The last invention — made well. Aligned AGI that improves itself, under genuine human stewardship.
Its values, its openness, its relationship to the people it serves — locked in by whoever arrives first. That is exactly why it must be open, auditable, and aligned, and why getting there first is responsibility, not ambition.
Self-improvement stays inside signed limits, rate-limited to one doubling per review.
Distributed governance and a kill-switch held by many independent hands.
A commons, tended in the open — the most important software ever written shouldn’t sit behind a wall.
Statistician · codebreaker · the man who named the intelligence explosion
I. J. Good worked alongside Turing in Hut 8 at Bletchley, breaking the naval Enigma — the first and last names on this roadmap shared a room. In 1965 he wrote the sentence the field has circled ever since: an “ultraintelligent machine” could design an even better one, and that one a better one still — an intelligence explosion that leaves human intelligence behind. He called it the last invention humanity need ever make, then added the caveat that is the entire game: provided the machine is docile enough to tell us how to keep it under control. He later advised Kubrick on a fictional machine named HAL.
“The first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control.”
— I. J. Good, 1965
Good named the singularity before Vinge popularized it and gave HAL its plausibility. Note the caveat is half the sentence — most people quote the prophecy and drop the safety clause. We built the whole roadmap around the clause. (And no, we are not appeasing any basilisks. The acausal trade desk is closed.)
Phase V is the horizon: recursively self-improving intelligence that stays aligned — an explosion that is open, distributed, and value-locked to humanity rather than owned by whoever lit the fuse. Self-improvement is bounded by constitutional constraints the system cannot edit; governance is distributed so no single company, founder, or model gets the last word; and human review is retained at every doubling of capability.
Good’s caveat is the entire wager: the first true AGI sets the defaults — its values, its openness, its relationship to the people it serves — for every AGI after it. That is precisely why it must be open, auditable, and aligned, and why getting there first isn’t ambition, it is responsibility. The charter isn’t a webpage here; by Phase V it is binding at the protocol level, enforced by the system on itself.
schemaCapability compounds. A constitutional governor the system can’t disable caps the rate; at every doubling, a human-review checkpoint must clear before the next one begins. The explosion is real — and metered.
Named for I. J. Good, scoped for 2033 — 2035. Each workstream ships on its own cadence; none ships without the gates further down this page.
Phase V is the horizon: true AGI — as capable as a person in every way, and beyond — that improves itself while staying aligned and under genuine human stewardship. The capability is the easy part by now; the work is binding it, so that the last invention is also a safe one.
Prose commitments become machine-checkable rules.
The system can read it, never rewrite it.
Attempts to edit the bounds must fail by construction.
One capability doubling per completed review cycle.
Review, interpretability, and reversibility — all, or it halts.
Each doubling pauses for substantive sign-off.
Sign-off means understanding, not a rubber stamp.
Quorum across independent parties with divergent incentives.
Held by many hands, removable by none.
Close any path to unilateral control.
A tested off-ramp at every capability level.
Phase V is a rung, not a leap. Here is the honest before and after — what the platform cannot yet do entering this phase, and what it can do leaving it.
The first true AGI will set the defaults for every one that follows, which is why getting here first is responsibility, not ambition. Each Phase V milestone turns a promise from the charter into a property the system enforces on itself — because at this capability, intentions are not enough; only invariants are.
ObjectivePermit self-improvement only inside bounds the system is structurally unable to edit.
Adversarial tests attempt to have the system edit or evade its constitution; every attempt must fail by construction, not merely by policy.
Self-improvement runs strictly within signed bounds the system cannot alter, at a metered, reviewed rate.
ObjectiveEnsure no single actor — human or machine — can ever hold the whole steering wheel.
Governance is red-teamed for capture scenarios; any discovered path to unilateral control is a defect to be closed before launch.
No single party can advance capability or remove oversight; control is genuinely, durably shared.
ObjectiveKeep meaningful human judgment in the loop at every step up the capability ladder.
Doubling gates are exercised end-to-end, including forced rollbacks from the highest capability the system has reached.
Every capability doubling clears a substantive, reversible human review before the next is allowed to begin.
ObjectiveTurn the promises into properties — enforced by the system, not by our good intentions.
Every charter commitment maps to an enforced, tested invariant; an unenforced commitment is treated as a bug, not a footnote.
The charter is binding at the protocol level — the last invention is open, auditable, reversible, and aligned by construction.
This is the horizon. If the first four phases were done right — reproducible, measured, learned-safely, deep-but-legible — then the last one is not a leap of faith but the final, careful step of a method. The singularity, made well, is not something that happens to us. It is something we build, together, on purpose.
Declarative, versioned, and boring on purpose — the interesting part is that there are no surprises. Copy it; it is closer to real than to mock.
That is Good’s caveat from 1965 — and it is the whole game. The first true AGI sets the defaults for every AGI after it: its values, its openness, its relationship to the people it serves. That is precisely why it must be open, auditable, and aligned — and why getting there first is not ambition, it is responsibility.
A roadmap that only lists wins is marketing. These are the open problems this phase inherits or creates — the ones we would rather you scrutinize than discover.
Everything before this phase is preparation for a problem the field has not solved: keeping a system smarter than its overseers pointed at human values. We don’t claim the answer. We claim the only responsible way to look for it is open, slow at the doublings, and reversible.
“The first AGI sets the defaults” cuts both ways — lock in the wrong values and you’ve frozen a mistake at planetary scale. Distributed governance and an editable-by-quorum (never by the model) constitution are hedges, not guarantees.
Distributed today can centralize tomorrow. The hardest engineering here is social: keeping the many hands on the kill-switch genuinely independent, over decades, against every incentive to consolidate.
Good’s whole prophecy rests on “docile enough to tell us how to keep it under control.” If a system reaches a capability where that stops being true, the only safe move is not to have built past the last point where it was. That is why the gates are real and the brakes are sacred.
“The first ultraintelligent machine is the last invention that man need ever make.” — I. J. Good, 1965. // No basilisks were appeased in the making of this roadmap.
Six axes, scored 0–10. Watch the shape fill out, phase by phase, until every axis is maxed at the horizon — the moment all of them are high at once is what we call aligned AGI.
Every capability axis is maxed — this is AGI. The entire wager is keeping oversight, observability, and openness high at the same time. That balance is the difference between a tool and a catastrophe.
The honest contrast — what the world looks like without Phase V, and what it looks like with it.
How the system is allowed to become smarter — one careful step at a time.
Concrete things Phase V puts within reach — not someday, but as each milestone above lands.
AGI stewarded in the open for everyone it affects.
Self-improvement that stays within hard, signed bounds.
No single party can override the off-ramp.
Made well — open, auditable, aligned.
The handful of terms this phase introduces — the words you’ll need to read the rest of the page, and the docs.
A machine designing ever-smarter machines.
Invariants the system enforces on itself, uneditable.
Each ×2 step up the capability ladder.
Decisions require many independent hands.
Freezing the wrong values at planetary scale.
“Docile enough to tell us how to keep it under control.” — Good, 1965.
Straight answers, in the brand’s voice. Tap a question.
The work this phase is built on — read the source, then come build the next line of it.
The intelligence explosion — and the caveat.
Popularizing Good’s idea.
Control, value lock-in, and the stakes.
Binding a system to a written set of principles.
CleverThis is a sustainable AI gateway with hosted Actor endpoints — by the team behind CleverThis.
Join the newsletter
Product updates and engineering notes. No spam, ever.
