Reading list

These are the sources the course leans on most, chosen so that each part has a path from a friendly first read to the frontier. Every entry says why it is worth your time and how hard it is:

• Beginner: readable with no ML background, or after the matching lesson.
• Intermediate: assumes the lesson; you will skim some maths.
• Advanced: research-level; read when you want to argue with the authors.

Within each part, the list runs from Beginner to Advanced. Dates are publication dates where the course’s research recorded them; leaderboards and docs change, so check them live. If you read only one source per part, read the first one.

Lessons: What a model is to Three ways to teach a model.

• Beginner · TimesFM-3 announcement (Google Research, 2026-08-31): “train once, bring your own history as context” in plain words, with the model from the course’s opening quote.
• Beginner · SetFit: efficient few-shot learning (Hugging Face, 2022-09-26): a small model fine-tuned on a handful of examples, in 30 seconds for $0.025; the clearest picture of fine-tuning versus prompting.
• Intermediate · The Foundations of Cost-Sensitive Learning (Elkan, IJCAI 2001): a few pages that turn a probability and a cost table into a decision threshold.
• Intermediate · Attention Is All You Need (Vaswani et al., 2017-06-12): the original transformer; read sections 1-3 and Table 1 on cost.
• Intermediate · BERT (Devlin et al., 2018-10-11): the encoder that made “pre-train once, add one output layer” the standard for classification.
• Intermediate · Language Models are Few-Shot Learners (Brown et al., 2020): GPT-3 and the zero-, one- and few-shot vocabulary the course uses.
• Intermediate · LoRA (Hu et al., 2021-06-17): the short, readable paper behind almost every cheap fine-tune.
• Advanced · Scaling Laws for Neural Language Models (Kaplan et al., 2020): Table 1 is the source of the 2N and 6N FLOP rules that tie model size to cost.
• Advanced · Strictly Proper Scoring Rules, Prediction, and Estimation (Gneiting & Raftery, JASA, 2007-03-01): the reference on why honest forecasts score best.

Lessons: Anatomy of Jev to From probability to automation.

• Beginner · TypeSafe docs: Primitives and Confidence (read 2026-10-28): the Choice, Score and Noul contract and the exact confidence formulas, from the vendor.
• Beginner · scikit-learn: Probability calibration: reliability diagrams, Platt, isotonic and temperature scaling, with runnable code.
• Beginner · An Independent Evaluation of TypeSafe’s Jev (Vals AI, 2026-10-06): the most careful third-party test of speed, accuracy and calibration, task by task.
• Intermediate · How Jev works (Victor Dibia, 2026-09-28): the option-scoring trick rebuilt on Qwen2.5-7B for about $4, with calibration numbers.
• Intermediate · Jev Decision Index (live leaderboard): 240+ decision models on one board, with ECE per task; learn to read it critically.
• Intermediate · Benchmarking Zero-shot Text Classification (Yin, Hay & Roth, 2019-08-31): NLI as a classifier, the start of the lineage Jev sits in.
• Intermediate · On Calibration of Modern Neural Networks (Guo et al., 2017-06-14): the paper that made temperature scaling the default fix.
• Intermediate · A Gentle Introduction to Conformal Prediction (Angelopoulos & Bates, 2021-07-15): the best practical guide to guaranteed prediction sets, with code and sample-size tables.
• Advanced · Language Models (Mostly) Know What They Know (Kadavath et al., 2022-07-11): base models calibrate, RLHF breaks it, a temperature of 2.5 repairs it.
• Advanced · NanoJev PR #14 (2026-09-21): a small, readable implementation of block-masked option scoring on a 0.6B decoder, with honest speedups and slowdowns.

Lessons: In-context learning and PFNs to Vertical foundation models.

• Beginner · TabArena leaderboard (live): the tabular board with slices, timing and leak warnings; see where tuned trees still sit.
• Beginner · GIFT-Eval leaderboard (live): time-series standings with leak flags; compare any vendor claim against it.
• Intermediate · Accurate predictions on small data with a tabular foundation model (Hollmann et al., Nature, 2025-01-08): TabPFN v2, its synthetic prior and its honest limits.
• Intermediate · nanoTabPFN and its code: train your own PFN in about a minute on one GPU and see why the prior matters.
• Intermediate · Nixtla’s Chronos experiment: the sharpest case that a statistical ensemble can beat a foundation model.
• Intermediate · ECMWF: AI forecasts become operational (2025-02-25): the one operational comparison behind weather’s orders-of-magnitude energy win.
• Intermediate · Matbench Discovery (live): the materials leaderboard, where 30-60M models already match a 730M one.
• Advanced · Transformers Can Do Bayesian Inference (Müller et al., 2021-12-20): the original PFN recipe and the proof that its loss targets the posterior predictive.
• Advanced · What Can Transformers Learn In-Context? (Garg et al., 2022): the cleanest experiment showing a transformer learning to be a learning algorithm.
• Advanced · Ahlmann-Eltze, Huber & Anders (Nature Methods, 2025): the baseline audit where deep models failed to beat linear ones on perturbation prediction.

Lessons: Designing the questions to Data rights.

• Beginner · Common Crawl: August 2026 crawl archive (2026-08-24): what is in a monthly crawl of the web and how to get it.
• Beginner · FineWeb-Edu dataset card (v1.4.0, 2025-07-11): LLM labels, an additive rubric and a small classifier run over the whole web; the Jev pattern at scale.
• Intermediate · The FineWeb datasets (Penedo et al., 2024-06-25): the full, ablated recipe for turning Common Crawl into clean text.
• Intermediate · Who Validates the Validators? (Shankar et al., 2024-04-18): the evidence for criteria drift, and why you label before you freeze a guide.
• Intermediate · Replacing judges with juries (Verga et al., 2024-04-29): why a panel of small, different models beats one big judge at over 7x lower cost.
• Intermediate · Pervasive label errors in test sets (Northcutt et al., 2021-03-26): even famous test sets are wrong at least 3.3% of the time; the case for label cleaning.
• Intermediate · Overview of guidelines for GPAI models (FLI, 2025-07-30): who counts as a general-purpose AI provider under the EU AI Act, and why a small decision model probably does not.
• Intermediate · Third Circuit rejects fair use in Thomson Reuters v. Ross (Haynes Boone, 2026-10-02): the first US appeals ruling on AI training, and the one closest to a decision model.
• Advanced · DeepSeek-R1 (2025-01-22): rule-based rewards only, chosen to avoid reward hacking; the reference for verifiable rewards.
• Advanced · Recent frontier models are reward hacking (METR, 2025-06-05): concrete cases of models gaming their checkers; read before you trust any reward.

Lessons: Where the time goes to Is 5x real?.

• Beginner · bootc documentation: what it means to ship an operating system as a container image.
• Beginner · Fedora wiki: F42 GCC 15 mass prebuild (2025-01-10): a real, hand-made failure taxonomy with counts, and how a toolchain campaign runs.
• Beginner · Wikipedia: Amdahl’s law: the formula and worked examples behind the honest 5x verdict.
• Intermediate · Log Detective in Packit (Fedora Magazine, 2026-05-13): how Fedora wired an LLM failure explainer into its CI, with its own caveats.
• Intermediate · fedora-copr/rpmeta: a small, typed, non-generative model already inside Fedora’s build system.
• Intermediate · LWN: Machine learning and stable kernels (2018-09-12): the precedent for a typed decision model in kernel maintenance.
• Intermediate · Fedora AI-assisted contributions policy (approved 2025-10-22): what automated validation may and may not decide.
• Advanced · Leakage and the reproducibility crisis (Kapoor & Narayanan, 2022-07-14): the leakage traps that bite outcome-labelled data, across 17 fields.
• Advanced · METR’s randomized trial of AI coding tools: 16 experienced developers and 246 tasks; the evidence behind the caution on coding agents.

Lessons: The bootstrapper’s stack to Serving for a fraction of a cent.

• Beginner · Unsloth: Train your own decision model: the shortest path to a Jev-like model; read its numbers as vendor claims.
• Beginner · mlx-lm LORA.md: every flag you need to fine-tune on a Mac.
• Beginner · Modal: cold starts: what serverless start-up and idle billing really cost.
• Intermediate · Synthetic data: save money, time and carbon (Hugging Face): a RoBERTa student matching GPT-4’s labels at $2.70 against $3,061 per million.
• Intermediate · Mapika/decider-2b model card: one person’s open decision model, trained in 56 minutes on one GPU, with its recipe and calibration caveat.
• Intermediate · ModernBERT (Warner et al.): the long-context encoder most bootstrapper plans start from, with throughput tables.
• Intermediate · ONNX Runtime: quantization: dynamic versus static int8, and when it can be slower.
• Intermediate · QLoRA (Dettmers et al., 2023-05-23): how a 4-bit base lets a 65B fine-tune fit one 48 GB GPU, with the authors’ own benchmark caveats.
• Advanced · Cost-Aware Model Selection for Text Classification (Valdes Gonzalez, 2026): fine-tuned encoders against prompted frontier models, with prices and latencies; the evidence for “one to two orders of magnitude”.

Lessons: The market map and Wedges, moats and proof.

• Beginner · YC Requests for Startups (Winter 2027 list, read 2026-10-28): the “Vertical Foundation Models” entry and its siblings in YC’s own words.
• Beginner · The Bitter Lesson (Sutton, 2019-03-13): one page, and the risk every specialized-model pitch must answer.
• Beginner · Bessemer: State of AI 2025: gross-margin benchmarks for AI companies, to sanity-check your own.
• Intermediate · SAP to acquire Prior Labs (SAP News, 2026-05): the lab-to-acquisition path, in the acquirer’s words.
• Intermediate · Microsoft separates decision-making from generative AI (InfoWorld, 2026-10-13): how a frontier lab matched the primitive and the price in under a month.
• Intermediate · Prior Labs docs: Models: the non-commercial terms to read before you build on TabPFN.
• Intermediate · Browserbase: What is Jev?: a threshold-plus-fallback deployment with measured gains.
• Intermediate · EIC Accelerator: the largest grant-plus-equity route for an EU founder.
• Advanced · Arize: decision model benchmark (2026-10-07): the first independent multi-vendor test, with calibration, cost and the rewording flip test; the template for your own evidence pack.

Lessons: The big debates and Talking SOTA.

• Intermediate · Fine-Tuned “Small” LLMs (Still) Significantly Outperform Zero-Shot Generative AI Models (Bucher & Martini, 2024-06-12): learning curves showing a few hundred labels beat GPT-4 zero-shot.
• Intermediate · Is Model Collapse Inevitable? (Gerstgrasser et al., 2024-04-01): accumulate real and synthetic data, don’t replace.
• Intermediate · GIFT-Eval (Salesforce): a forecasting benchmark built with a non-leaking pretraining set; the model for honest evals.
• Intermediate · Densing law (Xiao et al.): capability per parameter doubling about every three months, and what that does to cost moats.
• Advanced · Beyond Chinchilla-Optimal (Sardana et al.): why inference-heavy products should over-train small models.
• Advanced · Mitigating Bias in Calibration Error Estimation (Roelofs et al., 2020-12-15): why binned ECE misleads and what to report instead.
• Advanced · Beyond Binary Rewards (RLCR) (Damani et al., 2025-07-22): a proper scoring rule in the RL reward restores calibration; the frontier of “can calibration be trained in?”.
• Advanced · Tan et al. on language models as forecasters (arXiv 2406.16964): removing the LLM from three LLM-based forecasters matched or improved accuracy in most cases (26 of 26 for Time-LLM); why a general LLM is the wrong tool for numbers.