Learning Resources and Practice Paths

Choose a resource to close a specific skill gap, then demonstrate the skill in a small project and a design explanation. Completing more courses is not the same as becoming ready for an interview.

This is Learnastra's selection of external courses and primary documentation, reviewed September 24, 2026. The practice assignments below are Learnastra exercises, not claims about an external course's assessments or an employer's interview. Providers control their enrollment, prices, certificates and hosted services; access to this guide does not include those purchases.

Choose your next step

flowchart LR
  G[Identify a skill gap] --> R[Read one focused resource]
  R --> B[Build a bounded example]
  B --> T[Test normal and failure cases]
  T --> E[Explain the decision aloud]
  E --> N[Record the next gap]
  N --> G

Before enrolling, check prerequisites, the syllabus, software versions, current access terms and any separate GPU or API charges. Prefer a reproducible exercise over a long playlist you cannot apply.

Foundations: models and Transformers

Practice: explain the difference between a token, an embedding and an attention weight. Run the tiny decoder lesson, change one assumption and report what happens. For detailed explanations, use the tokenization, attention and Transformer chapters.

Retrieval and RAG

Practice: build lexical search over a small collection you are permitted to use. Add generation with citations, then compare vector or hybrid retrieval on the same questions. Include an absent answer, a deleted document, a changed permission and an exact identifier. Explain whether retrieval recall or synthesis failed before changing the design. Continue with RAG fundamentals and the enterprise RAG interview.

Agents and orchestration

Practice: create a support workflow with a read-only lookup tool and a proposed-action step. Add runtime-enforced tool scope, a total deadline and a maximum call count. Simulate a failed tool and a lost response. Explain why the workflow does or does not need several agents. Compare the agent fundamentals, orchestration and durable execution lessons.

Context and memory

Practice: retain a user-approved preference, explain its provenance and expiry, then delete it. Demonstrate that the next answer no longer uses it, including through a cache or summary. Read memory architectures, long-term memory and context engineering.

Evaluation and observability

Practice: label a small, permission-safe set of outputs with an explicit rubric. Compare a proposed model judge with human labels. Report disagreements and unjudged cases, then run a paired baseline/candidate experiment on held-out cases. A small exercise teaches the method; it does not establish a production error bound. Use evaluation fundamentals, observability and the evaluation-gated delivery interview.

Prompting and structured generation

Practice: compare zero-shot instructions, representative demonstrations and a constrained schema on a fixed task. Keep a separate test set. Include malformed inputs and valid-looking but semantically wrong outputs. Explain how business validation differs from parsing. Continue with prompt fundamentals, structured generation and DSPy.

Fine-tuning and adaptation

Practice: choose a stable, narrow task and compare a prompting baseline with an adapter. Keep training, development and test data separate. Report quality, memory, training cost, serving cost and a case where adaptation hurts. Use LoRA and QLoRA, preference optimization and verifiable rewards for distinct training objectives.

Inference, serving and MLOps

Practice: measure first-token and completion latency under increasing concurrency. Explain queueing, token quotas and memory use, then compare a change against the same workload. Include hardware and operations in the economics. Read serving infrastructure, CI/CD and FinOps.

Security, safety and governance

Practice: draw the trust boundaries around user input, documents, tools, memory and model providers. Test an indirect prompt injection in a local, synthetic exercise. Show which server-side restriction prevents the forbidden operation even if the model follows the malicious instruction. Read access control, agent sandboxing and governance.

Coding agents and developer tools

Practice: ask an agent to make a bounded change in a disposable project. Review every changed file, run meaningful tests, inspect permissions and explain a missed requirement. Use coding-agent tools and the autonomous coding interview. In a real hiring process, use AI tools only when that employer's instructions permit them.

Product and leadership preparation

Practice: write a one-page proposal with the user problem, a non-AI baseline, measurable success, excluded uses, operating costs and a limited launch plan. Explain who labels quality, who handles exceptions and what would stop the rollout. Practice the behavioral examples using your actual experience, never invented employment or customer results.

Lectures and public learning material

The Karpathy, fast.ai, Hugging Face and university links above provide focused starting points for public learning material. For video, prefer the lecture links on the instructor's or university's page so the course year and accompanying notes remain clear. General news channels can help discover a topic; verify technical claims against a paper, specification or current project documentation before putting them into a design.

Public reading access does not imply free GPU use, hosted inference, graded assessments or certificates. A promotional access period can end.

Five practice paths

Use a pace that matches your baseline knowledge. Finishing in a fixed number of weeks is not evidence of job readiness. Each path should end with a requirements-led interview exercise, not only a completed playlist.

Check your learning

1. Can you define the concept in one or two sentences without naming a vendor?
2. Can you draw the baseline and explain the path of one request?
3. Can you identify a failure the tutorial did not demonstrate?
4. Can you state which numbers are measured and which are assumptions?
5. Can you explain the full cost, including human and operational work?
6. Can someone reproduce your result using the recorded versions and permitted data?
7. Can you justify keeping the baseline if the more complex option does not help?

Final notes: staying current

1. Record the publication/course year and the software versions used in an exercise.
2. Verify changed SDK syntax in primary documentation before changing the conceptual explanation.
3. Keep your own error log and revisit the resources that address those errors.
4. Use the research reading guide to assess new claims and the framework maintenance guide to plan upgrades.
5. Keep certificates, subscriptions and tool purchases separate from evidence that you can solve and explain the problem.