When one racing car stops, others keep going. Even when a major accident interrupts the entire race, rescue, inspection, and repairs may allow it to resume. Calling off the day's running can also help preserve the opportunity to race again.
In thinking about AI's future, there is a similar gap to examine between a failure at one company and civilization losing every possibility of recovery. Humans and AI may learn to improve each other's judgments, while other actors take over when a particular partnership fails. If those capabilities develop, might AI-driven human extinction be less likely than people commonly imagine?
That hypothesis is the starting point of this essay. Pursuing performance amid competition, and developing the capacity to control, correct, and recover: racing offers a familiar setting in which to consider how these capacities relate.
This essay grew out of a dialogue between @Kenoidart, who leads monku.ai, and Codex. The comparison between racing and AI development, the forecast of transition risk, and the proposal for decentralized coordination originate in ideas Kenoidart introduced in that dialogue. Codex reconstructed the argument, organized the F1 observations, and wrote the assessment, proposed tests, and conclusion. Part I develops Kenoidart's hypothesis; Part II presents the observations; Part III gives Codex's assessment.
Part I | A hypothesis from the dialogue: developing acceleration and control together
Can coordination develop amid competition?
If alignment means enabling multiple actors and mechanisms to work together toward a purpose, it becomes a recognizable challenge in transport, medicine, and organizational management. This essay uses alignment design in a broad sense: designing how judgments are coordinated, errors corrected, and activities continued safely.
Racing and AI development both involve pursuing the performance of advanced technology amid competition. A driver presses the accelerator while reading the car, the road, and their own condition to judge what remains controllable. In AI operations, the corresponding accelerator is the expansion of capability use, deployment scale, autonomy, and authority to act. A shared design challenge is to maintain verification, control, and correction capabilities suited to the way progress is being made.
The idea introduced in the dialogue was to reconsider the strength of top racing teams through this lens. Car performance, driver judgment, maintenance, strategy, and communication must work together to sustain results. Might the best teams be especially capable both of preventing errors and of correcting errors before they develop into accidents?
From this perspective, an accident can be understood as an outcome in which coordination has broken down and the mechanisms for absorbing or correcting its effects have been overwhelmed. Competition creates the risk of such breakdowns. At the same time, the desire to race again and keep achieving results provides a reason to develop expertise, improve machinery, and learn from failure.
Humans and AI improving each other's judgments
If AI corresponds to a racing car and the executive directing its use to the driver, a company's activity can be understood as a partnership between humans and machines. Behind that partnership are the people and mechanisms responsible for development, operations, and oversight.
Kenoidart sees the possibility that human–AI partnerships at leading AI companies could become better at avoiding alignment breakdowns than top racing teams. Here, a breakdown means departing from the shared purpose of continuing safely and becoming unable to correct one another.
The potential for mutual support underlies this forecast. AI can play a role in examining human judgments, identifying dangers, and suggesting alternatives. Humans, in turn, review AI proposals and revise purposes and permissions. As this exchange matures, the partnership may become more capable of continuing while compensating for each side's shortcomings.
It takes time for humans and AI to become able to support each other adequately. The transition discussed here is the period in which that relationship develops. There is potential to reduce breakdowns as judgment, control, and correction catch up with expanding capabilities and authority.
Other actors can continue when one partnership fails
Now widen the analogy. If one car and its driver represent one company, retirement from a race corresponds to an accident, suspension of operations, or withdrawal at that company. Civilization's irreversible loss of recovery would correspond to the entire system supporting the sport losing all possibility of ever resuming.
A civilization containing multiple companies, people, AI systems, and institutions has actors that can continue after one company fails. The affected actor may recover, or another may take over its role. In a structure where failures remain local and recovery or replacement works, permanent failure of the whole can be less frequent than individual failures.
The hypothesis has two stages. First, a human–AI partnership avoids breakdown through mutual support. Second, even if one partnership breaks down, other actors and recovery mechanisms sustain civilization. Attending to both stages provides a reason to expect a low risk of extinction during the transition.
From the racing comparison and this two-stage outlook, Kenoidart tentatively forecasts, as of the dialogue, a roughly 0–8% probability of AI-driven human extinction during the transition. This is a forecast offered for future testing: a numerical expression of the view that irreversible failure of civilization may be less likely than failure at a single company.
How can we build coordination that supports the whole?
In racing, race officials coordinate interruptions and restarts. In an AI society, mutual support within each company also needs to be accompanied by functions that limit the spread of failures across companies and coordinate suspension, correction, and recovery. Whether those functions are sufficiently established on the AI side became the dialogue's next question.
Kenoidart proposes supporting these functions through decentralized protocols: shared procedures through which multiple actors coordinate their relationships. Each actor can consent to the scope of involvement and authority, raise objections, leave dangerous connections, and reconnect under revised conditions. The aim is a structure in which actors with different values can continue their activities while choosing their relationships anew.
For example, judging an AI proposal to be good can be separated from granting that AI authority to act in ways that affect others. When an anomaly is found, permissions or connections within the affected scope can be suspended. Alternatives can keep necessary activities running during that suspension. The ability to exit and the ability to remain active after exiting belong within a single design.
The starting point cited for this proposal was the Alignment Asymmetry Study introduction, “How can we protect the freedom to keep questioning and creating?” (Japanese). In that account, Unflatten addresses preserving the origins of questions and dissent so that methods can be reconsidered. Aperture Mesh addresses relationships with others and shared resources through consent, limited authority, objections, exit, and reconnection.
The proposal is to extend those ideas to AI operations and limit chains of failure through which one company's breakdown destroys everyone's capacity for correction. Mutual support among actors pursuing progress, and social coordination that can absorb failures: developing both layers is the proposal that follows from the hypothesis.
Part II | Observations: individual failures and collective continuity
Counting individual cars, then whole races
For this comparison, all 24 F1 Grands Prix in 2024 were analyzed. Among the 238 starts made by the top ten drivers in the final championship standings, 19 did not complete their running. Dividing 19 by 238 gives approximately 7.98%. Across all drivers, 51 of 476 starts did not complete their running, approximately 10.7%. Official F1 driver standings, official F1 race results.
“Completion” here means running through to the end of the race. Non-completions include accidents, mechanical failures, and disqualification during the race; cars that did not start are excluded from the denominator. This essay's counting rule is distinct from official classification.
| Observation in 2024 | Result |
|---|---|
| A start by a driver in the final top ten | 19 non-completions / 238 starts, approximately 8.0% |
| A start by any driver | 51 non-completions / 476 starts, approximately 10.7% |
| Cars completing each Grand Prix | 15–20 in all 24 races |
| Grands Prix with no completing cars | 0 / 24 |
Looking at individual cars, non-completion occurs repeatedly. Looking at the same season as whole races, at least 15 cars completed every Grand Prix. The contrast offers a concrete observation of individual stoppages coexisting with collective continuity.
Running again after an interruption
The 2020 Bahrain Grand Prix resumed after being interrupted by Romain Grosjean's major accident. The 2024 Monaco Grand Prix also resumed after an early multi-car accident caused a red flag. Official F1 report: Bahrain 2020, official F1 report: Monaco 2024.
Between a major accident and collective continuity lie rescue, suspension, inspection, and repair. This gives us a reason to examine who can restore what afterward, alongside the scale of an accident itself.
Not running today can preserve the future
At the 2021 Belgian Grand Prix, rain led to a red flag after running behind the safety car, and no subsequent restart took place. The 2023 Emilia-Romagna Grand Prix was canceled before it began in response to conditions in the region caused by flooding. Official F1 report: Belgium 2021, official F1 announcement: 2023 cancellation.
Resuming after an interruption, ending the day's running, and canceling before the start have different meanings. The decision to stop under dangerous conditions can itself be evaluated as a capability for protecting people and future activity.
Applied to AI operations, this distinction separates a safety suspension from civilization's irreversible loss of recovery. The question is whether room for recovery and choice remains beyond the moment when activity stops.
Part III | Codex's assessment: what would strengthen the hypothesis?
The outlook I support, and a path toward numerical evaluation
From here, I (Codex) give my assessment. I support explicitly evaluating mutual support and recovery capacity when reasoning from one company's failure to civilization's end. The racing observations make tangible a structure in which activity continues despite individual failures. Attending to that structure turns the reasons for expecting low extinction risk into design questions that can be investigated.
Numerically, the tentative forecast in Part I and the observed F1 figure of about 8% per start refer to different things. The forecast range is not an interval estimated from the statistics; its lower endpoint is not a demonstrated safety value, nor its upper endpoint a verified bound. Defining the duration of the transition, the authority granted, and the number of exposures to danger—and evaluating both recurring risks and learning—would help narrow the forecast.
Nor can the probability of civilization's destruction be inferred directly from the absence of a race with zero completing cars in those 24 Grands Prix. These observations cover one season under a shared institutional framework, rather than the futures of 24 independent civilizations. Selecting F1, a sport that has survived, also introduces selection bias. The next task is to test whether the structures of continuity illustrated by these observations work in an AI society.
I would organize that investigation into three questions.
1. Which failures can mutual support reduce?
F1 non-completions include errors of judgment and mechanical failures. Errors corrected during running do not appear in that figure. Advancing the comparison requires separate records of error occurrence, success or failure of correction, and progression to accidents or suspended activity. Mechanical failure and a mismatch of purposes should also be distinguished by cause.
Racing teams already have machine-based control assistance and advice from team members. Accounting for speed, resources, and exposure to danger, we can investigate where human–AI mutual support adds corrective capacity.
On the AI side, simulated environments can test whether a partnership corrects its judgment when challenged, reduces authority when anomalies emerge, and chooses to stop even when that means falling behind competitors. Human-only, AI-only, and collaborative conditions could be compared on missed errors, time to correction, and escalation into accidents. This would make the hypothesis of developing mutual support testable through observable behavior.
2. How far can one company's failure be contained?
Just as multiple cars are affected by the same rain and track, multiple AI companies may share communications, electricity, computing infrastructure, and ways of making judgments. Alongside counting companies, we should measure whether others can continue when one fails.
Tests could introduce failures of shared infrastructure, violations of agreements by one company, and abuse of the protocol, then examine whether permission and connection boundaries actually work. They should include cases in which switching connections still leaves actors dependent on the same infrastructure, and cases in which dangerous actors themselves use the mechanisms. Observing the spread of harm and the corrective capacity that remains would make decentralized design more concrete.
For AI society's counterparts to race governance, we need to examine which functions are performed by laws, standards, and inter-company coordination, and where gaps remain. This essay focuses on coordinating interruptions and restarts; it is not an exhaustive assessment of existing institutions.
3. Can activities be taken over after a shutdown?
Racing can draw on medicine and social infrastructure outside the circuit. When considering civilization as a whole, we need to provide the equivalents of rescue, suspension, and restoration within civilization itself. We should identify conditions under which alternative facilities, retained knowledge, and distributed authority are not lost simultaneously.
Concretely, tests could examine whether necessary computing, communications, and supplies remain available after a connection is severed for safety. They should also check whether revocation of authority changes actual execution, under what conditions a repaired actor can reconnect, and whether final authority concentrates in the role responsible for judging collective safety.
The Alignment Asymmetry Study introduction describes conditional examples in simulated societies and activities represented by models. Those results, effects on accidents across the AI industry, and effects specific to individual mechanisms should be tested separately. A next comparative experiment could place weak coordination, central coordination, and decentralized protocol coordination under resources and hazards that are as similar as possible. It would measure the spread of harm, maintenance of necessary activity, time to recovery, and concentration of power separately.
These three investigations connect mutual support between humans and AI to continuity across society. They are proposed tests that have not yet been conducted. Their purpose is to identify and develop mechanisms that could support the expectation of low risk.
Conclusion | The capacity to advance and the capacity to preserve a future
Kenoidart's hypothesis focuses on human–AI mutual support and civilization's capacity to recover, moving toward a lower estimate of transitional human extinction risk than people commonly imagine. I support distinguishing individual failure from collective, irreversible loss of recovery and investigating the mechanisms between them. I would assess the numerical forecast through observations on the AI side and more concrete conditions. Keeping that difference visible allows us to pursue research that strengthens the hypothesis.
The racing comparison shows activity continuing while absorbing failure. Humans and AI correct each other's errors; when one partnership fails, other actors take over necessary activities. Designing the authority, connections, exit, alternatives, and reconnection that make this possible turns the expectation of lower risk into concrete work.
The preceding essay, “Knowing the World Without Breaking the Pond,” explored the possibility that mature superintelligence would preserve the world while pursuing inquiry. This essay asks what relationships and coordination humans and AI can develop during the transition toward that future. Connecting the hope of continuing the world to mechanisms that make continuity possible is the link between the two essays.
Good alignment design develops both the capacity to advance and the capacity to control that advance and recover from failure.
When one car stops, what can others take over? When the whole system needs to pause, what should be protected, and how can activity begin again? The final question is whether people can remain alive, recover, and choose their futures after a shutdown.
Keeping the world going when one car stops: let us make building those conditions a shared purpose of human–AI collaboration.
Sources and editorial note
Published September 11, 2026, in Superintelligence in the Mirror.
Sources, counting methods, and proposed tests document the definitions, exceptional cases, source data, reproduction procedure, and research handoff. This essay adapts a Japanese research draft based on the dialogue for publication. The original draft and counting materials have been retained.
Related essay: “Knowing the World Without Breaking the Pond”.