Repository navigation
Releases
v1.1.0 - a record anyone can verify
The course itself is unchanged. All fifty lessons are the same, and every figure in them still comes from running the code. What this release adds is the machinery around them, so that more of what the repository claims is checked by a program rather than asserted.
One result in full. The front page now shows a complete result from phase 9 as the program prints it, including the table that locates the signal in a single vector. It is marked as an excerpt, and the build fails if it stops being a verbatim slice of what that lesson prints.
Figures outside a lesson are checked too. check_output.py only ever read lesson pages, so a number quoted on the front page was unguarded. It is not any more.
Every link is checked. check_links.py resolves every relative path and every heading anchor across all 69 pages, using GitHub's own anchor rules. It deliberately does not fetch external URLs: the course promises to run with the network closed, and a guard that fails because somebody else's server is down teaches readers to ignore it.
The record of completion, and a way to doubt it. The record is redrawn, and verify_certificate.py recomputes its credential from the file, or from fields typed off a photograph. One flag checks out the commit on the record and runs the whole course. VERIFY.md sets out every route, and the site carries the same thing as a page.
A machine to borrow. A devcontainer, so the first lesson runs in a browser tab for a reader who cannot install Python. It weakens nothing: every lesson still imports only the standard library, which is checked on every change.
One blue. The mark, the header, the record and the site now share a single accent colour, and the social preview exists in light and dark.
v1.0.0 - the course is complete
The course is finished. Fifty lessons across ten phases, from opening a terminal to writing a transformer block and proving its gradients by hand. Nothing is installed at any point: no account, no API key, no GPU, no paid service, and no network access while a lesson runs.
What is in it
One system, built the whole way through. It starts as a rule that tells a question from a statement and ends as something with tools, a retrieval index, a cost meter, an access boundary and a release sheet, with a transformer block built beside it.
| Lessons | 50, each a README.md, a solve.py and a check.py |
| Phases | 10, about 37 hours in total |
| Dependencies | none, enforced by a program rather than a promise |
| Network | closed while every lesson runs |
The curriculum lists every lesson. On the website the boxes tick off and remember where you stopped.
The five guarantees, each one a program you can run
python tools/run_lessons.py every lesson runs, on Linux, macOS and Windows
python tools/check_output.py every number shown is the number the code prints
python tools/check_imports.py nothing outside the standard library, anywhere
python tools/check_determinism.py the same output under three different hash seeds
python tools/check_glossary.py every term the course uses is definedThey run on every push, and nightly.
What this release adds
Phase 9 is complete, and it is the part of the course most worth arguing with:
- Lesson 3 takes the attention mechanism apart. A model with no query and no key, taking the mean of the values, comes within one question of the learned head. The weights are not where the signal is, which is why an attention heatmap is not an explanation.
- Lesson 4 adds position embeddings, then audits the held-back split and finds that 8 of its 10 questions had no answer in the training data. It ends by showing that which nine examples a model sees decides whether it learns an order at all, where how many it sees does not.
- Lesson 5 builds the full block, proves its backward pass against finite differences, and finds every architectural difference smaller than the spread between seeds of the same model.
Three of those five lessons end by failing to show what they set out to show, and they say so.
Also in this release
- A curriculum page listing all fifty lessons, generated from the lessons themselves and checked in CI so it cannot drift.
- Search metadata the site was missing: a description per page rather than one repeated across all of them, structured data for the course, the questions and the glossary, a
robots.txt, andllms.txtserved at the root. - A record of completion that runs all fifty lessons before it will draw anything, and refuses if one fails.
Starting
git clone https://github.com/Amey-Thakur/AI-ENGINEERING
cd AI-ENGINEERING
python tools/run_lessons.py phases/00-your-machine/01-the-terminalIf that fails because you have never opened a terminal, that is what lesson 1 is for.
A note on what it does not claim
The course measures its own results and reports them when they are unflattering. Phase 7 turns its tests on the course's earlier work and finds phase 3's classifier indistinguishable from a coin and phase 6's retrieval half indistinguishable from its absence. Those findings are left in. They are the reason the rest of the numbers are worth anything.