This course teaches one kind of AI from first principles. It covers models that answer a question you defined in advance with a typed decision and a probability, not with generated text, and that do it without a classifier trained for each task. Jev, launched by TypeSafe AI on 2026-09-15, is the best-known example. The course also covers the wider family of specialized foundation models: models trained once that then work zero-shot, or from a handful of examples supplied as input, such as TabPFN for tables and TimesFM-3 for forecasting, and their cousins in manufacturing, energy and biology.
You start with no machine-learning background. You finish able to train and ship a small decision model, judge a startup idea in this space, or hold your own in a conversation with a researcher.
I think Jev hit a nerve because it gives developers a typed decision and a probability instead of generated text, without training a classifier for every task. I’d like to see more research labs build specialized foundation models for manufacturing, energy, and biology that work zero-shot or from a handful of examples. The bar is matching or beating general-purpose models at orders of magnitude less inference compute, so they’re cheap enough to run on everything. For example, Google’s TimesFM-3 shows what this looks like for forecasting. It’s a 330M-parameter model that forecasts related time series at once with no task-specific fine-tuning and outputs the whole forecast in one forward pass. Training happens once, and each new user brings their own history as context. The hard part is getting high-quality, vertical-specific, verifiable rewards. You’ll need rights to the data and a clear win in accuracy and compute over existing specialist models.
These are the words of Diana Hu, from the “Vertical Foundation Models” entry in Y Combinator’s Winter 2027 Requests for Startups (YC Requests for Startups, undated, read 2026-10-28). A Request for Startups is YC’s public list of ideas it wants founders to apply with. Hu became a YC Managing Partner on 2026-06-11 (YC blog, 2026-06-11). The page itself carries no date. One source reports that web-archive captures first show the Winter 2027 list on 2026-10-27 (FourWeekMBA, 2026-10-28), and that date rests on that one source.
Both examples in the quote are real and recent. Jev opened on 2026-09-15. TimesFM-3’s weights went public on 2026-08-31, under a licence that does not allow commercial use (Pillitteri, 2026-09-01), so “cheap enough to run on everything” does not yet apply to it in a product unless you go through Google Cloud.
The course takes the quote apart sentence by sentence:
The quote says
Where you learn it
“a typed decision and a probability instead of generated text”
Part 1, then Part 2
“without training a classifier for every task”, “zero-shot or from a handful of examples”
Part 1 and Part 3
“orders of magnitude less inference compute”
Part 1, then the Builder track
“Training happens once, and each new user brings their own history as context”
Part 3
“verifiable rewards”, “rights to the data”
Part 4
“a clear win in accuracy and compute over existing specialist models”
You start from zero. Arithmetic is enough. Every term is defined in plain words the first time a lesson uses it, and every formula comes with a worked example you can check by hand.
You want to reach the state of the art, not stop at the basics. Every lesson climbs the same ladder: the intuition in plain words, a worked example with small numbers, the precise version, then where the field stands in October 2026 and what is still unsolved.
You are bootstrapping. No venture money, no big-tech compute. Whenever cost comes up, the lesson gives the cheap path with real numbers: a consumer GPU, Apple Silicon, a rented GPU for a few hours, or frontier-model labels on a budget.
You build things. The running example is Linux distro building: “will this package build pass?”, “which failure class explains this log?”. The case study asks whether a Jev-like model could help someone build a new distro 5x faster, and answers honestly.
Every lesson separates what changes outcomes by an order of magnitude from what is only 10% polish, and says plainly when something is only polish. A taste, from the course’s research:
Scoring the allowed answers instead of generating text. Fine-tuned small encoders matched or beat GPT-4o and Claude Sonnet 4.5 prompting on 3 of 4 classification datasets, at $5-33 per million requests against $192-2,702 (Valdes Gonzalez, 2026-02-06).
Automating only the cases the model is sure about. If a small model handles 80% of cases and sends the rest to a model 100 times more expensive, you pay about 4.8 times less; at 99% it is 50 times less. That only works if the probabilities are honest.
Distilling once, then running forever. A small model trained on a large model’s labels matched GPT-4’s 94% accuracy on one task and ran at about $2.70 per million sentences against $3,061 (Laurer, Hugging Face, 2024-02-16).
Only polish: choosing between the best modern encoder models. They differ by 0.5-2 points on the standard GLUE benchmark.
Every lesson follows the same template, so you always know where to look:
BuilderFounderExpert
Track badges like the ones above, showing which tracks the lesson is core for.
The idea in one line, the whole lesson in a sentence.
Why this matters, the 10x angle.
The body, three to six sections that climb from intuition to the state of the art and its open problems, with small worked examples and short Python snippets.
Say it like an expert: a few sentences you could say out loud to a researcher or an investor.
Bootstrapper’s angle: the cheap path, with numbers.
Check yourself: three to five questions, with the answers hidden until you open them.
Go deeper: the best three to six sources, primary sources first.
Claims are cited where they appear. When a number comes only from the company selling the product, the lesson says so (“TypeSafe says…”); when only one source reports something, it says that too; and when sources disagree, both sides stay visible. Systems are drawn as plain-text diagrams using the C4 model; how to read the diagrams explains the boxes and arrows in five minutes.
All three tracks share Parts 1-4 and the case study, then branch into a section of their own. These are recommended paths, not fences; read anything you like.
Models as learned functionsParameters, training versus inference, loss and generalization, and why a foundation model is trained once and reused everywhere.
Probabilities, Bayes and expected costDistributions, conditional probability and Bayes worked by hand, expected cost, and why a probability beats a bare label.
Generating text versus scoring optionsTokens, logits and softmax; the autoregressive loop versus scoring known options in one pass; cost, latency and failure modes.
Inside Jev: the contract and the claimsState, typed questions (Choice, Score, Noul), distributions and confidence; what is known versus claimed; where it fits next to LLMs.
How decision models are builtThe lineage from zero-shot classifiers, rerankers and reward models; scoring many options in one pass; output heads; a reference design.
When 0.9 should mean 90%Reliability diagrams, ECE, Brier and log loss; why LLMs miscalibrate; temperature scaling; testing a black-box API.
Turning probabilities into automationCosts and thresholds, abstention, conformal prediction, cascades and routing: automating the confident majority.
Learning a task from its inputWhy a model can learn from examples in its context; networks pre-trained on synthetic data; Bayesian inference paid for once.
Foundation models for verticalsBiology, energy and weather, manufacturing and materials: real wins, overstated claims, and verifiable rewards per field.
Finding data on the open internetCrawls, dumps, APIs, issue trackers and logs; extraction, filtering, deduplication and decontamination; cost per million pages.
Labels from AI and from the worldFree outcome labels, LLM labelling and juries, weak supervision, synthetic data and active learning, on a small budget.
Where distro maintainers' time goesKinds of distro building, from Linux From Scratch to bootc images, and where the hours actually go, with evidence.
The recurring decisions in distro workTyped Choice, Score and Noul questions from everyday distro work, and the labelled data that already exists for each.
Designing the co-pilotThe first decisions to model, the dataset, the base model, training, evals and a success metric, sized for one person.
Is 5x faster real?An honest verdict using Amdahl's law, a fast-plus-slow agent design, and what speeds up for which kind of distro.
Who is building what, and for how muchDecision-model and specialized-model companies, funding, business models, pricing, open versus closed, and where YC's call came from.
Wedges, moats and the proof buyers wantBootstrapper-sized wedges, moats from data rights to cost, the risks, and the evidence that convinces customers and investors.
The field's big argumentsGeneral versus specialized, foundation models versus tuned baselines, calibration, synthetic data, fast versus slow AI, open versus closed.
