A useful learning plan needs an outcome. You might want to judge a chatbot's answers more carefully, understand how a neural network learns, or make an assistant complete a repeatable task. Each goal calls for a different starting point. The resources below offer several routes; the suggested exercises are our own.
Start with better judgment about AI answers
Anthropic's AI Fluency collection includes tutorials on hallucinations, bias, sycophancy, tokens, and context. It also offers learning material tailored to builders, educators, students, and other audiences. These topics are useful before you depend on a chatbot for research or drafting.
Try this: ask an assistant to summarize a short document you know well. Check each factual statement against the document. Then repeat the task with a clearer audience, length limit, and instruction to identify missing information. Keep notes on which changes actually improve the result.
Use MIT to connect concepts with experiments
MIT 6.S191: Introduction to Deep Learning pairs lectures with slides and software labs. The published material covers sequence modeling, computer vision, generative modeling, and reinforcement learning. Labs include music generation and fine-tuning a language model. The course expects elementary calculus and linear algebra; Python familiarity is helpful.
Try this: follow one lecture and its related lab before starting another topic. Change one setting in the example, record what happens, and explain the result in a few sentences. A notebook containing your observations gives you something concrete to revisit.
Look inside a neural network
Andrej Karpathy separates his educational videos into technical and general-audience tracks. His Neural Networks: Zero to Hero playlist is the technical route, beginning with building micrograd. His official website also links introductory explanations of language models and his practical video, How I use LLMs.
Try this: work through a small implementation with the video paused. Explain what the inputs, predictions, and loss represent before running the next step. When you need help, write down the exact operation you cannot yet explain.
Move to existing models with Hugging Face
The free Hugging Face LLM Course teaches Transformers, Datasets, Tokenizers, and the Hub. It progresses from using pretrained models to fine-tuning, sharing demos, and more advanced work with language models. The introduction recommends good Python knowledge and an introductory deep learning course first.
Try this: choose one small language task, prepare a handful of examples, and run a pretrained model. Inspect the failures before considering fine-tuning. Keep the original examples as a comparison set when you change the model or its configuration.
Give your first agent a narrow job
The Hermes Agent Quickstart walks through installation, provider selection, and an initial conversation. It recommends getting the basic chat working before adding features such as gateways or routing. The documentation is publicly readable; running models or connected services may involve separate costs.
Try this: use a folder of sample notes and ask the agent to produce an index. Define the desired output first, then inspect whether every note appears. For an AI-assisted coding project, apply the same approach: one small feature, a clear completion condition, and a check of the generated result.
Choose your next session
- Everyday AI use: start with AI Fluency and the document-summary exercise.
- Technical understanding: pair MIT lectures with Karpathy's implementation videos, then move into Hugging Face.
- Agent practice: complete the Hermes setup and one limited task before extending the workflow.
Finish each session with an artifact you can inspect: an annotated answer, a working notebook, or a checked output file. Use the questions that remain to choose your next lesson.
Editorial note: This original English guide was inspired by Asya Karpova's AI learning roundup on vc.ru. Resource descriptions were checked against the providers' pages. The exercises and suggested learning sequence are SHAMSI AI's editorial recommendations.