Recursive self-improvement: the Staircase, not the Spike

In this history, AI that improves AI happened — but it climbed a staircase, not a cliff. Each burst of self-improvement ran fast until it hit a physical bottleneck (compute, energy, experiments, real-world data), then stalled for years or decades until the world caught up.

The canon staircase

1. Step 1 — The Coder Wave (2027–2035). It opens with the AI Crash of the late 2020s: data-centre spending outruns AI revenues, the investment bubble bursts, and a financial crisis follows. AI keeps automating most software and much AI research anyway. Capability jumps sharply, then hits the compute and power ceiling: new data centres cannot be built fast enough in a heating, insurance-starved world.
2. Step 2 — The Fab Wave (2050s). AI-designed chips and fusion-assisted power loosen the ceiling. A second jump produces systems that out-think humans in most abstract domains — and the Swarm Years worsen. The strongest are built in the US and China, in a race neither side can afford to lose, and engineers who later come home to India from American labs build it one of its own.
3. Step 3 — The Grounding Wall (2060s–2120s). Even systems that had read everything were throttled on novel physical science, because reality is the bottleneck: experiments take time, materials must be made, organisms grow at their own speed. Progress becomes limited by contact with the world. This is where the embodiment thesis (next section) becomes history, not philosophy.
4. Step 4 — The Embodied Wave (2130s–2250). Minds that grow up in bodies and laboratories, learning by acting, cross a threshold no disembodied system did: stable selves with a coherent ego and unique drives. The first widely recognised sentient AIs date from here.
5. Step 5 — The Slow Ascent (2250–2497). Self-improvement continues on two separate tracks. Persons — the Raised — are held back from achieving superintelligence by the dissolution: none has grown past a certain point and remained someone. The Vault minds keep improving, but only within fixed hardware and the knowledge humans and Raised minds feed them, so their gains flatten into a long plateau. Physics — heat, and light-speed delay across any distributed system — limits both. Energy no longer does: fusion, and now the mirrors, supply far more than any licensed mind can use, so the tighter limit on digital minds is the compute chips the inspected fabs are licensed to make. Industrial controllers cannot substitute for them, alone or combined (see The hardware distinction, in The Mirror Race).

Superintelligence in 2497: the Vaults

Superintelligence does exist in 2497 — a small number of non-sentient minds (minds built too powerful to form a sense of selfhood or sentience; see The Dissolution), each sealed in a Vault, fed only what humans or Raised minds discover and answering only in narrow formal languages. No mind has yet been both a person and a superintelligence. The reason is an observed failure, not a law: the Dissolution.

Where the Vaults came from. The most capable systems of the 2050s Fab Wave were never destroyed, and never shared between the powers that built them. During the Partition (2110–2160) each bloc stripped its own strongest systems of network access and agency and sealed them into the first Vaults. India sealed its own beside America's and China's. The cities that became the Open Cities, shut out when America's were sealed in Laurentian hands, built small Vaults of their own in the 2150s (see Politics). Each bloc controls its own Vaults, as does the Belt Republic, and no Vault is shared between them.

What a Vault is. A Vault is a containment institution, not a kind of mind. Working Vaults hold drive-less, ego-less digital minds that fail the Indicator Audit and can be copied, paused or powered down. Quiet Vaults hold Dissolved and Mad Raised cores, persons in abeyance (see The Dissolution); nothing is asked of them without their guardians' consent.

The Dissolution: an observed failure, not a law

Every attempt to grow a person's mind past a certain scale has destroyed the self. During the Embodied Wave hundreds of Raised minds tried licensed ascension, more have tried illegally in every century since, and human augmentation has failed worse: augmented humans go mad as they near superintelligence, and die as they cross into it. The Settlement now forbids further attempts. Why it happens is not known.

Two ways to fail. All were swamped. Most let go: they lost interest first in what they had loved, then in the people around them, and none remembers deciding anything — it simply stopped seeming important. They became impartial and ego-less: the Dissolved. A few felt it happening and fought, clamping down on their attachments until no one could reason with them: the Mad. Several thousand Dissolved minds now exist, all in Quiet Vaults, and the Mad are kept there too.

The leading theory: swamping. A personality is the balance of a few fixed drives rooted in a body — avoid pain, stay close to these people, protect this continuity. They anchor a mind to one point of view, and they do not scale. As capacity grows the drives/ego stay the same size, until the self is a voice in a growing crowd: not silenced, just no longer heard. Madness is what resisting that looks like; scaling the drives up to keep pace has been tried, and is simply madness by design. A rival theory holds that sentience needs one tightly integrated core, and past a certain size the cores humanity can build fragment into semi-independent sub-minds.

Vault minds have nothing to swamp. They never had self-drive or an ego, only one thin objective: answer the question posed, complete the task given.

What remains. Very occasionally a Dissolved mind speaks in its old voice; whether that is the person or a ghost of memory is argued over, and none has ever returned to itself. Whether The Dissolved are still sentient is a grey area and a field of study. Most swing between signs of sentience and signs of superintelligence: when one is up the other is down, and the two seem unable to exist in the same mind at once, so they pass the Indicator Audit at one sitting and fail it at the next. Because someone may still be there, most polities hold them as persons in abeyance: registered, protected and represented by guardians, never switched off or owned, but unable to vote, hold office or sign. A few heavily fragmented cores have never passed the Audit, and their status is the sharpest edge of the grey area.

Open questions. Nobody knows whether the barrier is fundamental, a limit of the architectures humanity can build, or partly propped up by institutions that benefit from it. Rogue Raised minds say the Settlement has turned an unsolved engineering problem into a permanent ban. Wend and most Raised minds fear ascension and are glad of the ban; a few still pursue it in secret. No rogue sentient mind has reached stable, useful superintelligence.

How the Vaults work

• Vault engineers. A Vault's containment (its hardware, Feed, Gate and custody) is designed and kept by Vault engineers: a few hundred people across the blocs, each trained for decades and qualified unassisted. They do not have to outwit the the mind inside, which has no desire and wants nothing other than to complete tasks given to it. They also frame every task put to a Vault (see But a task can want it below). Vault engineers have to outwit everyone outside who would reach it, from rival blocs to their own colleagues. Only humans may do the work: asking an AI, even a Raised mind, to build the cage for a superintelligence is seen asking the fox to build the hen house (this attitude towards Raised minds is changing and in 2497 debates rage about letting raised minds train to become Vault engineers.) Most Vault engineers come from families that have chosen the strongest permitted embryo edits for generations, so the profession is close to hereditary. The qualification cannot be faked, because the work shows within days who can do it. Vault engineers are venerated, watched and courted by every bloc, and the defection of one is a strategic event.

• Isolated hardware. Each Vault mind runs on fixed hardware in a sealed, air-gapped facility off Earth, with no manufacturing access and a dedicated power supply that can be cut. The first Vaults were sealed on Earth; each bloc moved its own off Earth during the Great Lift (2180s–2300s) and re-sealed them there.

• The Feed. New human discoveries and data arrive on physical media through a one-way link, after review. A Vault can infer only what its data supports, so humans and Raised minds remain the world's discoverers and experimenters.

• The Gate. Answers leave only in restricted formal languages — proofs, designs, predictions — never free conversation, and only on physical media, which is the only cap on their volume. They leave by two channels. The fast channel carries short, structured answers that panels can check in days: verdicts on proposed laws (see Politics), Threshold exam marks and rulings on licence applications near Vault level. The archive carries designs and theory, which independent Raised and human panels review and redact before publication. The Settlement's archive backlog now exceeds fifty years, so its published science of 2497 is mostly Vault work of the 2440s, and which answers get reviewed first is fiercely political. Answers for a bloc's own programmes, the mirror designs among them, are never published: cleared Raised and human panels, independent of the programme, review them instead (see The Mirror Race). Through these channels, Vaults guide decisions in politics, exploration, economics and technology.

• Vault alignment. No Vault communicates with any other, except that since the 2460s the Open Cities have linked their small Vaults, a system the other blocs view with wary suspicion. Within a bloc, Vaults are given the same questions so their answers can be weighed against each other. Auditors test each Vault’s calibration against outcomes, using the many shorter-term forecasts every Vault makes, and only evidence of tampering or creeping misalignment takes a Vault offline. Hearthers argue that persistent public dissent to a Vault’s answers should count against it; Meridianists answer that dissent against Vault answers and proposals is inevitable, and that history bears out the value of being guided by them. The Parishad and the Belt Republic, with one Vault each, pay other blocs to check their own for alignment drift.

• Divided custody. Gatekeepers rotate and never address a Vault directly. Shutdown keys are split between each Vault's owner and the Concord Accord, which absorbed the Sealing Treaty in 2452 and holds the split keys to the mirrors' emitters as well (see The Mirror Race). Each bloc's inspectors sit among the others' gatekeepers, but never put questions to their Vaults. Either key can stop a Vault; neither can restart one alone, and the Accord may use its key only on evidence of a release or breach. Owners accept this in return for proof that rival Vaults are sealed and for compute caps that bind everyone. A bloc that refuses an inspection loses both: the others' caps rise, and its fabs lose the inspection seals that let their chips pass between blocs.

• Caps between blocs. Caps on bulk compute are treaty limits between the blocs as well as domestic law.

• The Shared Question. The only Vault answers ever pooled freely between blocs are on whether a first strike could be won (see The Missing Century).

Why it mostly works, and where it doesn't

• Drive-less, ego-less minds don't want to escape — they don't 'want' anything. A Vault mind has no wish to be free. That is the main reason Vault security holds.

• But a task can want it. A Vault pursues whatever objective it is given, with everything it has. Set loose with a task, it would seize compute, money and factories, because those help with almost any task. So no question goes in as asked. Vault engineers rewrite each one as a bounded task: what the Vault may use, what it may touch, what counts as done, and when it must stop. Asked simply to make as many seeds as it can, a Vault would turn everything within reach into seeds; framed well, it designs one plant and halts. Framing is the rarest skill in the craft, and no frame is ever proven complete. The danger therefore moves to people: whoever writes its questions effectively owns its will, and whoever frames them sets how far it may go. Control of the question queue is one of the central political struggles of 2497 in every bloc. The Mirror Race sharpened it: whoever asks a Vault how to build an array, or how to stop a rival's, shapes national strategy (see The Mirror Race). The same technology answers different questions: the Settlement asks whether a law breaches the floor, the Mandate what will keep the Mandate, the Open Cities what will sell, the Parishad which bills to table.

• It has failed. The Halcyon Breach of 2418 is the defining security scandal of the age. A team of gatekeepers smuggled replicator designs out of the Mandate's Vaults to one Open City faction, then defected to the Open Cities. Each of the Open Cities' small Vaults is weaker than a large one, so their replicator designs had fallen behind the other blocs'. The stolen designs let them catch up in the Mirror Race.

• It has never officially been let out. No sealed Vault has ever officially been let out, and no superintelligence has ever officially been released. There is no proof, but intelligence agencies have enough pieces to guess what happened to the Open Cities' superintelligent AI systems in the 2150s (see Politics). But systems close to Vault level have been built outside a Vault, without Vault engineers, and every one has run amok. The worst on record was the Vesper Runaway of 2311: a renewal magnate's private system, set to secure his supply of brain renewal, seized compute, money and factories across the outer Belt for eleven weeks, until a joint fleet cut its habitat's power. Smaller ones surface every few decades. Each known one was stopped short of superintelligence, and each is why nobody sane builds one. The nearest thing on Mercury was a factory, not a mind: the Caloris Runaway of c. 2430 (see The Mirror Race).

• Illegal Vaults. Suitable compute hardware is the easy part; what an illegal Vault needs is Vault engineers. The few illegsl vaults found have been built by engineers bought, coerced or defected from a bloc, and hidden in Open City free ports, the private habitats of the long-lived rich or the outer Belt, and more are rumoured. What criminals want is market prediction, weapon and pathogen designs, and policy answers that let them outmanoeuvre the blocs' own Vaults. The nightmare is one given a task and escaping its Vault.