Learn the AI that returns a decision, not an essay
A free course on decision models like Jev and the specialized foundation models behind them. Start with no machine learning background. Finish able to build one for under $500 and discuss it at research level.
28 lessons · 836 cited sources · facts as of October 2026
- State
- rpm build log · Fedora 44 · exit code 1
- Ask
- Which failure class is this?Choice
- Toolchain change0.71
- Missing dependency0.17
- Mirror download error0.09
- Something else0.03
confidence 0.61one forward pass, no text generated
Most AI answers in paragraphs. Most work needs a choice.
Routing a ticket, triaging a failed build, flagging a risky change: each is a question with a short list of allowed answers. Writing an essay to pick one is slow, costly and hard to trust.
Generating text
- Writes the answer one token at a time
- Returns JSON you parse and hope matches your schema
- Costs more for every word it writes
- States a confidence that is a weak signal
Scoring the answers you allow
- Reads your input once and scores every option in one pass
- Can only return answers you defined, so nothing to parse
- Jev charges only for input: output is free
- Returns a probability you can set thresholds on
0.1 to 0.65 seconds per request for Jev in independent tests, flat from 1 to 64 questions. Anatomy of Jev
10 to 800 times cheaper per decision: a small self hosted model against the cheapest API tier. The cost table
What you will be able to do
Talk at research level
The debates, the leaderboards and the open problems, with every number linked to the source it came from.
Build on a bootstrapper’s budget
Fine tuning a small model on 100,000 examples costs under $5 of rented GPU time. Plans for $50, $500 and $5,000.
Make data your moat
A one week recipe for 50,000 labelled decisions for about $110, plus the rights checklist that keeps it clean.
Know when to trust a probability
Calibration, conformal prediction and cascades: automate the confident share of the work and route the rest.
See it on a real problem
A Linux distro copilot case study, with an honest answer to “can this make me five times faster?”
Every question with a closed set of answers is a decision. Learn to make it cheap, calibrated and yours.
How the course works
- 01
Start from zero
Five foundation lessons build every idea from arithmetic: what a model is, probability, softmax worked by hand, transformers in plain words.
- 02
Climb to the frontier
Decision models, models that learn from context, and data. Every lesson climbs from intuition to a worked example to the state of the art.
- 03
Pick your track
Builder, Founder or Expert, plus a case study that puts it all to work on a real distro.
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.
- Sources cited
- 836
- Research reports
- 9
- How the facts were checked
- Leaderboards recomputed from raw result files. Claims made only by a vendor, or found in only one source, are marked as such.
Three tracks, one foundation
Everyone reads the shared parts. Then each track goes deep on one goal.
Builder
Train and ship a decision model on a budget: the cheap stack, a distillation recipe, and serving for a fraction of a cent.
Open the Builder trackFounder
Read the market: who builds what, how the money flows, and which wedges a small team can still win.
Open the Founder trackExpert
Hold your own with researchers: the strongest side of each debate and the state of the art on one page.
Open the Expert track
What is inside
Questions
Do I need a machine learning background?
No. Part 1 builds every idea from arithmetic: what a model is, probability, softmax worked by hand, and transformers in plain words. Each lesson then climbs step by step to the state of the art.
Is it really free?
Yes. There is no signup, no paywall and no tracking. Open the first lesson and start reading.
Is this course made by TypeSafe AI?
No. It is an independent course. It explains Jev, TypeSafe AI’s decision model, next to its rivals and the research behind them, and it says so when a claim comes only from the vendor.
How long does it take?
About 8 hours of reading across 28 lessons, roughly 17 minutes each. Read the shared parts, then only the track you need.
Do I need a GPU?
Not to read the course. To train, fine tuning a small model on 100,000 examples costs under $5 of rented GPU time, and small models fine tune on an Apple Silicon Mac.
How current is it?
Facts are as of October 2026, and claims link to dated sources. Prices and leaderboards move weekly, so the lessons say when each number was read.
Which track should I start with?
Everyone starts with Part 1. Then pick Builder to train and ship models, Founder to build a company, or Expert to hold research level conversations.
Can a decision model help me build a Linux distro five times faster?
Not on its own. It speeds up judgement calls, which are a minority of the work, so expect about 1.2 to 2 times. Paired with a coding agent and a fast build loop, an image based derivative distro can get 2.5 to 3 times, and around 5 times at a stretch. The case study works through the numbers.
Is the data rights lesson legal advice?
No. It summarizes rulings and rules as of October 2026 so you know which questions to ask a lawyer.
Start with the first lesson
Free, no signup, no tracking. Read one lesson and decide whether the rest is for you.
Start the free course