learn ai in the right order.
Start with the original sources. Build small things as you go. Skip the noise until you need it.
why learn this
Everything else is easier when you understand what the model is doing, what it is not doing, and where the core abstractions came from.
before this
- - basic programming
- - high-school algebra is enough for the first time through
start with
you should be able to
- explain what a neural network learns
- describe attention without hiding behind the word attention
- explain tokens, embeddings, next-token prediction, and context windows
- know why transformers replaced earlier sequence models
save for later
- - training infrastructure details
- - frontier safety reports
- - fine-tuning papers
practice
Build a tiny autocomplete model
beginner · python, pytorch or tinygrad, a small text corpus
make: a notebook or script that trains a tiny character-level model and samples text
you know it works when: show training loss, sample output, and a short note on what improved after tuning
why learn this
Prompting is the first control surface. Good prompts make later systems easier to evaluate, debug, and operate.
before this
- - foundations
- - basic comfort using chat models
start with
you should be able to
- write task prompts with clear inputs, outputs, and constraints
- use few-shot examples when behavior is hard to describe
- separate instruction quality from model capability
- know when prompting is no longer enough
save for later
- - model-specific tricks that do not transfer
- - long prompt libraries without evals
practice
Build a prompt pack for one workflow
beginner · claude, chatgpt, or any model playground
make: five reusable prompts for a real workflow, each with inputs, expected output, and examples
you know it works when: include before/after outputs and explain which prompt changes improved reliability
why learn this
Retrieval and generation fail together in real products. Treating them as one layer makes it easier to debug relevance, citations, and answer quality.
before this
- - foundations
- - prompting
- - basic arrays and vectors
start with
you should be able to
- explain semantic search and cosine similarity
- choose a chunking strategy for a document type
- inspect retrieved context before blaming the model
- separate retrieval quality from answer quality
- create a small eval set for grounded answers
save for later
- - custom embedding training
- - graph RAG
- - vendor-specific frameworks
practice
Build a cited knowledge assistant
intermediate · next.js or python, embeddings API, sqlite or a vector store, model API
make: a local app that searches a document folder and answers with citations
you know it works when: include retrieval logs, citations, and ten grounded-answer evals with pass/fail notes
why learn this
Context is the working memory of an AI system. Good context design is often the difference between a demo and a useful product.
before this
- - prompting
- - retrieval & rag
you should be able to
- decide what should be in prompt, retrieval, tool result, or memory
- identify context that is stale, redundant, or distracting
- identify a prompt-injection boundary and its mitigations
- explain how context rot changes system design
save for later
- - giant-context brute force
- - memory systems without evals
practice
Build an AI research assistant
intermediate · typescript, model API, search or local documents, markdown output
make: an assistant that gathers sources, compresses notes, and writes a sourced brief
you know it works when: show the final brief, source list, and what context was kept or discarded
why learn this
Agents are useful when work requires decisions, tools, feedback, and multiple steps. They also fail in new ways.
before this
- - prompting
- - context engineering
- - basic API work
start with
you should be able to
- choose workflow vs agent intentionally
- design tools with inputs and outputs the model can use
- handle uncertainty, retries, and partial failure
- measure whether the agent is getting more reliable
save for later
- - autonomous everything
- - multi-agent architectures before one agent works
practice
Build a GitHub issue triage agent
intermediate · typescript, github api, model API, structured outputs
make: an agent that reads an issue, labels it, asks clarifying questions, and drafts a fix plan
you know it works when: run it on ten issues and record correct labels, bad labels, and failure reasons
why learn this
Without evals, AI product work becomes vibes. Evals turn errors into a system you can improve.
before this
- - prompting
- - rag or agents
start with
you should be able to
- write examples that represent real product failures
- separate unit evals, human review, and production monitoring
- use error analysis to choose the next change
- know when an eval is being gamed
save for later
- - leaderboards
- - generic benchmarks that do not match your product
practice
Build an eval set for an LLM feature
intermediate · typescript or python, json fixtures, model API, simple report output
make: a repeatable eval harness with examples, expected behavior, and scoring notes
you know it works when: compare two prompts or models and explain the regression you would ship or reject
why learn this
Production AI is mostly product engineering: reliability, cost, latency, observability, feedback, and iteration.
before this
- - retrieval & rag or agents
- - evals
- - basic full-stack development
start with
you should be able to
- identify the failures users actually feel
- track cost, latency, quality, and feedback
- design fallbacks for model and retrieval failures
- know what must be monitored after launch
save for later
- - premature fine-tuning
- - complex orchestration before instrumentation
practice
Ship a small AI product slice
advanced · next.js, model API, database, auth provider, logging
make: a deployed AI feature with auth, usage limits, feedback capture, and basic observability
you know it works when: include a demo link, failure-mode notes, and one iteration driven by feedback or evals
why learn this
Capable AI systems create product, security, policy, and social risks. Builders need enough safety literacy to make better decisions.
before this
- - foundations
- - agents
- - evals
start with
you should be able to
- explain why agentic systems need different controls
- identify misuse, overreliance, privacy, and security risks
- read a system card or safety report critically
- design a basic risk review for an AI product
save for later
- - policy fights without technical grounding
- - abstract doom or hype pieces
practice
Write a launch risk review
advanced · markdown, product spec, eval results, abuse-case checklist
make: a concise risk review for one AI feature before launch
you know it works when: include risks, mitigations, monitoring, and a decision on what blocks launch