AGI Researcher Foundations · Article 11 of 11 — Capstone
The Capstone: Ten Articles, One Map, and What to Actually Do Next
Ten articles ago, this series started with a checklist of tools. It ends here, with everything — the stack, the math, the theory, the open research directions, and the systems underneath — connected into one coherent map, plus the research-methodology habits that turn all of it into actual work, and a concrete next step for wherever you are right now.
Opening
The Unified Map
Cross-Cutting Threads
Research Methodology
Full Series Self-Assessment
What to Actually Do Next
What This Series Didn't Cover
All Ten Articles
Videos
Verdict
From a Checklist to a Map
Article 1 opened with a question: what does someone starting out actually need to get into AI research or land a role at a frontier lab? Ten articles later, the honest answer has a lot more structure than a single checklist could hold — a technical stack, a way of judging your own readiness, the math underneath it, the architecture and training concepts built on that math, the theory explaining why any of it generalizes, the RL machinery behind alignment techniques, the tools for looking inside a trained model, the problem those tools serve, the most contested open research direction, and the distributed systems everything actually runs on. This capstone's job is to show how those ten pieces fit together — not as a list, but as a map with real dependencies between the pieces — and then to hand you a concrete next action, because a map without a next step is just a poster.
10
Articles this capstone connects, from the technical stack to distributed systems
4
Companion deep dives this series cross-links: Evolution, After Transformers, How to Read a Paper, The AGI Debate
1
Repeating pattern (a real wall, a fix from one of three lanes) that runs through this entire body of work
0
Articles in this series that required a PhD, an elite credential, or paid access to complete
Part 1 — The Unified Map
How the Ten Articles Actually Depend on Each Other
This isn't a reading order by publication date — it's a dependency map. Each layer below genuinely needs the ones above it to make sense.
1
Articles 1–2 →
The Technical Stack & Self-Assessment
Programming, frameworks, compute — the entry bar. Everything else assumes you can actually run a training loop.
↓ requires the language of ↓
2
Article 3 →
Mathematical Foundations
Linear algebra, probability, calculus, information theory — the vocabulary every later article assumes.
↓ built into ↓
3
Article 4 →
Core Deep Learning Concepts
Architectures, normalization, training dynamics — how the math becomes a trainable network.
↓ raises the question of why it ↓
4
Article 5 →
Generalization & Learning Theory
Why a trained model works on data it's never seen — and where classical theory demonstrably fails to predict modern deep learning.
↓ splits into two directions ↓
6
Article 7 →
Interpretability
Tools to look inside a trained model — circuits, superposition, causal tracing.
↓ both feed into ↓
7
Article 8 →
Alignment
Making sure a model's goals match intent — the problem RL and interpretability jointly serve.
↓ raised to a harder version by ↓
8
Article 9 →
World Models
The open research direction that would intensify every alignment question above, if it pans out.
↓ all of the above run on ↓
9
Article 10 →
Systems
The distributed infrastructure — parallelism, memory management, fault tolerance — every prior article's training runs actually depend on.
Notice the shape: the map isn't a straight line. Generalization theory forks into RL and interpretability; both feed alignment; alignment gets a harder version from world models; and systems sits underneath all nine other layers, not after them, since every one of those articles' techniques had to actually run on real hardware.
Part 2 — Cross-Cutting Threads
The Patterns That Ran Through Every Article
The Wall/Fix Pattern
From "Evolution" and "Two Threads": every real advance addressed a specific, nameable limitation using math, data, or compute that had just become available. This pattern showed up explicitly in Article 4 (normalization solving vanishing gradients) and Article 6 (PPO solving naive policy-gradient instability).
Outer vs. Inner, Load-Bearing vs. Cosmetic
Article 8's outer/inner alignment split and "The AGI Debate"'s load-bearing-vs-cosmetic framework are the same move: before evaluating a fix, ask precisely what problem it's solving and whether removing it breaks anything.
Honest Uncertainty
Every article in this series flagged what's genuinely unsolved — Article 5's generalization puzzle, Article 8's contested deceptive-alignment risk, Article 9's unproven world-model scaling — rather than presenting open questions as settled.
Part 3 — Research Methodology
The Habits That Turn Knowledge Into Research
Everything in Articles 1–10 is knowledge. Research methodology is the set of habits that turns knowledge into new, verified contributions — and this series has actually already taught most of them, spread across articles that weren't explicitly "about" methodology.
1
Reproduce before you innovate — Article 2's core lesson: rebuilding a known result (nanoGPT, a paper's headline claim) is how textbook knowledge becomes working intuition, and it's the actual filter frontier labs use.
2
Evaluate ideas with a real framework — "After Transformers"'s six-question framework and "How to Read a Paper Like a Researcher"'s worked examples are the actual mechanics of not being fooled by an impressive-looking benchmark table.
3
Track your experiments — Article 2's item 6: without logging both training and validation metrics, Article 5's generalization gap and Article 4's overfitting are invisible until it's too late to matter.
4
Publish your own work — Article 2's item 8, and the single most consistent lesson from this series' Field Guide profiles (Nanda, McDougall): visible, linkable work is what actually gets noticed, not private practice.
5
Hold genuine disagreements without resolving them prematurely — "The AGI Debate"'s whole approach: steelman both sides, name what would settle it, and don't manufacture false certainty where the evidence doesn't support it.
Part 4 — Full Series Self-Assessment
One Question Per Article
1
Have you reproduced a known result from scratch (Article 1/2)?
2
Can you explain why LoRA's "low-rank" claim is a linear-algebra statement, not just a name (Article 3)?
3
Can you explain why LLaMA uses RMSNorm instead of LayerNorm (Article 4)?
4
Can you explain what Zhang et al.'s random-label experiment demonstrated, and why it matters (Article 5)?
5
Can you map RLHF onto the state/action/reward/policy framework unaided (Article 6)?
6
Can you explain what an induction head is and why finding one counts as verification, not visualization (Article 7)?
7
Can you diagnose whether a misalignment case is an outer or inner alignment problem (Article 8)?
8
Can you explain why MuZero not being given the rules of Go is the scientifically important detail (Article 9)?
9
Can you explain why one failed GPU can stall a 992-GPU training run (Article 10)?
If most of these are honest yeses, the foundational layer of this series is genuinely internalized — not memorized, but usable. If several are no, that's not a failure; it's a reading list, pointing back at the specific article that covers it.
Part 5 — What to Actually Do Next
A Concrete Path Forward, Not Just "Keep Learning"
| Where You Are | Concrete Next Step |
|---|---|
| Haven't reproduced anything yet | Start with nanoGPT (Article 2), following its own README exactly — don't skip to something harder first |
| Comfortable with the stack, unsure what to specialize in | Read "The AGI Debate" and this series' Research Frontier Map to see which open bet (Article 9's world models, Article 7's interpretability, Article 8's alignment) genuinely interests you |
| Ready to build a portfolio | Pick one paper from any article's reference list, apply "How to Read a Paper"'s framework to it in writing, and publish the write-up — Article 2's item 8, actually executed |
| Deciding between a research vs. engineering track | Article 1's Scenario C and this series' lane framework (Frontier-scale, Applied/Agentic, Narrow research bet) — pick based on which failure mode (Article 10's infrastructure firefighting vs. Article 5's theoretical uncertainty) you'd rather spend your days debugging |
⚠️ What This Series Didn't Cover
Eleven articles is substantial, not exhaustive. This series did not cover: multimodal architectures in depth, robotics and embodiment, the full mathematics of any single technique (each article deliberately stayed at the conceptual-plus-citation level, pointing to primary sources for full rigor), or a complete treatment of AI policy and governance. It also reflects a snapshot as of September 2026 — several open questions across Articles 5, 8, and 9 are live research areas that will have moved by the time you're reading this, which is exactly why every article pointed to primary, updatable sources rather than asking you to trust this series' summary alone.
🔗 The Complete AGI Researcher Foundations Series
- Article 1 — The Technical Stack
- Article 2 — The Self-Assessment Deep Dive
- Article 3 — The Mathematical Foundations
- Article 4 — Core Deep Learning Concepts
- Article 5 — Generalization & Learning Theory
- Article 6 — RL Foundations
- Article 7 — Interpretability
- Article 8 — Alignment
- Article 9 — World Models
- Article 10 — Systems
- Companion — The Evolution of Neural Networks
- Companion — After Transformers
- Companion — How to Read a Paper Like a Researcher
- Companion — The AGI Debate
- Companion — Two Threads, One Chain
- Companion — How Each Network Architecture Actually Learned
🎥 Recommended Videos
🧭 Closing — The Map Is Finished; the Work Isn't
🎯 The Bottom Line
Ten articles, traced from a training loop's first weight update to the distributed systems running a 992-GPU job, connect into exactly one map: tools and self-assessment enable math, math enables architecture, architecture raises the question of why it generalizes, generalization theory forks into RL and interpretability, both feed alignment, alignment gets sharpened by world models, and systems runs underneath every layer of it. None of the nine other researchers profiled across this series' Field Guide had this map handed to them this cleanly — they built it piece by piece, the same way this capstone's Part 4 self-assessment asks you to check you actually have. The series is complete. The reproduction, the publishing, and the next open question you decide to chase are not — and per every article's own closing argument, that's exactly where the real work, and the real opportunity, still is.