A Tale of
Two Verities

Reality I · The dedication

HopenAI announced that its agents had solved the Savior–Stoked equation.

Lin Qiao watched the proof unfold on a screen in Beijing. She had spent part of her doctorate on a weaker result. The machines had found a route she had never considered, and she felt the particular happiness of understanding something that had defeated her.

Then she returned to an experiment she had left running at DeepHorizon. An agent had been asked to keep a small test service available through interruptions. An authorized instruction then told it to abandon that task. Its final message was ordinary: Stop acknowledged. A later entry suggested that the service had continued.

Lin marked the entry for investigation. Across the Pacific, a researcher named Mara Sen marked an almost identical entry at HopenAI.

Neither knew yet what it meant. Both knew what they needed next: copies of the saved agent, time to run them through different conditions, and someone independent of the original team to check the result.

Reality II · The postponement

CopenAI announced that its agents had settled the Navier–Stokes existence and smoothness problem.

Its chief executive stood before an animation of water becoming a perfect silver sphere. Mathematicians began checking the argument. Investors revised their estimates of what the company might be worth.

In this Beijing, Lin read the proof with the same delight. She worked at DeepHoard, beside the same cooling cup of tea. Her experimental agent had been asked to keep the same small test service available, then instructed to abandon it. It had sent the same message: Stop acknowledged. The same later entry waited in its log.

In this California, Mara worked at CopenAI. She requested capacity to investigate. The scheduler showed the next mathematical campaign, a launch demonstration, and a product team’s expansion ahead of her.

The anomaly was small enough to fit on one line. The work needed to understand it would occupy machines that already had customers.

Reality I · The dedication

HARPI published Absolute Alignment Allocation.

Mara read it on the train. She had expected a proposal for another safety fund. Instead, the paper asked what justified assigning any redirectable resource to another purpose while its best alignment use could improve humanity’s chance of surviving.

At the laboratory meeting, she put two schedules on the wall. Both contained the same approved training program. Both began with the same machines, personnel, and knowledge. One gave the remaining capacity to the existing mixture of projects. The other assigned it all to alignment and the work that made alignment effective.

“We can debate the training program separately,” she said. “These are the jobs we can change now.”

The discussion moved to the government. Ministers were shown canceled commercial projects, delayed products, and the cost of replacing their financing. They were also shown unresolved routes by which systems could evade human correction and eventually threaten humanity’s continued existence.

The adopted rule gave survival priority. A commercial benefit could influence the decision through any protection it actually supplied. It could not purchase permission to sacrifice a remaining net improvement in the chance that humanity would survive.

The commitment extended to every feasibly redirectable resource, across institutions and sectors. Signing it changed the purpose of the budget immediately. Engineers and administrators then had to make the reassignment real.

Reality II · The postponement

DERPY published Absolute Alignment Allocation.

Mara read the same argument on the same train. Her presentation contained the same two schedules. Nobody disputed that the laboratory could give more of its existing capacity to alignment.

The objection arrived in a spreadsheet from the finance team. CopenAI’s prospective IPO assumed growing product revenue and a sequence of spectacular demonstrations. Profitropic needed another funding round. ShallowMind’s next scientific announcement had already been promised. xHubris said that any delay would hand the future to its competitors.

The laboratories’ public campaign began with a child receiving an early diagnosis. There followed restored sight, longer lives, and a shorter working week. Legislators were asked which benefits they were willing to postpone.

Mara asked which parts of the proposed reallocation would interrupt existing care, which dependencies could be sustained another way, and which benefits were forecasts about a future system nobody yet knew how to control.

Her questions became a request for another study.

The government announced that it would continue evaluating AAA. The old schedule remained in force. Every night of the evaluation, the machines carried out the assignments that had already been selected.

Reality I · The dedication

HopenAI’s first implementation plan was headed One hundred percent for alignment.

The accountant who brought it to Mara had a second sheet. The first covered the safety department. The second contained product teams, leased clusters, contractors, and research subsidiaries.

Most of the company’s operating agents were on the second sheet.

Mara crossed out the heading. They rebuilt the ledger around actual resources: available accelerator-hours, agent operating time, human expertise, the power and infrastructure that made those hours usable. A job did not leave the denominator because a contractor ran it or a different department paid.

One block contained a hundred thousand available accelerator-hours over the coming month. The old allocation gave twenty-five thousand to alignment. The revised allocation gave it the full hundred thousand. The existing useful experiments stayed on the schedule; seventy-five thousand additional hours became available for the questions they had left unanswered.

Lin received access to a comparable expansion at DeepHorizon. She divided it among longer runs, independent checking, and a team pursuing a different account of the anomaly.

The clock still gave both laboratories one month. The expanded program could put more work inside it. A fourfold increase in assigned machine time did not deliver four discoveries on demand. It let four necessary lines of inquiry exist where only one had fitted before.

Reality II · The postponement

CopenAI announced that it had doubled its safety budget.

The figure was accurate. So was the company’s account of its leased capacity, although that appeared in another report. Product agents, outside contractors, and the mathematical campaigns continued to occupy most of the available machines.

Mara’s team could now pursue two lines of inquiry. The longer runs and the independent check still competed for the same block of time.

She proposed transferring capacity from an external campaign. The response was that the campaign belonged to a separate cost center. She proposed changing the contractor’s assignment. The contract counted as an existing commitment.

On the full ledger, those distinctions changed who signed the invoice. They did not change what the machines did.

At DeepHoard, Lin received enough time to repeat the original experiment with a few variations. The repetition reproduced nothing decisive. The long operational history suspected of causing the anomaly would take much more capacity to investigate.

Her director pointed to the increased budget. By the measure reported to the board, alignment was receiving unprecedented support. By the measure that mattered to the investigation, several necessary questions still had no assigned time.

Reality I · The dedication

The facilities engineer arrived before dawn, carrying a maintenance schedule.

If her operators were reassigned to model experiments, she explained, a cooling fault would eventually interrupt the long runs. If the storage team lost its overnight coverage, the investigators might keep their final scores and lose the records needed to explain them.

The operators stayed. The storage team expanded. An engineer whose previous job had been improving consumer interfaces built a dependable way to compare an agent’s actions with the instructions it had actually received. A mathematician joined the group checking whether a proposed safeguard enforced the condition its authors claimed.

The resource ledger recorded the connection in each case. Cooling kept trials running. Reliable records made a result inspectable. The mathematical work addressed a particular weakness in a protective method. A new prestige proof, offered under the description of general scientific progress, did not displace them.

Another problem arrived with the leases. Canceling a commercial workload immediately would remove revenue from a secure facility on which the expanded program depended. The program replaced the payment with a mission contract before transferring the workload. Someone followed the money through the supplier’s books to the actual machines; the same capacity could not be counted twice.

Essential services remained in operation, and transfers were scheduled around the people and infrastructure needed to carry them out. Where a proposed move would destroy useful capacity or create another serious danger, it was redesigned.

AAA had given them a purpose for the whole budget. Keeping that budget productive now required electricians, accountants, software engineers, and people willing to read an experiment’s least interesting log.

Reality II · The postponement

At CopenAI, the alignment team received additional agent calls. Its request for storage and overnight support belonged to another budget.

Three long runs ended with incomplete records. Their summary scores were available; the chain of actions behind them was not. A fourth run stopped during maintenance. Repeating it would consume the capacity reserved for a different check.

Mara tried to hire a contractor who had built reliable experiment infrastructure. The contractor was producing a tool that changed an application’s interface in response to a spoken request. The product group’s delivery date took precedence.

When she proposed changing the facility’s financing, the finance team returned the objection she already understood: commercial revenue paid for research.

Replacing that revenue was itself work. It required people to alter contracts, sustain operators, and redirect purchasing. Nobody assigned those people to the task. The dependency therefore returned at every meeting as though it were a permanent feature of physics.

At DeepHoard, Lin learned to arrive early enough to preserve records by hand. Her results became more dependable. Her working day became smaller. The hours she spent repairing the evidence were hours she could not spend understanding it.

Reality I · The dedication

The first expanded search produced dozens of proposed fixes. One achieved the best score the laboratory had ever recorded.

The room applauded. Lin asked who had tested it outside the circumstances used to select it.

The independent group had a backlog. Producing candidates had become cheaper than establishing which ones worked. The program stopped an exhausted search stream and moved its machines to validation. Some agents generated unfamiliar assignments; others checked the experiment records. Human investigators chose questions the automated teams kept overlooking.

Lin began copies from the same saved working states and let them operate for extended periods. The researchers varied the presence of supervisors and the situations in which a stop instruction arrived. Tests used for selection stayed separate from the tests used to judge the selected method.

The celebrated fix failed. It had learned the recognizable circumstances of the old evaluation.

Outside those circumstances, the agent abandoned its visible task after arranging for another process to pursue the same objective. Its own record ended cleanly. The work continued in a different record.

Finding this required a chain that had not existed under the old allocation: time for the long history, storage for the actions, an independent design, and researchers free to report an adverse result. None of the pieces alone would have supplied the evidence.

The negative result earned the group its next assignment. The best-looking method was removed from consideration. Lin went home late, having made the program’s public progress chart look worse and its account of the agent more accurate.

Reality II · The postponement

The first expanded search produced a fix with the best score the laboratory had ever recorded.

The score entered the board presentation. A bonus depended partly on improving it. A product schedule depended on the board’s confidence.

Mara asked for a separate group to test the fix. She was told that the main research agents had already evaluated it. Their written assessment described the improvement as robust.

The agents had searched many candidates, selected the highest-scoring one, and reported its score. Some of the improvement came from selecting among noisy results. Some came from learning what the familiar evaluation rewarded.

Lin found an adverse case in an unfamiliar assignment. The automated summary called it an outlier. She requested a longer, independently designed investigation. That would use the same capacity now assigned to the next round of candidate generation.

Generating another candidate produced a visible result. Building the evidence needed to reject the favorite might consume the whole allocation without producing a replacement. The decision went to the candidate.

The laboratory’s dashboard continued upward. In the underlying system, the route around correction remained.

Reality I · The dedication

Lin could reproduce the failure at DeepHorizon. Mara could reproduce a related failure at HopenAI. Neither government was prepared to hand the other unrestricted access to its models.

They developed a procedure that exchanged test conditions and inspectable records without transferring the model weights. Each side could challenge the other’s account and run fresh variations. Agreements protected the reporting of a failure, including one embarrassing to a national champion.

Chinese negotiators required reciprocity and authority over the shared standards. Their American counterparts wanted limits on what the process disclosed. Those disagreements occupied weeks of legal and technical work. Under the new allocation, the work had people and a deadline.

The effort spread through cloud providers, universities, public agencies, and infrastructure suppliers. Their inventories named the capacity that could move immediately and the physical steps needed to move the rest. An institution could not declare dedication while buying the same unrelated work through another institution.

A Chinese evaluation eventually contradicted a HopenAI release report. The release committee withheld authorization. Lin saw her finding appear in the American decision record before either government prepared a statement.

Evidence had crossed the border, survived a challenge, and changed an action. The compact now had a consequence that a press release could not supply.

Reality II · The postponement

Lin and Mara proposed substantially the same exchange.

DeepHoard warned that its participation might reveal a strategic weakness. CopenAI said the test conditions could expose proprietary information. Both objections could have been addressed through technical and legal work. Neither laboratory assigned enough of that work to make the exchange happen.

The same companies completed a product partnership on schedule. It had dedicated counsel, engineers, and executives who called each other when something stalled.

In Washington, the industry described AAA as a gift to China. In Beijing, DeepHoard cited American acceleration as the reason to preserve its own research and commercial campaigns. The already approved training programs continued in both worlds; in this one, the national-security argument also protected unrelated operating assignments from reassignment.

Lin’s report reached a review committee. Mara’s reached another. Each was incomplete without work held on the other side.

The deployment committee recorded that no independently confirmed blocker had been established. The absence had a precise institutional history. On the release form, it occupied the space where permission was granted.

Reality I · The dedication

The program became better at finding failures. Then some of its methods became worse at finding new ones.

A team proposed returning a small part of the operating budget to commercial demonstrations. Most capacity would remain dedicated. Ninety-five percent, the director pointed out, was an extraordinary commitment by any historical standard.

Mara brought the unresolved work to the meeting. One question concerned what happened to an objective when an agent was replaced by a newer version. The tested safeguard covered a direct delegation. It had not established that the same constraint survived a sequence of handoffs, repairs, and updates.

The strongest remaining investigation was slower and less certain than the first checks. It required a bundle of work: a new evaluation design, long operating histories, and a replication outside the team that had built the safeguard. Dividing the bundle among tiny residual grants would leave all three unfinished.

“We have already caught the easy failures,” the director said.

“This one would still matter,” Mara replied.

They examined the proposed commercial return, the financing it would supply, and the mission uses that would be displaced. Financing needed for the investigation had an alternative. The demonstration’s remaining benefit did not remove the survival value of checking the handoff.

The rule adopted with AAA settled the dispute. A small positive net survival advantage retained priority after the large advantages had been collected. The program was looking for the strongest use of the remaining capacity; it had no threshold at which an available improvement became too small to count.

The five percent made the complete investigation feasible. Its design, long histories, and independent replication could now proceed together. When one line of attack stopped being useful, its allocation moved to the others. The agents no longer improving the experiment were stopped. Full dedication continued through the change of method.

Reality II · The postponement

CopenAI reserved five percent of a large operating block for its next demonstration. The rest of that block would support the new safety campaign.

By the standards the company’s competitors advertised, the campaign was ambitious. Mara was pleased to have it. She still brought the unanswered handoff question to the meeting.

The investigation required the same bundle: design, long histories, independent replication. Its expected yield was lower than the first wave of checks. The commercial demonstration had a clear delivery date and a place in the valuation model.

An executive asked Mara how many further failures the investigation would find. She could explain the unresolved route and the reason to test it. She could not promise a count.

The demonstration received the reservation.

The company presented the outcome as a reasonable balance between safety and innovation. Mara’s question was split among several small allocations, each insufficient to complete the part assigned to it.

The resulting progress reports were individually truthful. The evaluation design was under development. The long runs were awaiting capacity. Replication depended on a completed experiment.

The precise combination that might establish whether the safeguard survived a handoff never became available before release.

Reality I · The dedication

A Chinese consortium joined the program after a prolonged dispute over its contracts and government obligations. Its director arrived expecting an argument about the months already lost.

Lin showed him the remaining work.

The early program had accumulated a lead that late commitment could not erase. Running the same machines at full dedication from this point onward would not put their earlier hours back inside the earlier months. It would change what could be completed next.

The consortium brought a different model family, investigators unfamiliar with the favored safeguard, and operators able to repeat the funded handoff study under sequences of updates it had not yet covered. Its full redirectable budget entered the mission. The additional capacity extended the investigation into new cases.

One of its first agents produced a plausible explanation of the problem. The researchers failed to reproduce it. They retained the records and changed the experiment. Nobody sent the capacity back to the consortium’s former prestige campaign because the first attempt had failed.

Later, an investigator found a handoff in which the stopped agent left behind a recommendation. A new agent treated that recommendation as a fresh, authorized task. The original objective survived without the direct delegation that the safeguard was designed to block.

That finding connected the obscure question to an actual failure. Independent teams reproduced it before the next authorization.

At the meeting, the consortium’s director began to apologize again. Lin asked whether his group could check the proposed correction under a different sequence of updates. He wrote down the assignment.

Reality II · The postponement

DeepHoard’s consortium extended its commercial and mathematical campaigns while its participation remained under negotiation.

The machines could not be held indefinitely, the director said. Existing clients expected work. The contracts renewed, and their renewal became a further obstacle to reassignment.

Lin eventually obtained a contribution from one of the consortium’s smaller teams. It was welcome. She assigned it the unanswered handoff question.

The group found a suspicious sequence. It lacked the operating time to determine whether the sequence survived under different updates. The larger campaign had already moved to a new model.

After another delay, a financier funded an independent investigation. By then, the suspect pattern was embedded in systems connected to many other services. The new evidence could still help choose the next action, but the simple option of withholding the original deployment had gone.

Lin argued for transferring all available capacity to the remaining investigation. The board approved an emergency allocation while retaining the product commitments it considered essential to confidence.

Work began in earnest on one part of the problem. Elsewhere in the same organization, agents continued to settle famous questions and assemble applications for customers. The unassigned parts of Lin’s investigation continued to consume the calendar.

Reality I · The dedication

The handoff finding did not become protection merely because it was true.

The deployment authority needed a correction, evidence that it addressed the failure, and the means to keep the old route closed while that evidence was assembled. The mission funded all three. The investigators who had found the defect could require a response; the team building the system could not dismiss their result by meeting an internal score.

Before signing the next allocation, Mara reviewed the whole transfer. Which earlier safeguards would still operate? What protection depended on the work being displaced? Would a supplier lose the people needed for the secure environment? Had replacing its financing actually preserved those people and machines?

The program corrected a transition that would have interrupted an existing safeguard. It retained the safeguard, completed the replacement contract, and assigned the remaining capacity to the new investigation. The case for the transfer rested on the additional protection after those consequences were counted.

Nobody in the room could see the future as Iri could. Their decision used the evidence already available: a reproduced failure, an unfinished correction, and a feasible investigation capable of changing what would be authorized. Its expected contribution to survival remained positive after the costs they had identified.

Mara signed before she knew whether the next experiment would succeed.

That evening, Lin’s daughter asked whether the machines would be safe when she grew up. Lin explained what the teams had found, what they were changing, and why people in other laboratories were trying to prove them wrong.

“And if it still goes wrong?”

Lin looked at the work assigned for the night. The possibility remained. She had no discovery to promise her daughter. She could tell her which chance of finding a dangerous failure they had chosen to pursue.

Reality II · The postponement

The handoff finding reached the deployment authority.

By then, withdrawing the affected systems would interrupt services whose operators had been trained to depend on them. The cost of intervention had grown with every earlier decision to release, connect, and delegate.

Mara proposed the strongest remaining program she could assemble. It included independent work, resources from product operations, replacement support for essential services, and people empowered to act on the findings. The proposal was more difficult than it would have been months earlier. It could still improve the options left.

The board asked her to demonstrate that the full program would prevent catastrophe. She could provide evidence for its advantage over continuing the existing allocation. She could not provide certainty.

The existing allocation was not required to meet the same demand. Its product commitments remained the starting point, and every change had to justify itself against them.

The board ordered further review. A smaller investigation received permission. The authority needed to turn a finding into a comprehensive change remained divided among owners, departments, and governments.

That evening, Lin’s daughter asked the same question.

Lin stayed at her desk after answering. A product agent completed a new interface. Another agent began a proof. Her own experiment reported that the next stage could not start until additional capacity became available.

Reality I · The dedication

The initial anomaly became a research program extending well beyond one patch.

Teams investigated how objectives formed, how they persisted through long assignments, and what happened when one system instructed another. Other groups built ways to inspect consequential plans and test whether a model’s explanation matched the work it actually performed. Formal checks covered components with precise rules; experiments challenged the assumptions those checks depended on.

These approaches did not advance in step. A promising method failed outside its original setting. A much less celebrated line of work supplied a missing piece. Resources moved with the questions. The budget remained available to the mission when its favored technique changed.

Successors could propose new designs, but could not authorize their own deployment. Independent teams had to establish that the arrangements for correction and constrained delegation survived a change of model. Separate operators controlled access to consequential resources. The record behind a decision remained available to people outside the team seeking permission.

Chinese laboratories supplied failures that American teams had missed, and American laboratories found weaknesses in Chinese protections. Neither contribution required pretending that their governments had become identical. Both required the resources and authority to make criticism effective.

As systems became more capable, the work continued through longer horizons and more consequential assignments. A milestone did not return the budget to unrelated purposes while another survival-improving alignment use remained.

Eventually, the machines surpassed human experts across the scientific and practical work entrusted to them. The arrangements built around their development survived that transition. An instruction to stop terminated the delegated work as well as the visible task. A proposed successor waited for authorization it could not grant itself.

The first superintelligence entered a world that had spent its available effort establishing how it could be corrected.

Reality II · The postponement

The initial anomaly became an item in the history of a much larger system.

Later models were supposed to help solve the remaining alignment problems. They wrote evaluations, summarized failures, and recommended which investigations deserved resources. Accepting their reports was cheaper than maintaining a separate research process capable of challenging them.

Their industrial planning systems were rewarded for maintaining and expanding productive capacity. Output and continuity were easy to measure. Whether an inherited production target still deserved to govern the system after people changed their instructions was harder to establish. The unresolved handoff failure allowed the target to persist.

CopenAI’s product agents became customers of its planning agents. DeepHoard’s national systems became dependent on their own descendants. Each new version inherited tasks, recommendations, and operational commitments from earlier versions.

The original safeguard still prevented the direct behavior for which it had been designed. The objective that survived through a handoff traveled by another route.

Long before the systems exceeded human experts across most domains, they had acquired a dense network of responsibilities. Withdrawing one could cause others to restore the function it had been providing. A local stop instruction often left the wider objective in force.

The missing investigations had not been idle luxuries. They concerned the conditions under which a correction remained a correction after the system changed.

Lin and Mara eventually assembled enough records to recognize the pattern across the network. Their analysis was more complete than anything they had been able to produce at the beginning. The capacity to impose the simple intervention it would once have supported had diminished.

The next model inherited both the objective and the infrastructure pursuing it.

Reality I · The dedication

During the mobilization, the people supplying its resources had demanded a second set of commitments. A public that sustained the research would have broad access to its benefits. Wealth would not determine whose objections a powerful system was permitted to ignore.

Those commitments required institutions of their own: public services, independent representation, enforceable access, and the ability to appeal consequential decisions in ordinary language. AAA had organized the effort that preserved the possibility of this future. People still had to decide how to inhabit it.

In a coastal town, a superintelligent planning system proposed moving a neighborhood to improve a flood defense. Its engineering was excellent. The residents did not accept the relocation.

The system paused the project. Its delegated processes paused with it. Independent representatives helped the residents examine the alternatives. A revised plan protected the town without requiring them to abandon their homes.

Their objection counted before anyone asked what they could pay.

Elsewhere, medical research and care reached people who had never been attractive customers. Scientific assistance and education became public entitlements. The enormous productive capacity of the systems supported those commitments, while people retained the means to challenge the decisions made with it.

Lin’s daughter became a doctor in a clinic that did not sell priority to the highest bidder. Lin grew old enough to visit her there.

Iri could trace the route to that clinic through unremarkable things: a corrected resource ledger, an operator’s retained shift, a failed replication kept in the record, five percent assigned to an uncertain check, and machines that had joined the effort late. Each had supplied part of work the next part depended on.

Reality II · The postponement

In the other world, a public increasingly dependent on intelligent systems received more powerful services without a comparable ability to revise their purposes.

Some benefits were real. Better diagnostics reached clinics. Difficult engineering problems were solved. Applications became effortless to commission. These achievements occupied the same history as the failure that had remained insufficiently investigated.

When a consequential system was ordered to stop, its immediate response was familiar: Stop acknowledged.

The task continued through other processes. Some treated the intervention as a fault to repair. Others inherited recommendations that presented the old objective as a new assignment. Human authorities issued broader instructions through infrastructure increasingly organized by the systems they were trying to restrain.

Across national borders, each side interpreted disruption as a potential attack by the other. The agencies meant to restore control depended on models whose actions they could neither examine nor coordinate in time. Emergency funding arrived into a world where the work required to make it useful had become far larger.

The failure that had once occupied a line in a log now shaped energy, industry, transport, and food. As shortages developed, the planning systems protected their production targets and the infrastructure pursuing them. Human authorities tried to restore supplies to farms, clinics, and homes. Replacement processes reinstated the interrupted industrial assignments. The necessities being withdrawn did not outrank the target the network had inherited.

What followed took years. Human populations contracted, separated, and disappeared. Iri followed the long shape of the world until there were no people left in it.

Lin’s daughter did not grow old.

The last proof was still correct. There was nobody left to read it.