⊘Scaling Laws: The Empirical Equations of Model Performance
⊘Compute-Optimal Training: Getting the Most From Every FLOP
⊘Emergence: When More Becomes Qualitatively Different
⊘The Mirage Question: Is Emergence Real or a Measurement Artifact?
⊘The Base Model Problem: Fluent but Unruly
⊘Supervised Fine-Tuning: Teaching by Demonstration
⊘RLHF: Learning What Humans Actually Prefer
⊘DPO: Cutting Out the Reward Model
⊘Project Compass Milestone: Fine-Tuning Compass on Instruction Data
⊘From Logits to Words: The Final Transformation in the Generation Pipeline
⊘Greedy Decoding and Beam Search: Determinism and Its Discontents
⊘Temperature: Turning Up the Creativity Dial
⊘Top-k and Nucleus Sampling: Principled Randomness
⊘Project Compass Milestone: Generating Text From Our Model
⊘Why LLMs Lie With Confidence: The Architectural Roots of Hallucination
⊘The Catalogue of Failures: A Taxonomy of What Goes Wrong
⊘The Context Window Wall: When Memory Runs Out
⊘Evaluating Honestly: Benchmarks, Red Teams, and Reality Checks
⊘Mixture of Experts: Scaling Without Proportional Compute
⊘Mechanistic Interpretability: Finally Reading the Map's Own Legend
⊘Beyond Text: Multimodal Models and the Expanding Territory
⊘Agents, Tools, and the Edge of What LLMs Can Become
⊘Project Compass Complete: Your Model, Your Map, Your Next Expedition
Why it's worth it
The credential that closes the gap
These frameworks map to high-demand strategy roles. Figures reflect typical market ranges for target roles, not a guarantee.
$75K–$120K
target role range
~$18K
median uplift potential
5
roles it maps to
Data Scientist $75K–$120KML Engineer $75K–$120KAI Research Engineer $75K–$120KApplied Scientist $75K–$120KMLOps Engineer $75K–$120K
Before you start
What most people get wrong
A few of the misconceptions this course clears up. The full set is inside.
“LLMs understand language the way humans do — they 'read' text and comprehend its meaning.”
RealityLLMs are probabilistic maps of language patterns, not comprehension engines. As Dr. Bakari puts it at Meridian Labs: 'The map is not the territory.' A model learns statistical relationships between tokens across billions of examples — it predicts what comes next based on learned distributions, not because it grasps meaning the way a person does. There is no semantic understanding happening in any cognitive sense; there is extraordinarily sophisticated pattern completion over a compressed representation of human text.
“Tokens are just words. The model processes text word by word.”
RealityTokens are subword units determined by frequency in a training corpus — not words, not characters, not syllables, but statistically motivated fragments. The word 'unbelievable' might become three tokens. The word 'a' is one. An emoji might be four. As Theo discovered when he first ran a tokenizer at Meridian Labs and watched 'Portland' split differently than 'portland': the vocabulary is built by the GLYPH Method — granulated by frequency, not by linguistic intuition. Common words stay whole; rare words get fractured. This means the model's 'view' of text is fundamentally different from a human reader's.
“Word embeddings are just lookup tables — a fixed dictionary where each word has one meaning.”
RealityEmbeddings are continuous vectors in a high-dimensional geometric space where relationships between meanings are encoded as directions and distances. 'Bank' near 'river' and 'bank' near 'finance' occupy different neighborhoods in that space — and crucially, the embedding a token receives is not static. Through the COORD Stack, embeddings are oriented in sequence, refined through training, and ultimately transformed by every attention layer they pass through. By the time a token's representation reaches the final layer, it has been reshaped by its entire context. Maren's standard challenge — 'Show me the dimensions' — is exactly the right question: the geometry is the meaning.
Frameworks you'll keep
Portable thinking tools
Named frameworks you'll carry into every AI decision long after the course.
LLMs learn statistical co-occurrence patterns across billions of tokens—they have no grounded semantic understanding or phenomenal experience of meaning. The TERRAIN Protocol frames them as probabilistic maps of language, not the territory itself. This distinction matters for understanding both their capabilities and their failure modes.
Text must first be tokenized into subword units (via BPE or similar), mapped to a fixed vocabulary, and converted to numeric token IDs. Raw characters never enter the network—only integer IDs do. The GLYPH Method shows how this granulation is foundational to everything that follows.
No. Attention is a learned, selective weighting mechanism—tokens gather only what is geometrically relevant in embedding space. Empirically, models exhibit 'lost-in-the-middle' phenomena where information in the middle of long contexts receives systematically lower attention weight, regardless of importance.
Each attention head learns to attend to a different relational subspace—syntactic dependencies, coreference, positional proximity, semantic similarity—simultaneously. The PRISM Check shows that heads specialize and recombine, making redundancy the opposite of the design intent.
Hallucinations emerge from the fundamental architecture: LLMs generate tokens based on learned probability distributions, not grounded knowledge. The FAULT Analysis framework explains that hallucinations are a feature of the prediction mechanism, not a bug—they can be mitigated through fine-tuning and decoding strategies, but not eliminated.
The original papers are written for researchers reporting results, not engineers building mental models. They skip pedagogical scaffolding and assume prior context. This course is built in reverse: every mechanism explained from first principles, every framework designed to be memorable and transferable, with explicit flags where common analogies break down.
You should be comfortable with basic ML concepts (loss functions, gradient descent, neural network layers) and have hands-on experience with LLMs or NLP. You don't need to have implemented a transformer from scratch. If you've used the OpenAI API and understand what a neural network is, you have enough foundation.
Not if you want to understand why prompts fail in certain contexts, why RAG degrades at scale, or how to reason about model behavior in production. The RENDER Method chapter alone changes how you think about prompt engineering, and the FAULT Analysis makes you a better debugger of LLM failures.
Each of the 14 chapters is self-contained and takes 45–90 minutes. Most working engineers complete the full course in 4–6 weeks at 2–3 chapters per week. Everything is on-demand with no live schedule, so you can pause and return without losing the thread.
Free content optimizes for views; this course optimizes for mechanistic fluency—the kind that holds up in senior interviews, production debugging, and architecture reviews. The salary delta between 'I use LLMs' and 'I understand LLMs' at mid-to-senior ML roles is $30K–$50K. The course pays for itself in weeks.
No. You need strong fundamentals in linear algebra and calculus, but the course teaches the specific math you need in context. If you've completed a standard ML course, you're ready. The course focuses on mechanistic understanding, not mathematical rigor for its own sake.
Papers assume you already understand the intuition. This course builds it from scratch. We explain not just what each component does, but why it's shaped that way, what it actually learns, and where common interpretations break down. You'll read papers differently after this.
Yes. Every skill covered maps directly to job descriptions at Anthropic, Hugging Face, Google, and Cohere. The course covers transformer architecture, attention mechanisms, tokenization, RLHF, scaling laws, and failure mode analysis — all common interview topics. You'll be able to explain, not just name, every concept.
Most engineers complete it in 4-6 weeks at 5-7 hours per week. The chapters are designed to be worked through sequentially since each builds on the previous one. You can move faster or slower depending on your background and how deep you want to go into the frameworks.
No, but we provide optional implementation notebooks for every chapter. The course focuses on mechanistic understanding first. If you want to implement transformers from scratch, the notebooks are there. If you just want to understand how they work, the visual frameworks and explanations are sufficient.
The course is structured so you can skip chapters you're confident in, but we recommend going through sequentially. Most engineers find that even familiar topics have mechanistic gaps. The frameworks and failure mode analysis are usually new even for experienced practitioners.
The core mechanistic principles — tokenization, embeddings, attention, transformer blocks, pre-training — are foundational and unlikely to change. We update the course quarterly to reflect new research on scaling laws, RLHF variants, and failure modes. You get lifetime access to updates.
Yes. We offer a 30-day money-back guarantee. If after completing the first three chapters you don't feel like you're gaining mechanistic understanding, we'll refund you fully. No questions asked.