Alex Stremsky · papers
The Convoy Papers
The frontier-AI race selects for the holder of a decisive lead one would least want, and a formation of rival developers that keeps anyone from gaining such a lead can be engineered rather than wished for.
The Council Papers here define a proposal for this, and The Constellation Papers describe the physics and the theorem underneath.
Pick by interest and by the time you have.
Where to start
The frontier AI race, and what to do about it
- 1 MIN The Council Papers section below.
- 10 MIN the abstract and §3.1 of A Council of Rivals.
- 30 MIN the whole of A Council of Rivals, then the abstract of The Verification Model.
Verification and compute governance
- 1 MIN The Council Papers section below.
- 10 MIN the abstract of The Verification Model, then the abstract and §3.1 of A Council of Rivals.
- 30 MIN the whole of The Verification Model, with §3.1 of A Council of Rivals for the margin it serves.
Why a lead that is defended starves its holder
- 1 MIN the abstract of Why the Winner Starves.
- 10 MIN the abstracts of Why the Winner Starves and The Physical Prices of Knowledge.
- 30 MIN the whole of Why the Winner Starves, with the abstract of The Physical Prices of Knowledge for the bounds it uses.
The physics of knowledge, and a knowledge-based civilization
- 1 MIN The Constellation Papers section below.
- 10 MIN the abstract of The Physical Prices of Knowledge.
- 30 MIN the whole of The Physical Prices of Knowledge, and a skim of The Singularity Race (2016) for where it all began.
The papers
They stem from a 2015–2016 series of posts on kurzweilai.net under the title The Singularity Race (the forum is gone; the thread is preserved on archive.org).
They belong in two groups.
The Council Papers
Frontier AI is a race that selects for the actor most willing to cut safety and to value absolute power. If such an actor gains a decisive advantage, the result is either a singleton — one actor with unchallengeable control — or a desperation war, started by rivals who see the singleton coming and strike first. And even if no one gains that advantage, the race forces every participant to cut safety, which is unacceptable on its own.
These papers propose a system for regulating the speed of AI development that is designed to work in real-life conditions: it relies on the participants' own interests, not on their benevolence, wisdom or perfect rationality.
A Council of RivalsPDF
The central body of the proposal: a design that avoids the dangers of the frontier AI race without blocking development, and the mathematics behind it.
- a bounded release window — the lead any member may hold over the rest — kept under half the lead the others could tolerate before turning to force
- a fault-tolerance margin under which compliant members together always outweigh any possible coalition of defectors
- verification run as a published budget, not a promise of omniscience
- a frontrunner compensated for slowing, rather than throttled
- endgames pre-committed, so that no vote is needed when the moment comes
The Verification ModelPDF
How the Council establishes the real situation that its mechanisms act on.
- replaces the question "is this member verified?" with "how much can the Council fail to see, and is that less than its margin tolerates?"
- nine observation channels, each priced by how much it can miss
- eight ways to conceal capability, each with its counter and with what the counter leaves unseen
- a reference algorithm that enforces the bound at the level of coalitions without having to detect a coalition
- the attacks found so far that the model cannot defend against, stated by name
Other elements of the design
Papers in preparation:
- the bounded window that allows fast development while avoiding its risks
- the Commons: competition at the frontier edge, cooperation behind it
- the structure of membership that makes the Council's work possible
What would break this design
A design that does not look for what would break it, and publish the list, has zero fault tolerance and is not worth considering for the real world. The falsifiers that would take the whole structure down:
- a decisive lead that does not depend on knowledge
- a verification blindness that cannot be bounded at any price
- a frontrunner for whom no compensation beats defection
Each paper also names its own.
The Constellation Papers
The epistemological foundations, from the physics up, of a knowledge-based, potentially post-Singularity civilization. Which structures and choices serve it better, which worse, and which are outright bad.
The Physical Prices of KnowledgePDF
The base for all the other Convoy Papers. Prices the elementary operations of knowledge at Landauer's rate, with del Rio's bound, and from that derives the existence of — and the dependencies of — an optimal size and density for bodies of knowledge. It also examines their metabolism and dynamics, and finds the bounds within which each side of the Kurzweil–Modis dispute is right.
Why the Winner StarvesPDF
A leader whose lead is made of knowledge, and who defends it, starves its own further growth — regardless of its intentions; the result is the same for a guardian and for a tyrant. The convoy — a formation of knowledge systems that are not equalised, but do not drift too far apart either — is the only regime in which a leader's security and its growth stop being at war.
Built on the bounds of The Physical Prices of Knowledge; shown with a set of models that compare configurations of such a formation.
Other
In preparation: the morals and ethics of transhuman AI; the construction and basic ruleset of such a civilization; its milestones and other aspects.
All papers are on Zenodo under CC-BY 4.0, with DOIs and versioned supplements; the models can be run from them.
Alex Stremsky (alex@stremsky.org)
Last updated 2026-10-04.