Start here
A powerful AI can seem helpful in a chat and still be the wrong thing to trust with real decisions. The question is whether people can still change its course when it goes wrong — not only whether it sounds nice today.
Questions
Where should a new reader start?
Read the short essays. They begin on Start Here — or open the first essay directly. If you already know your role, use the Guided Tour.
What is this site?
Orientation for the Towards Superintelligence Alignment project: short essays, knowledge cards, a field map, experiments, Lean checks, and the book PDF.
Is this a new knowledge platform?
No. It is an authored reference layer for one framework. Discussion and contributions can happen elsewhere.
Does this project claim alignment is solved?
No. It offers a framework and a conditional dependency spine, not a certificate that current systems are safe. See what we are not claiming.
What is the central thesis?
Alignment means preserving grounded, human-correctable value-bearing processes across capability growth, ontology shift, successor creation, and selection pressure — while civilization can still correct.
What are the six thesis claims?
The Introduction's reader contract: boundary, value-bundle, grounding, correction, successor, and basin. Chapter 48 revisits them with status labels and open gaps; it does not treat them as discharged. See the six thesis claims hub.
What is correction-channel integrity?
The condition that legitimate human correction still causally changes future system behavior before irreversible damage. See the correction-channel integrity card.
What is a value bundle?
A compressed direction of control that changes behavior across contexts. The framework cares whether this structure survives transformation — not whether the slogans stay the same. See value-bundle transport.
Why start with boundary discovery?
Because the deployed actor may be a loop of model, memory, tools, dashboards, and people. Auditing only the model can miss the real controller. See boundary discovery.
Does Lean prove alignment?
No. The Lean dependency spine checks conditional structure. It does not prove that real systems satisfy the bridge assumptions.
Do the experiments validate the thesis?
No. They are methodology-building and sanity checks. Negative results bound what the manuscript may claim. The Experiments page splits simulations, external tests, and Witness (stop-checks on records we did not write); coverage and negative results sit next to them. Narrative prose is in docs/EXPERIMENTS.md.
Where should safety engineers start?
Use the Engineer / evals path, then read correction-channel integrity and the deployment-gate card.
Where should policy readers start?
Use the Funder / policy path. It emphasizes decision triggers, scope assumptions, artifacts, and uncertainty. Appendix C is the institutional translation entry.
What is the Field page?
A map of who in alignment research works on which problems, relative to this project's bridges. Start at Field. The longer briefing is one click further.
Where is the full book?
The PDF is bundled with this site — Download PDF from the book page or Start Here. The book page lists contents in PDF order. It opens in the browser without GitHub login.
What is the quiz?
A field-knowledge quiz. Select all that apply; get immediate feedback. After enough answers, it names which researcher knows most about what you already know — coverage, not agreement.
Will there be translations?
Not yet. English is the working language of the book and this site. Other-language editions are not in progress.