One Roadmap.
Two Tracks. 52 Weeks.

Start here. Tokens, training, context windows, hallucination, emergent behavior — in plain language from the person who built Tesla Autopilot's AI. Watch this before anything else in the program.

Visual, intuitive explanation of transformers and attention. Watch back-to-back with Karpathy.

Best non-technical AI vocabulary course. Covers AI, ML, deep learning, and agents without jargon. Complements Karpathy for the consulting track. Do modules 1–3 this week.

Read by end of Week 1. The clearest explanation of what an agent is, how it differs from a chatbot, and the key architectural patterns. Reframes how you think about Hermes and every agent tool.

The definitive non-technical AI strategy course. Covers AI project lifecycle, organizational adoption, when AI should augment vs. automate vs. inform. Do Weeks 1–2 of the course content this week.

Gold-standard explanation of the full agent stack — planning, memory, tools. Read sections 1–3. This is what Hermes is actually doing.

Business leader view of planning, strategy, scaling, and responsible AI. Quick but useful for consulting framing.

Capabilities, limits, use cases, and societal/business impact of GenAI. Complements "AI For Everyone" with a GenAI-specific lens. Do modules 1–2 this week.

Clearest non-technical explanation of APIs. After this, the concept that "Hermes reads your Gmail via API" will be intuitively obvious. Read once, done.

How GPT models are trained, fine-tuned, and aligned (RLHF). Gives you the model lifecycle from pretraining to production. Conceptually dense but no prerequisites needed.

Practical AI productivity workflows and responsible use. Adds Google's perspective alongside Anthropic and OpenAI framing.

Enterprise AI adoption, agents, workflow redesign, and value capture patterns. Read the executive summary and the "Where value is being captured" sections closely.

Annual facts on AI capability, economics, policy, and responsible AI. Read chapters 1 and 4 (Technical Performance, Economy). Return to other chapters throughout the program.

Real-world examples of focused GenAI implementation with measurable outcomes. Grounds the theory in concrete cases.

9 chapters with hands-on exercises in real time. Covers XML tags, role assignment, chain-of-thought, few-shot, output formatting, avoiding hallucinations. Built by the Claude team. Better than any paid course at this level. Start here.

The best 90 minutes you can spend on prompt engineering. Covers the two core principles with live demos. Principles apply directly to Claude.

Complete the "Prompt Engineering" and "Agentic AI" courses this week and next. All free, certificates included. Underutilized — third-party coverage is sparse so practitioners who complete this have a real advantage.

Bookmark and return to throughout the program. The evaluation section — how to measure prompt quality systematically — is what separates practitioners from casual users.

MIT's legendary course on shell basics, shell tools, and Git. Lectures 1–3 are essential. Run every command as you watch. This is the first real technical skill in the program.

Visual, browser-based Git tutorial. You see commits and branches as actual diagrams. Complete the full Introduction Sequence. Best Git resource available at any price.

After this video you can read any API response or config file. Zero prerequisites. Do it early — it unlocks everything else.

Hands-on: repos, commits, branches, pull requests. Run after Learn Git Branching to apply concepts in real GitHub. Create your learning portfolio repository this week.

Structured vs. unstructured data, transactional vs. analytical workloads, storage types. Quick but covers essential vocabulary.

Data roles, services, relational and non-relational data, analytics workloads. Provides the mental model for every data platform conversation with a client.

People/process/technology/data maturity model for AI adoption. The data readiness section is immediately applicable to client diagnostics.

Best browser-based SQL tutorial. Complete lessons 1–12 (through joins). Write every query yourself — don't just read. By the end you'll have the 80% of SQL you'll use 95% of the time.

Deeper treatment of relational modeling, normalization, indexes. Do Weeks 1–2 of the course for conceptual depth beyond SQLBolt. Return for Week 3 (design) in Week 9.

Common vocabulary for enterprise data management: quality, governance, master data, metadata, lineage. Read the overview — you don't need the full book, just the framework.

Practical lakehouse governance patterns for structured/unstructured data and models. Grounded in what actually gets implemented vs. what stays in theory.

Modern data architecture, governance, and trusted data quality. Gives you the architectural vocabulary for every data platform client conversation.

Most practical Python resource for non-developers. You write working scripts that automate real tasks within the first few chapters. Ch 1–3 this week (intro, flow control, functions).

Python taught specifically through AI use cases. Do modules 1–2 this week as a complement to ABS. Directly bridges Python to the Claude API.

Lists, dictionaries, strings, pattern matching, file I/O, PDFs. Chapter 9 (organizing files) is directly useful for automating Hermes file workflows. Run every example.

Data manipulation, grouping, sorting, missing data — in browser. By the end you can filter, group, join, and summarize data. The bridge from Excel to Python analysis.

What APIs are, why they matter, REST basics, hands-on usage. After this you can call any public API and parse the response into a table. Essential prerequisite for Claude API work.

Examples for API use, embeddings, evals, agents, structured outputs. Use selectively — the "How to call the API" and "Embeddings" examples are most relevant this week.

Cloud models, service types, shared responsibility, scalability. Quick but builds the vocabulary for every cloud architecture discussion. Do Week 14.

Mental model for generating business value from AI/ML/GenAI transformation. Covers people, process, platform, governance. Consulting-track anchor for cloud strategy advice.

Read the Basic Syntax page once. Markdown is used in Hermes SKILL.md files, GitHub, Obsidian, and most AI tools. Learn it once, use it everywhere.

Best free statistics resource. Descriptive stats (Wk 17), probability and sampling (Wk 18), regression and correlation (Wk 19). Do each topic section rather than the full course.

Plain-language explanations of regression, trees, metrics, PCA, neural nets. The single best free statistics-to-ML bridge available. Watch 3–4 videos per week in Wks 17–24.

45 minutes that teaches you to decide if ML is appropriate, frame success metrics, and translate business problems into prediction targets. Do this in Week 21. Required before every ML use-case conversation.

Animated videos, visualizations, hands-on exercises. Weeks 22–23 cover supervised ML, evaluation metrics, and overfitting. Week 24 covers unsupervised learning and forecasting. The best free ML foundation course.

43 rules for ML engineering from Google's production experience. Covers metrics, pipelines, monitoring. Highly practical — every rule is a lesson from a real failure. Read in Week 23 alongside model evaluation.

Forecasting concepts and hands-on time-series exercises. Do in Week 24 alongside unsupervised learning. Forecasting use cases are extremely common in consulting engagements.

Complete lessons 9–14 this phase (LLM fundamentals, embeddings, RAG, agents, fine-tuning). Earlier lessons were covered in Phase 1. Use as structured overview before going deeper.

Transformers, tokenizers, datasets, fine-tuning, open-source AI ecosystem. Do the "Transformer Models" and "Fine-Tuning" modules in Week 26–27. Essential for understanding open-weight models (Llama, Qwen, Mistral) running on Ollama.

Business-friendly RAG definition, grounding, and external knowledge bases. Read in Week 28 as the conceptual foundation before building a RAG system.

Most credentialed RAG course available. Retriever architecture, semantic search, vector databases, chunking, evaluation. Do Modules 1–3 in Week 29, 4–5 in Week 30.

Extends RAG to iterative, agentic retrieval. The architecture that makes your newsletter research agent improve over time. Take after the main RAG course.

Hands-on tutorial for Q&A over unstructured documents. Build your first working RAG pipeline. Do in Week 30.

Agent concepts, tool use, orchestration. Do modules 1–3 in Week 31. The most comprehensive free agents course, building on the HF ecosystem you already know from Phase 5.

Official MCP course from the protocol's creators. Client-server architecture, how to connect AI applications to external tools, building and deploying an MCP server. Do in Week 32.

Hands-on MCP from beginner to implementation, with community challenges and a free certificate. Take after the DeepLearning.AI version for reinforcement.

Human-centered AI product design, user needs, mental models, feedback loops. The UX layer that determines whether a technically correct AI product actually gets adopted.

Four agentic design patterns: reflection, tool use, planning, multi-agent collaboration. The mental model for writing Hermes SKILL.md files and system prompts that actually work.

Programmatic evaluations for model outputs. The foundation of systematic quality testing before production launch.

Systematic evaluation for RAG: retrieval relevance, answer faithfulness, citation accuracy, refusal behavior. Separates retrieval failure from model failure.

Experiment tracking, deployment, monitoring, CI/CD for ML. Do Weeks 1–3 of the course (tracking, orchestration, deployment). Full 9-week course is for engineers.

Common LLM app risks: prompt injection, data leakage, insecure outputs. Read all 10 items. Then red-team your personal RAG assistant against items 1–3. Required before any production AI deployment advice.

Critical security risks for autonomous/tool-using AI agents. Essential for advising on Hermes-style deployments in enterprise contexts.

Core trustworthy AI risk framework: Govern, Map, Measure, Manage. The practical anchor for cross-sector AI risk management. Read the full Core and the Govern function in detail.

GenAI-specific risks and actions aligned to NIST AI RMF. The 12 unique GenAI risks (confabulation, data privacy, IP, CBRN, etc.) and how to control them.

Operationalizes fairness, reliability, privacy/security, inclusiveness, transparency, accountability. The most practical operationalization of responsible AI principles available.

Current EU AI Act application milestones and enforcement dates. Read alongside OECD Due Diligence Guidance to understand what triggers legal review.

International trustworthy AI principles and practical enterprise guidance for applying them. The policy baseline you need for multi-jurisdiction advisory work.

Weeks 0–4 of CS50x (computation, algorithms, memory, data structures, web) give you the conceptual vocabulary for every technical conversation. The most efficient path to genuine technical literacy as a non-developer founder.

2–7 minute videos on web concepts: frontend vs. backend, APIs, databases, deployment, Docker. Builds complete mental model of software architecture faster than any course. Watch 2 per day across Week 48.

Karpathy's thesis that neural networks are a fundamentally different programming paradigm. Understanding this reframes what a software company means in 2026 and what moat is actually buildable.

Business and product layer of AI: when to use models vs. fine-tune, moats in AI products, the difference between demos and products, how AI changes software economics. Required reading for any AI founder.

Week 50: return to these reports, this time through the lens of your chosen domain (your startup idea, or a consulting industry vertical). Extract all domain-specific data, use cases, and adoption patterns.

Best business and strategic analysis of tech and AI. Thompson writes at the intersection of business model, product strategy, and technology — exactly the thinking for Week 50–52 synthesis. Free posts twice a week.

Week 52: watch the video that started the program. If you catch things you missed — understand concepts that confused you before — you've made real progress. You will. This is the best 1-hour calibration test available.