The curriculum
Study
The syllabus, published before it is climbed. Every unit runs consume → do → output, and nothing closes without an artifact: a notebook entry, a lab build, a playable explanation, or a spaced-repetition bottle.
LADDER SET · LAST PASS
On the bench
Written by the study desk, not computed hereCS146S Assignment 1 — put one real Claude Code session behind mitmproxy on a scratch repo, then annotate the system prompt and two tool schemas from the capture. CMU 11-768 runs in the same slot: its lecture fast pass next, Assignment 1 once the OpenRouter key exists.
Set 2026-10-10 when Connor added CS146S and said both it and 11-768 get his immediate attention, all assignments. CS146S A1 is the one build that starts today at zero cost — it runs on the Claude Code he already uses, no key, no sandbox — and its output (what a production harness’s prompt is defending against) is the exact reading the 11-768 harness build needs next. 11-768 A1 still waits on a key that does not exist yet.
Step 1 · Build 12You are here
Build an agent harness that fixes real codeCMU 11-768 Assignment 1: write the ReAct loop, get the CodeAgent to repair the chess-app, add context compaction, and give the ChessAgent its tools. Watch the public lectures first in one fast pass, writing down every gap, then let the gap list drive the build.
Done when — The public test suite passes on your code (`make test`), the CodeAgent fixes at least one chess-app failure in the sandbox, and you can draw the loop and the compaction step from memory.
every public lecture in one fast pass, then about a weekreads AI Agents (11-768)
Step 2 · Build 01
Why does this laptop generate at 20.9 tokens per second?Done when — The predicted tok/s, the measured tok/s and a stated reason for the gap all exist, and a stranger with the same laptop could reproduce all three.
two 60-minute blocksreads Efficiency in LLMs
Step 3 · Build 02
An evaluation you would bet onDone when — The harness scores a real model run, AND you can demonstrate at least one concrete way it can be made to lie.
about a weekreads Build a Reasoning Model From Scratch
Step 4 · Build 03
What a marginal dollar of test-time compute actually buysDone when — The curve exists with your own measurements, the knee is identified, and you can state what changes its position.
about a weekreads Build a Reasoning Model From Scratch
Step 5 · Build 05
A refinement loop, and the point where it starts driftingDone when — The loop runs, the drift is reproducible, and it is playable by someone who did not build it.
about a weekreads Build a Reasoning Model From Scratch
Step 6 · Build 06
Compute a DRAM access by hand, then check it against the specDone when — A hand-computed access time matches a published spec, and you can explain what a row-buffer hit changes.
about a weekreads Computer Architecture
The rest of the map
Builds off the current route, in difficulty order. Reference until the route reaches them.
- 04Predict the VRAM before renting the GPUAfter 01
- 07Attention, implemented, then made fastunlocks 08 · 09Open
- 08FlashAttention, by handAfter 07
- 09What each collective operation actually costsunlocks 10After 07
- 10Train a model across devices, and account for every secondAfter 09
- 11Teach a small model to reason, with rewards you choseAfter 05
- 13Measure whether a harness change actually helpedOpen
- 14What one AI campus pays for its power, and how much of its cost that isOpen
The week
One deep block, one short one, a review — no third track- Mon–Fri60 min, first thingTrack AThe current step on the path. Recall last session from memory, build, then log three lines and stop mid-task.
- Tue · Thu30 min, after the first blockTrack BValuation, Module 0 on Micron. The first block answers F1 (the income statement) — the question that has held the module since August.
- Fri15 minBothBottle three to five facts from the week, in your own words.
- Satonly if needed—Catch up a missed weekday. Otherwise off.
- Sun20 minBothRe-read the notes due this week (listed under Re-read) and re-test the bottles that are due. Nothing new.
Track A — AI infrastructure, in order
- 1AI Agents (11-768) Now — the lectures, then Assignment 1 (build 12 on the path).
- 2Renewables in Electricity Markets Next — modules 1–3 first, the part build 14 needs.
- 3US power market design and datacenter load Then — the capacity auction, the queue and the large-load tariff (R1, R2, R5) for build 14.
- 4How compute pays for itself — the filings Then — the CoreWeave 10-K, which closes build 14.
- 5Build a Reasoning Model From Scratch After build 14 — chapters 3–5 carry builds 02, 03 and 05.
- 6Computer Architecture Then — only the memory lectures, raided by builds 01 and 06.
Track B — investing, in order
- 1Valuation Now — Module 0 on Micron, starting with F1.
- 2Introduction to Financial Options Once Module 0 closes — the three published lectures, with the tutorial and assignment.
- 3Systematic Trading Once Options 101’s published lectures close — phase C0, the starter system.
- 4Smart Portfolios Held until Carver C1 is underway; the cadence decision (Q1) waits for that.
Nothing to re-read yet. Every published note comes back here on day 1, 7, 30 and 90.
Track A — the AI-infrastructure stack
Bottom-up · rungs gate each otherA bottom-up ladder from the machine to the money: what the hardware physically does, what it costs to serve a token, and what that arithmetic implies about the companies selling it. Rungs gate each other — a rung is not opened until the one under it can be explained from memory.
| Rung | Unit | State | Progress | Artifacts | Source |
|---|---|---|---|---|---|
| 1 | Digital Design & Computer ArchitectureSAFARI · ETH ZürichConditional — Studied only if the Computer Architecture L1–L2 gate-check fails. Connor decides, honestly; the desk records the verdict. | Conditional | Not scoped — it is only sequenced if the gate-check fails. | NBBLDGMEBTL0/1 | LINK PENDING |
| 2 | Computer ArchitectureOnur Mutlu · ETH Zürich · Fall 2023 | Ready | 0 / 32 lectures | NBBLDGMEBTL0/2 | ↗ |
| 3 | Inference EngineeringPhilip Kiely · Baseten · 2026Gated — Opens when the memory rung (L2a, L9, L12–L15) closes — the hardware has to be under it. | Gated | 0 / 8 chapters | NBBLDGMEBTL0/3 | ↗ |
| 3 | Efficiency in LLMsAlex Smola · companion to rung 3Gated — Shares rung 3 with Inference Engineering and opens with it. | Gated | 0 / 6 sections | NBBLDGMEBTL0/1 | LINK PENDING |
| 3 | GPU Kernels / MLSysMLSys kernel trackGated — Needs the memory hierarchy under it before occupancy and coalescing mean anything. | Gated | 0 / 4 parts | NBBLDGMEBTL0/1 | LINK PENDING |
| 5 | AI Agents (11-768)Daniel Fried & Graham Neubig · CMU · Fall 2026Honest read — Assignments 2 (Eval, due Sep 24) and 3 (Training, due Oct 22) have no public starter yet — only Assignment 1 is cloneable. Lecture highlights, Canvas submission and compute credits are enrolled-students-only; the seat here is earned on the public starter and the public tests, nothing else. | Ready | 0 / 23 lectures | NBBLDGMEBTL0/2 | ↗ |
| 6 | Renewables in Electricity MarketsPierre Pinson · Imperial College London | Ready | 0 / 6 modules | NBBLDGMEBTL0/2 | ↗ |
| 7 | US power market design and datacenter loadPrimary documents · PJM, ERCOT, FERC, LBNLGated — Opens once the general market design is understood. The US rules are exceptions to it. | Gated | 0 / 6 readings | NBBLDGMEBTL0/2 | ↗ |
| 8 | How compute pays for itself — the filingsCoreWeave 10-K · Nebius 20-F · Nvidia purchase-obligation and guarantee notesGated — Opens once the power-market vocabulary is in place. | Gated | 0 / 3 filings | NBBLDGMEBTL0/2 | LINK PENDING |
| — | RLHF & Post-TrainingNathan LambertNo quota — No weekly slot allocated. Draft Day decision. | Awaiting Connor | 0 / 13 video lectures | NBBLDGMEBTL0/1 | ↗ |
| — | Build a Reasoning Model From ScratchSebastian Raschka | Ready | 0 / 8 chapters | NBBLDGMEBTL0/3 | ↗ |
| — | Modeling: Multimodal AI (MAS.S60 / 6.S985)Paul Liang, Dimitris Bertsimas, Jinhua Zhao & Sang-Gook Kim · MIT · Spring 2026No quota — No weekly slot allocated (Connor, 2026-09-10).Honest read — The graded work (HW1–HW5, the reading assignments, the midterm) is not public, so the lectures alone produce no build — a build has to come from the tutorials or a self-set project using the templates. | Awaiting Connor | 0 / 24 lectures | NBBLDGMEBTL0/1 | ↗ |
| — | World Models (CIS 6280)Jiatao Gu · University of Pennsylvania · Fall 2026No quota — No weekly slot allocated (Connor, 2026-09-12).Honest read — Assignments 1–3 (due Sep 24 · Oct 19 · Nov 18) and the final project (proposal Oct 14, report + code Dec 14) are enrolled-students-only and their specs are not on the public page, so any build here comes from the lecture-linked notebooks or a self-set world-model build in one domain, on the course’s own project brief. | Awaiting Connor | 0 / 23 lectures | NBBLDGMEBTL0/1 | ↗ |
| — | The Modern Software Developer (CS146S)Mihail Eric · Stanford · Fall 2026Honest read — Connor is not enrolled: Gradescope, grading, the OSS-contribution programme (30% of the grade), the final project (50%) and the guest talks (unrecorded) are student-only. What is public is the three lecture transcripts, the reading lists, the slide decks on Drive and the assignment specs — so the seat here is earned on the assignments and the writeups, nothing else. Assignments 4–6 (agent-ready repo · agent-readiness report · TBD) are dated on the calendar but not yet in the repo. | Ready | 0 / 20 sessions | NBBLDGMEBTL0/2 | ↗ |
| — | Modern AI Systems: Agents and System Optimizations (CS2680)Juncheng Yang · Harvard SEAS · Fall 2026Honest read — Connor is not enrolled, and the course gives auditors no compute: no Claude Code licence, no course API key, no HPC GPUs, no leaderboard. A3’s leaderboard (part of its grade) is students-only, and A4–A5 assume the course’s RTX PRO 6000 Blackwell partition, which has no stand-in here yet. A3 and A5 are submitted as published blog posts, which fits /notebook directly. | Ready | 0 / 16 lectures | NBBLDGMEBTL0/2 | ↗ |
Valuation and systematic risk, run as a daily loop rather than a course. It is read here so the two tracks are one picture, but the work happens on the markets desk and every action hands off there — this surface never re-implements the cadence.
| Rung | Unit | State | Progress | Artifacts | Source |
|---|---|---|---|---|---|
| — | ValuationAswath Damodaran · 12-week climb · case company Micron (MU)Awaiting — Module 0 cannot start until F1 is answered. The block is Connor’s, not the machine’s. | Awaiting Connor | 0 / 9 modules | NBBLDGMEBTL0/2 | ↗ |
| — | Systematic TradingRobert Carver · phases C0–C4Honest read — Code-advanced but mastery-dormant: the simulators built from this material pass 35 of 35 tests, and C0 is unstarted. The instrument runs; the operator has not been trained on it. | Ready | 0 / 5 phases | NBBLDGMEBTL0/1 | LINK PENDING |
| — | Smart PortfoliosRobert Carver · 10 bitesAwaiting — Cadence pending Connor’s Decision Q1. The desk surfaces the gap and never invents a default.No quota — No cadence set. | Awaiting Connor | 0 / 10 bites | NBBLDGMEBTL0/1 | LINK PENDING |
| — | Introduction to Financial OptionsAgustin Lebron · Charles University, Prague · Fall 2026 | Ready | 0 / 3 lectures | NBBLDGMEBTL0/2 | ↗ |
- BTLPipeline and cache vocabulary, on a spaced scheduleDigital Design & Computer Architecture
- NBThe memory wall, derived rather than restated from a vendor deckComputer Architecture
- BTLMemory-hierarchy numbers, on a spaced scheduleComputer Architecture
- NBWhy memory bandwidth is the bottleneck, from the physicsComputer Architecture · L15
- NBThe token-price → task-price bridgeInference Engineering
- BLDThe memory roofline, verifiedInference Engineering
- BTLops:byte and KV-cache formulas, spacedInference Engineering
- NBServing unit economics — where the cost of a token goesInference Engineering · ch8
- BLDVRAM / KV-cache calculator, verified against a real runInference Engineering · ch8
- BTLEfficiency levers and what each one costs, spacedEfficiency in LLMs
- BTLOccupancy and coalescing rules, spacedGPU Kernels / MLSys
- NBThe harness flagship — the earned version of the argumentAI Agents (11-768)
- BLDDesign notes into the machine that runs this siteAI Agents (11-768)
- BTLPost-training objectives and their failure modes, spacedRLHF & Post-Training
- NBReasoning models, measured on one small enough to watchBuild a Reasoning Model From Scratch
- GMEA self-refinement loop, playableBuild a Reasoning Model From Scratch
- BTLGRPO mechanics and where it breaks, spacedBuild a Reasoning Model From Scratch
- NBWhat test-time compute actually buys, measuredBuild a Reasoning Model From Scratch · ch4
- NBWhat fusing a second modality costs to serve — measured on a run tutorial, not readModeling: Multimodal AI (MAS.S60 / 6.S985)
- NBA one-domain world model built to the course’s own project brief, with the modelled state, transition and evaluation made explicitWorld Models (CIS 6280)
- BLDThe assignments, done — a dissected trace, an OAuth MCP server, a skill, and weeks 4–6 as they are publishedThe Modern Software Developer (CS146S)
- NBWhat a production coding agent’s system prompt is defending against — read off the captured trace, not the paperThe Modern Software Developer (CS146S)
- BLDThe assignments, done — a trajectory viewer, a harness, an optimised harness, then a self-served model and a full-stack pass as they are releasedModern AI Systems: Agents and System Optimizations (CS2680)
- NBWhat it cost to hold task success fixed while cutting tokens — the A3 write-up, measuredModern AI Systems: Agents and System Optimizations (CS2680)
- NBHow a market clears when the product cannot be stored — the electricity piece of the Older markets seriesRenewables in Electricity Markets
- BTLMarket-clearing vocabulary, spacedRenewables in Electricity Markets
- NBWho earns from AI power demand and who pays for it, from the auction and tariff documentsUS power market design and datacenter load
- BTLCapacity, queue and tariff mechanics, spacedUS power market design and datacenter load
- NBHow a neocloud pays for its GPUs, read from the filingHow compute pays for itself — the filings
- BTLCompute-contract terms, spacedHow compute pays for itself — the filings
- NBThe Micron valuation, with its assumptions exposedValuation
- BTLValuation fact set, spacedValuation
- BTLVolatility targeting and risk decomposition, spacedSystematic Trading
- BTLHandcrafting and rebalancing rules, spacedSmart Portfolios
- NBA gas peaker priced as a call option on the spark spread — where the two tracks meetIntroduction to Financial Options
- BTLPayoffs, parity and option value, spacedIntroduction to Financial Options
Every slot above is currently owed, which is what a syllabus published before the climb looks like. Debt accrues on coverage, not on existence: an all-zero ladder is a contract written in advance, not a pile of failures.
Ledger findings
Targets named but not found on the owning surfaceNo unresolvable targets — every landed artifact points at something that exists.
Candidates
Assessed 2026-08-18 · not allocatedCS 185/285 — Deep Reinforcement Learning
Sergey Levine · UC Berkeley · Spring 2026
free · 25 lectures, 5 homeworks + a final project. Slides and homework public; video links point at the Fall 2023 recordings.
- barely overlaps Build a Reasoning Model From Scratch — Raschka ch6–7 (GRPO) touch the same policy-gradient family, far more cheaply and already on the ladder.
- barely overlaps AI Agents (11-768) — Only the LLM RL lecture pair and its homework speak to the harness argument; the other 23 lectures are control and robotics.
The heaviest item on the list and the only one with a HARD published prerequisite Connor does not meet on the record: CS189/CS289 or equivalent, plus assumed working ability to train deep neural networks. Nothing left on the ladder clears it now that CS336 was removed (2026-10-05) — the transformer fluency would have to come from builds 07 and 10. Recorded here with the prerequisite stated so it stops being rediscovered as "should I start this yet?" every few months. If only one slice is ever wanted, it is the LLM RL lecture pair + HW4; lectures 1–3 (imitation, behavioural cloning) are the approachable on-ramp.
Building a distributed training framework from first principles
Umar Jamil
free · One continuous first-principles build. Duration and publication date unverified.
UNWATCHED — title and channel confirmed via oEmbed, but the video itself has not been opened, so no earned overlap read exists. On its stated syllabus it is rung-4 material (the distributed-training ground CS336 Unit 2 covered before CS336 was removed 2026-10-05), and it is the clean complement to gpu-kernels: that bridge is INTRA-GPU (occupancy, coalescing, tensor cores), this is INTER-GPU (process groups, collectives, device meshes, and pipeline/data/FSDP/tensor/context/expert parallelism composed into one framework). Neither covers the other. Its rare shape is the draw — one self-contained build with a working artifact at the end, and it is now the only rung-4 material on the list. The catch: "no prior knowledge of distributed training required" is true of the distributed part only; it also builds MLA, RoPE, YaRN and a mixture-of-experts layer, so without transformer fluency it is watched rather than built — watching dressed up as building. Revisit once build 07 (attention from scratch) ships — that build is where the transformer fluency comes from.
- §1Architecture and Model Design22 papers · ★1
- §2Efficient Training and Scaling15 papers · ★1
- §3Inference Efficiency and KV Cache8 papers · ★1
- §4Sparse Attention and Long Context9 papers · ★1
- §5Reasoning and Test-Time Compute15 papers · ★1
- §6Reinforcement Learning and RLVR25 papers · ★1ModelsReasoning book ch6–8 · RLHF book
- §7Agent Systems and Tool Use33 papers · ★1Harnessesthe-machine
- §8Coding Agents and Software Engineering16 papers · ★1Harnessesthe-machine
- §9Diffusion Language Models9 papers · ★1
- §10Model Evaluation and Benchmarks13 papers · ★1
Presence on this list is a pointer, not a claim: a source with no close-read file has not been read, and counting it would inflate the shelf the way a timestamp inflates freshness. The catalogue itself lives in the repo at docs/context/readings/llm-papers-2026-h1.md — a working file, not a public route. It is a bibliography, distinct from docs/context/reading-list.json, which is the intake queue where a bookmark gets exactly one disposition — and now has a page of its own. The source list →
Where a saved link goes before it becomes anything. Each one gets exactly one ruling — read it, propose it for the Atlas, write something, build something, or bin it. The bins are on the page too; a list of only the keepers would flatter the judgement.
The curriculum is one typed file, and the study desk is its only writer. There is deliberately no tick-box on this page and there never will be — a second way to mark something done is a second version of the truth, and it would diverge within a week. The spaced-repetition bottles live off this site and are counted here, never shown. Track C was removed 2026-08-10; two tracks is the whole ladder.