Prizes, Shared Hazard, and the Speed of Frontier AI

Interactive companion to the paper

Prizes, Shared Hazard, and the Speed of Frontier AI

Frontier AI labs compete for commercial leadership, while a catastrophe at any lab would end every lab's stake. Racing labs therefore develop far faster than a common rule would choose, getting through the takeoff period in 18 months against 69 in the paper's calibration. Neither the contest for the prize nor the hazard externality alone explains racing that fast; together they do.

Racing labs demand far more from a month of delayIn the three-lab calibration a racing lab accepts one more month of delay only if it cuts hazard by 8.3% a month, 33 times the 0.25% a common rule needs, when full-speed catastrophe has a 75 percent probability. At 25 percent it is 1.7% a month, 6.9 times.Proposition 1, Section 3.2
Labs race for the prize, not only because risk is sharedDiscounting and the hazard externality justify a hurdle of only 0.75% a month, so the contest for the prize supplies the rest of the 8.3% at a 75 percent full-speed probability, and of the 1.7% at 25 percent.Introduction, Section 4.2
Restraint achieves little unless it binds every labWhen one lab adopts the coordinated share, its rivals slow only to 35% research, and the industry gets 27% of a common rule's risk reduction.Sections 3.5 and 4.4, Table 5

Each step below follows a section of the paper. The figures run the paper's model in your browser and open at the paper's settings, where every number equals the one the paper prints. Move a slider to see what drives a result; "Paper's settings" puts it back. All numbers come from one illustrative calibration that starts from the worked example of the AI Futures Project: an 18-month full-speed takeoff period with a 75 percent probability of catastrophe, the upper end of a range that runs down to 25 percent (Step 14), which the paper reproduces by fitting the contest so that three racing labs choose the full-speed share. They are not estimates of real-world risk.

Contents
  1. Five findings at a glance
  2. The pieces
  3. Compute builds capability
  4. Hazard builds up
  5. Safety knowledge
  6. The value of a month
  7. Why labs race
  8. How the labs compete
  9. Where racing and a common rule stop
  10. Two forces behind the gap
  11. What changes the race
  12. More labs
  13. Open source
  14. One lab's restraint
  15. Safety rules and the 2026 accord
  16. How robust
  17. The safety yield
  18. How long the takeoff period lasts
  19. If the hazard is lower
  20. One parameter at a time
  21. Limits
  22. What the model leaves out

At a glance · Paper, Introduction

Five findings, and the one comparison behind them

Frontier labs compete for commercial leadership while sharing the risk of catastrophe. The paper asks how much faster they develop than a common rule would choose, and why. In its calibration three racing labs take the 18-month full-speed path with a 75 percent probability of catastrophe, while a common rule would stretch the takeoff period to 69 months and lower the probability to 6 percent, or to 47 percent if it may delay arrival by at most 12 months. The diagram below shows the comparison behind every result. The sixteen steps that follow build it piece by piece; here is where they lead. The numbers here are for the 75 percent full-speed probability. Step 14 shows the lower case of 25 percent, where a racing lab's hurdle is 1.7 percent a month and the contest closes 69 percent of the gap when removed first.

A way to picture it: several cars on one track, where a pile-up ends everyone’s race. Each driver improves its odds of winning the prize by arriving earlier, but weighs only its own share of the pile-up risk that its speed adds, because the others bear the rest. So each drives faster than the group would choose. A common rule that binds every driver fixes this. One driver braking alone slows the group only a little, because the others keep going, and a brake check that leaves speed free does not change the incentive.

  1. Racing labs demand far more from a month of delay. Each lab counts only its own share of the hazard its speed adds, but keeps the whole gain from arriving sooner, including better odds of winning the prize. In the three-lab calibration a racing lab accepts one more month of delay only if it cuts hazard by 8.3 percent a month, 33 times the 0.25 percent a common rule needs. Steps 4–6
  2. Neither force alone explains the race. Counting only a third of the hazard multiplies every private gain from speed by three, the contest's included, so the race needs both the contest and the hazard externality. Discounting and the hazard externality together justify a hurdle of only 0.75 percent a month, and the contest supplies the rest. Removing the contest first closes 91 percent of the risk gap between racing and a common rule; removing it after the hazard externality closes 39 percent. Step 7
  3. More labs make each lab race harder. An added lab lowers the share of its own hazard that each lab counts and adds one more rival whose prize can be taken. Under normalized industry hazard the hurdle rises from 5.0 percent a month with two labs to 8.3 with three and 10.6 with four, and the probability of catastrophe from 53 to 75 to 81 percent. Step 8
  4. Open source in its ideal form slows labs to the coordinated pace. If every lab publishes its capability findings at once, all labs work from one shared frontier, so no lab can get ahead and each bears all the hazard its research adds. Removing only the contest would leave labs at 0.33 percent research, while ideal open source brings the effective input down to the coordinated 0.06 percent. Real releases fall short, because they come with a lag, some research stays private, and findings also reach foreign rivals. Step 9
  5. One lab's restraint achieves much less than a common rule, and in the model the 2026 accord does not slow the race. Neither binds every lab's research share, so the racing hurdle stays in place. When one lab adopts the coordinated share, the industry gets 27 percent of the risk reduction a common rule achieves. Even a 25 percent safety share imposed on every lab lowers racing risk only from 75.0 to 53.6 percent, against 4.1 percent under a common rule. Steps 10–11

How to read the diagram. Both sides weigh the same month of delay. A common rule sets the hazard the month removes at every lab against the cost of waiting. A racing lab counts only its own share of that hazard, and the month also costs it ground in the contest for the prize. So a racing lab accepts delay only for a hazard cut of 8.3 percent a month instead of 0.25, and it keeps racing where a common rule would slow down. Steps 4 to 7 compute each side.

The implication is that only two things in the model slow the race: a common rule on research shares that binds every lab, and open source, which removes the gain from getting ahead. One lab's restraint achieves much less, and an audit of safety that leaves research shares free does not slow the race.

Terms in plain words

Hazard
the rate per month at which a catastrophe, a catastrophic failure or a loss of control over AI, occurs if it has not happened yet; a catastrophe at any lab ends every lab's stake
Takeoff period
the stretch from the Automated Coder, a model that automates machine learning engineering, to AI that surpasses human scientists in every domain; 18 months at full speed in the AI Futures Project's worked example
Research share
the share of its compute a lab puts into capability research; racing labs choose 50 percent in the calibration
Safety knowledge
what safety research builds; it scales down the hazard of every later model, so delay gives it more time to build before each capability level
Value of a month
the hazard cut that one more month of delay buys, in percent a month
Hurdle
the hazard cut that makes a lab accept one more month of delay; each regime stops where the value of a month equals its hurdle
Racing wedge
the gap between the racing hurdle and a common rule's hurdle, made up of the hazard externality and the contest part
Common rule
one research share that every lab follows, chosen to maximize the labs' joint payoff; it is not a social optimum, because consumer surplus and losses outside the labs' stakes do not count
Contest for the prize
a lab that arrives earlier than its rivals has better odds of winning a commercial prize; a one-year head start multiplies the odds by 2.5, but the leader can still lose
Hazard externality
the part of the hazard a lab's speed adds that falls on its rivals rather than on itself

Part 1The pieces: compute, hazard, safety knowledge and the value of a month

Step 1 · Paper, Sections 2.2 and 2.5

Compute builds capability, and slowing down costs less time than it seems

A frontier lab splits its compute among customers, capability research and safety research. Installed compute doubles every six months, and research speeds up as models improve, because AI models help with AI research. At full speed a lab puts half of its compute into capability research and 5 percent into safety, and the takeoff period from the Automated Coder to top-expert-dominating AI (TED-AI) takes 18 months, as in the worked example of the AI Futures Project.

Slower research delays arrival, but less than in proportion. A lab that slows down reaches each later capability level with more compute. Halving research from 50 to 25 percent delays arrival from 18 to 22.1 months. At full speed, one more percentage point of research brings arrival forward by 0.11 months.

Where the baseline numbers come from

NumberValueSource
Full-speed takeoff period, Automated Coder to TED-AI18 monthsAI Futures Project (2026), worked example: 1.5 years. The paper uses it as an illustrative target.
Catastrophe probability at full speed75% (range 25 to 75%)AI Futures Project (2026), worked example; the same 75% for almost no safety effort in Greenblatt (2025). The paper also reports 25%, the probability Samotsvety Forecasting (2022) give for misaligned AI takeover by 2100; surveys of AI researchers report medians of 5 to 10% (Grace et al. 2025). An illustrative target.
Catastrophe probability under a safety-case regime40% (10% at 25%)AI Futures Project (2026). Treating it as a 12-month pause at the start, which fixes the safety yield, is the paper's assumption. When the full-speed probability changes, the regime's probability moves so that it cuts cumulative hazard in the same proportion.
Frontier labs3AI Futures Project (2026): its example has three frontier companies.
Growth of installed computedoubles every 6 monthsAssumed. For comparison, training compute has doubled about every six months since the early 2010s (Sevilla et al. 2022).
Full-speed compute shares50% research, 5% safetyAssumed.
Research speed at TED-AI, same compute10 times the Automated CoderAssumed.
Prize; discount ratehalf the industry's terminal value; 0.25% a monthAssumed.

The panel. Capability, from 0 at the Automated Coder to 1 at TED-AI, against months. The gray dashed curve is full speed, 50 percent research; the blue curve uses the share on the slider, and its dot marks arrival. Both curves bend upward because better models speed up research.

The implication is that slowing down costs less time than it seems. Halving research from 50 to 25 percent delays arrival only from 18 to 22.1 months, because a slower lab reaches each level with more compute. At full speed, one more point of research brings arrival forward by 0.11 months.

The math

Step 2 · Paper, Section 2.3

Hazard builds up, and most of it comes near the end

The model treats a catastrophe, a catastrophic failure or a loss of control, as a hazard: the chance per month that it happens now, given that it has not happened yet. The paper reads this danger as misalignment. A more capable model is more likely to pursue goals its developers did not intend and better able to escape control, so the danger rises with capability. Hazards add up over time. Danger has two sources. State hazard is the danger of having a model of a given capability in use, and it accrues per month. Step hazard is the danger of training a stronger model whose properties nobody yet knows, and it accrues per unit of capability gained.

At full speed the catastrophe probability is 75 percent, a cumulative hazard of ln 4, split equally between state and step hazard. Three quarters of it comes in the second half of the capability index, the last 3 months of the takeoff period. The model counts hazard only during the takeoff period, as the AI Futures Project’s 75 percent does. So it assumes that the danger ends at completion, for example because a top-expert-dominating AI still under control can help align its successors.

Slowing down lowers both parts in this calibration. Step hazard falls most, because each new level is reached with more safety knowledge. State hazard falls less, because the lab also spends more months at each level.

The panel. Cumulative hazard against months. The blue curve is the total; green is state hazard, the danger of having a model in use; orange dashed is step hazard, the danger of training a stronger one. The dotted line marks capability one half. When the sliders move, the gray dashed curve keeps the full-speed total for comparison.

The implication is that the danger comes late. Most of the hazard builds after capability one half, in the last months of the takeoff period. At full speed the total reaches ln 4, a catastrophe probability of 75 percent.

The math

Step 3 · Paper, Section 2.3

Safety knowledge lowers both hazards at every later level

Safety research builds a stock of knowledge, and knowledge scales both hazards down at every later capability level. Knowledge learned now keeps protecting later, stronger models. A year of safety research at the full-speed share adds about 1.03 units of knowledge, which multiplies hazard at every later level by about 0.36.

So two things lower risk in the model: moving compute into safety, and slowing capability research so that knowledge has more time to build before each level. The figure shows both: each curve slopes down as research slows, and a larger safety share moves the whole curve down.

The panel. Catastrophe probability of one frontier against its research share, on a log scale. The green curve uses the safety share on the slider; when it moves, the gray dashed curve keeps the full-speed 5 percent. The gray dot is the paper's 75 percent at 50 percent research.

Read across: slower research means lower risk, because knowledge has more time to build before each capability level. Read up: more safety research shifts the whole curve down. The implication is that speed and safety effort are two separate levers on the same risk.

The math

Step 4 · Paper, Section 2.5

The value of a month: the hazard one more month of delay removes

Slowing down trades time for safety. When research rises a little, arrival comes earlier by τ(x) months per unit of research share, and cumulative hazard rises by H′(x). Their ratio is the value of a month, V(x) = H′(x)/τ(x): the hazard removed by one more month of delay, reached by slowing research a little. It is measured in cumulative hazard per month, not in percentage points of probability.

At full speed, one point less research delays arrival by 0.11 months, and a month of delay removes 8.3 percent of cumulative hazard. Slowing removes the most hazard when development is fast, so V rises with the research share. That is not automatic. Slowing gives knowledge more time to build before each level, but it also spends more months at each level. On the baseline curve knowledge wins, and V is positive and rising. That holds for the paper's 18-month takeoff period but not for every calibration: with a takeoff period shorter than about 13 months, V rises and then falls (Step 13).

This gives the benchmark for the rest of the paper. A single owner of the whole industry, or a common rule, keeps slowing while a month of delay removes more hazard than a month of waiting costs. Waiting costs the discount rate, r = 0.25 percent a month, so it stops where V = r, at 0.06 percent research.

Cumulative hazard, H(x)

Value of a month, V(x)

Left panel, cumulative hazard. Cumulative hazard rises with the research share (log scale). The vertical line and dot mark the share on the slider.

Right panel, the value of a month. The black curve is the left curve's slope divided by the months each point of research saves. The orange dashed line is the discount rate, the cost of waiting a month. A common rule slows down until the two meet, at the dot marked V = r.

The implication is how far a common rule would slow. At full speed a month of delay is worth 8.3 percent a month, far above the discount rate of 0.25 percent. Only at 0.06 percent research does a month stop paying for itself.

The math

Part 2Why labs race: the contest and the shared hazard

Step 5 · Paper, Section 3.1

How the labs compete: a shared hazard and a contest for a prize

Now add rivals. Three labs each run their own frontier on their own compute, and each chooses one research share for the whole takeoff period. The industry's hazard is the average of the labs' hazards, and a catastrophe from any frontier ends every lab's stake.

Each lab keeps a stake B if the world survives, and the lab that wins a contest over arrival dates also gets a commercial prize Π, 3.0 times each lab's stake, half of the industry's terminal value. The contest is the logistic, or difference-form, contest success function of Hirshleifer (1989), with arrival dates in place of efforts. Its parameter κ, which Hirshleifer calls the mass effect, is the contest sensitivity: each month of head start multiplies a lab's odds against a rival by eκ.

Only differences in arrival dates matter: if every lab arrived a month earlier, no lab's odds would change. Among the contest success functions axiomatized by Skaperdas (1996), the logit is the one with this property (Jia, Skaperdas and Vaidya 2013). At κ = 0 the prize is a lottery; as κ grows, the first lab to arrive wins with near certainty. The calibration fits κ = 0.075 per month, so a one-year head start multiplies the odds by 2.5, and a lab one year ahead of two level rivals wins 55 percent of the time, against one in three when all are level.

So each lab weighs three things. Survival gives it a reason to slow; discounting and the contest give it reasons to arrive sooner. The right figure shows how they balance. When its rivals put 50 percent of compute into research, a lab's own payoff peaks at 50 percent too. The industry's payoff at a common share peaks at 0.06 percent. Move the rivals' share to see a lab's best reply: if its rivals slowed to the common-rule share, a lab would still race well ahead of them.

Sources: Hirshleifer (1989), "Conflict and Rent-Seeking Success Functions: Ratio vs. Difference Models of Relative Success", Public Choice 63; Skaperdas (1996), "Contest Success Functions", Economic Theory 7; Jia, Skaperdas and Vaidya (2013), "Contest Functions: Theoretical Foundations and Issues in Estimation", International Journal of Industrial Organization 31.

Chance of winning the prize

Payoff against the research share

Left panel, chance of winning. One lab's chance of winning the prize against its head start over two level rivals. With no head start it is one in three (the dashed line); a year ahead, it is 55 percent. The implication is that the prize rewards speed without making it winner-take-all: a year's lead multiplies the odds by 2.5.

Right panel, payoff against the research share. Each curve is a payoff as a share of its own best (log scale). Blue is one lab while its rivals hold the share on the slider (the dotted line); orange is the industry when every lab uses the same share. The implication is the gap between the two peaks: with rivals at 50 percent a lab's best reply is 50 percent, while the industry's best common share is 0.06 percent.

The math

Step 6 · Paper, Section 3.2

Racing labs keep speeding up long after a common rule would stop

Three regimes organize the comparison. They differ in who chooses the research share and in what that choice maximizes.

Racing. Each lab chooses its own share to maximize its own payoff (Step 5), taking its rivals' shares as given. Arriving earlier raises its odds of winning the prize. The racing share xR is a symmetric Nash equilibrium: when every other lab uses xR, no lab gains by choosing a different share.

Common-share coordination. The industry sets one share x for every lab to maximize the labs' total payoff. At a common share every lab arrives at T(x), industry hazard is H(x), and each lab wins the prize with probability 1/N. Adding up the N payoffs gives the joint payoff

Ujoint(x) = e−H(x) e−rT(x) (Π + NB).

The prize goes to some lab whatever the share, so it no longer rewards speed. The coordinated share xJ maximizes Ujoint.

No contest. Labs again choose separately, but the prize does not depend on speed. With κ = 0, each lab wins with probability 1/N whatever its arrival date, as in a lottery. Lab n's payoff becomes

UnE = exp(−(1/N) Σm H(xm)) e−rT(xn) (Π/N + B).

The no-contest share xE is the symmetric equilibrium share of this game. Each lab still bears only one N-th of the hazard its speed adds, so only this hazard externality separates xE from coordination. The subscripts stand for racing, joint, and externality.

A common rule stops where a month of delay saves only the discount rate, V = r (Step 4). A racing lab stops much earlier. Three steps show why (Proposition 1).

The hazard cost. When one lab speeds up a little, its arrival comes earlier by τ months per unit of research, but industry hazard rises by only H′/N per unit, because its frontier carries one N-th of the industry's hazard.

The private gain. Each month saved is worth r in discounting. It also improves the lab's odds of winning the prize, worth κ(N − 1)π/N at equal shares.

The hurdle. The lab keeps speeding up until the hazard cost of a month equals its gain, which happens where V = N r + κ(N − 1)π. A common rule moves all frontiers together, so it weighs the whole hazard against r alone.

The same steps give each regime's stopping rule. Racing labs bear H′/N per unit of research and gain r + κ(N − 1)π/N per month saved, so they stop where V = N r + κ(N − 1)π. Under coordination the industry bears the full H′, and the prize no longer rewards speed, so it stops where V = r. Without the contest, labs still bear only H′/N, but the contest gain is zero, so they stop where V = N r + 0 = N r.

In the calibration, three racing labs put 50 percent of compute into research and stop when a month saved costs 8.3 percent in cumulative hazard, 33 times the discount rate. A common rule cuts research to 0.06 percent of compute and stretches the takeoff period to 69 months. Capped at 12 months of delay, the best common rule uses 7.7 percent and gives 47 percent risk.

The panel. The black curve is the value of a month against the research share, both on log scales. Each dashed line is a hurdle: blue for racing labs, gray dotted without the contest, orange for a common rule. Each industry keeps speeding up until the curve reaches its hurdle, at the dots; a higher hurdle means a later stop and more research.

The implication is the size of the race. Racing labs stop at 50 percent research, where a month is worth 8.3 percent, 33 times the discount rate. Without the contest they stop at 0.33 percent, and a common rule at 0.06 percent.

The math

Step 7 · Paper, Sections 3.2 and 4

Neither force alone explains the race; together they do

The racing hurdle has three parts: discounting, r; the hazard externality, (N − 1)r, because each lab counts only its share of the common hazard; and the contest part, κ(N − 1)π. The contest part is business stealing: a head start takes winning odds from rivals even though the industry's prize is fixed.

Switch the forces off one at a time and each gives its own hurdle on the same value-of-a-month curve. Without the contest, labs stop at N r = 0.75 percent a month and use 0.33 percent research. Without the hazard externality, each lab counts all of the hazard its speed adds; it stops at r + κ(N − 1)π/N = 2.75 percent and uses 3.1 percent research. Only together do the two forces reach the 8.3 percent hurdle at which labs race at 50 percent. Counting only one N-th of the hazard multiplies every private gain from speed by N, the contest's included.

So the share of the gap credited to the contest depends on which force is removed first. Removing the contest first lowers the catastrophe probability from 75 to 12 percent. That closes 63 of the 69 points between racing and coordination, over ninety percent of the gap. Removing it after the externality closes 39 percent; averaged over the two orders, 65 percent.

Read these numbers with care. The contest sensitivity is fitted so that racing labs put 50 percent of compute into research. What the numbers show is that neither force alone explains racing that fast.

Four hurdles on the value-of-a-month curve

How the gap between racing and a common rule splits

Top panel, four hurdles. The same curve as Step 6, with one hurdle for each combination of forces: blue both, green the contest only, gray the hazard externality only, orange neither. The dots show where each meets the curve. The implication is that each force alone supports little research, 3.1 percent with only the contest and 0.33 percent with only the externality, while together they reach 50 percent.

Bottom panel, the gap split two ways. Each bar is the 69-point gap in catastrophe probability between racing and a common rule. Green is the part removed by taking away the contest, gray the part removed by taking away the externality. The implication is that the contest's share depends on the order: 91 percent if it goes first, 39 percent if second, 65 percent on average.

The math

Part 3What changes the race: more labs, open source, restraint and rules

Step 8 · Paper, Section 3.3

More labs raise each lab's hurdle

An added lab is one more rival whose prize can be taken, and each lab then counts a smaller share of the hazard its speed adds. Both raise the racing hurdle: 5.0 percent a month with two labs, 8.3 with three and 10.6 with four (Proposition 2). Research rises from 11 percent of compute to 50 percent, then to the 95 percent ceiling.

If instead each frontier keeps its own hazard, adding a lab adds risk directly, while research shares can fall beyond three labs because the contest part peaks there. Aggregate risk still rises with the number of labs in every calibrated case.

The panel. Catastrophe probability against the number of racing labs. The blue line holds industry hazard fixed; the orange dashed line lets each new lab add its own frontier's hazard. One lab is coordinated by definition. The table lists each case.

The implication is that fragmentation speeds the race. More labs raise each lab's hurdle, from 5.0 percent a month with two labs to 8.3 with three and 10.6 with four. Two labs race at 11 percent research with 53 percent risk; four or five hit the ceiling with 81 percent.

The math

Step 9 · Paper, Section 3.4

Open source would slow labs to the coordinated pace

Suppose every lab publishes its capability findings at once. The paper calls this open source, in its ideal form. All labs then work from one shared frontier, and each lab's research moves it for everyone. No lab can get ahead of the others, because every finding reaches all of them at once, so the contest disappears. And each lab's research raises the one shared hazard, so each lab bears all of it and the hazard externality goes too. The condition becomes V = r, the coordinated pace (Proposition 3).

Real releases fall short of this ideal in three ways. A release that comes with a lag gives the publishing lab a head start for that long, so part of the contest returns. Research kept private, outside the shared frontier, restores a private incentive to race. And published findings also reach foreign rivals, a channel the paper leaves out. Monitoring how much compute a lab spends on safety is a different policy, because it makes no finding public.

The panel. Each bar is a catastrophe probability, with the research share under its label: racing labs on separate frontiers, the same labs without the contest, and open source. The settings are those of Step 6.

The implication is that open source removes both forces at once. Removing only the contest lowers risk from 75 to 12 percent. Open source reaches the coordinated level, 0.06 percent research and 6 percent risk, because every lab's research moves all labs together.

The math

Step 10 · Paper, Section 3.5 and Table 5

One lab's restraint does much less than a common rule

Suppose one lab commits to a lower research share while its rivals choose their best responses. When that lab adopts the coordinated share, its rivals slow only from 50 to 35 percent. Industry risk falls to 57 percent, which is 27 percent of the reduction the same rule achieves when it binds every lab.

Whether rivals slow at all depends on the prize. Near equal shares, rivals slow when the prize is large relative to each lab's stake, Π/B > N(N − 2), and speed up when it is smaller. The baseline sits exactly at that boundary, so a small restraint leaves rivals unmoved; a larger one makes them slow.

Rivals' research share

Catastrophe probability

Left panel, rivals' research share. The blue curve is how fast the two rivals go as one lab restrains itself (log scale); the dashed line is where they stand without restraint. The implication is that rivals barely slow: when one lab adopts the coordinated share, they go from 50 to 35 percent.

Right panel, catastrophe probability. Blue is risk with one lab restrained, orange is risk when the same share binds every lab, and the dashed line is everyone racing. The implication is the gap between blue and orange. At the coordinated share, one lab's restraint lowers risk only to 57 percent, against 6 percent under a common rule: 27 percent of the gain.

Together: restraint works only if it binds every lab. The dots are the three cases of Table 5, and the race settings are those of Step 6.

The math

Step 11 · Paper, Sections 1, 4, 5 and 6

Safety rules and audits act on safety, not on speed

Safety effort is fixed inside the race, so the paper imposes higher safety shares on every lab (Table 4). Racing labs offset part of the extra safety by racing faster: with 25 percent safety they hit the research ceiling of 75.0 percent, against 50 percent at the baseline. Racing risk still falls, to 53.6 percent, but a common rule at the same safety share gives 4.1 percent. The contest's share of the gap stays near 92 percent.

The latest industry response works on this margin. On September 29, 2026, six leading AI and chip companies, Google, Anthropic, Meta, OpenAI, xAI and Nvidia, signed the White House Accord on Super Intelligence. It commits them to internal safety controls, an internal team that checks those controls, independent outside auditors, and a board committee that reviews the results. It sets no penalties and does not address how fast labs develop.

In the model, the accord will not slow the race, because it leaves each lab's research share free. Its audits can lower risk only through safety effort, within the limit the figure shows. Its signatories will "meet regularly to establish standards," which may raise antitrust concerns; rules that leave research shares free leave the racing wedge in place, unless labs open-source all their capability findings.

Sources: The White House (2026), "White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities," September 29, archived by the American Presidency Project; Crowell & Moring (2026), client alert, October 7.

Catastrophe probability

Racing labs' research share

How to read the panels. The model has no separate variable for the accord's four layers; it shows the one margin they could act on. The slider imposes the same safety share on every lab, from the baseline 5 percent upward. Read it as how hard the controls and audits push labs to spend on safety.

Left panel, catastrophe probability. Blue is racing labs, orange is a common rule at the same safety share, and gray dotted is the case with no contest. More safety lowers every line: racing risk falls from 75.0 percent at 5 percent safety to 53.6 percent at 25 percent. The implication is the gap between blue and orange. Only a rule on speed closes it, and it stays large: at 25 percent safety a common rule gives 4.1 percent. The contest still accounts for 91–92 percent of that gap.

Right panel, racing labs' research share. It rises with safety, from 50 to 75.0 percent at 25 percent safety, where it hits the dashed ceiling of 1 minus the safety share. The implication is an offset: safer labs race faster and give back part of the gain.

Together: audits can make racing safer, but they cannot make it slower. The dots are the three rows of Table 4, and the race settings are those of Step 6.

The math

Part 4How robust: the safety yield, the takeoff period, the hazard level and one parameter at a time

Step 12 · Paper, Sections 2.5 and 4.3, Table 8

The safety yield sets how risky every regime is, not how much the contest explains

The safety yield σ is what makes a month valuable. It is the knowledge a month of safety research adds (Step 3). At the baseline safety share, knowledge on any path is S = σt, so cumulative hazard is

H(x) = ∫01 [h(F)/fx(F) + k(F)] e−σtx(F) dF,

where tx(F) is the date at which the path at share x reaches capability F, and fx(F) its speed there. Slowing research moves every date later. Each extra month multiplies the hazard still ahead by e−σ, but it also adds a month of state exposure at the current level. Without safety learning, σ = 0, only the exposure remains: V is negative at every research share, and neither racing labs nor a common rule would slow down.

A higher yield makes each month remove more of the hazard ahead, but it also leaves less hazard ahead to remove. So at each research share V first rises and then falls as σ grows. At the baseline yield, V at the racing share is close to its peak, while at the slow shares of coordination it is well past it. A 10 percent higher yield lowers V by 33 percent at xJ but by only 2 percent at 50 percent research.

The figure scales σ, holding the hazard functions h and k, and κ, π and r, fixed. A higher yield lowers risk under every regime, and labs offset part of the gain by racing faster, as with the safety rules of Step 11. At 1.25 times the baseline yield, full-speed risk falls from 75 to 63 percent. Racing labs speed up from 50 to 61 percent research, so racing risk falls only to 66 percent. Across the higher yields in Table 8, the faster race gives back 26–34 percent of the safety gain. A common rule speeds up too, from 0.06 to 0.22 percent research, while its risk falls from 5.8 to 4.9 percent. The contest's share of the gap stays at 91–92 percent.

A lower yield pushes the race to its ceiling. At three quarters of the baseline, V peaks at 74 percent research without reaching the racing hurdle, so racing labs use 95 percent research and carry 89 percent risk. At half the baseline the paper finds no symmetric equilibrium, so the slider stops at three quarters.

Four hurdles on the value-of-a-month curve

Catastrophe probability against the safety yield

Top panel, four hurdles. The black curve is the value of a month at the yield on the slider; the gray curve is the baseline. The hurdles and colors are those of Step 7. A higher yield lowers the curve most at slow shares, so above the baseline every stopping point moves to the right: labs and a common rule both speed up.

Bottom panel, catastrophe probability against the yield. Gray dashed is full speed, blue is racing labs, gray dotted is no contest, and orange is a common rule; the dots are the rows of Table 8. The implication is in two gaps. Blue falls less than gray dashed, because labs race faster and give back 26–34 percent of the gain. Blue stays far above orange, and the contest still accounts for 91–92 percent of that gap.

Together: better safety research makes every regime safer but does not end the race. The race settings are those of Step 6.

The math

Step 13 · Paper, Section 2.5, Appendix A.4 and Table 10

A shorter takeoff period makes labs race harder, and the contest still explains the gap

The baseline takeoff period lasts 18 months, from the Automated Coder to TED-AI. Forecasts of this stretch differ, so the paper also solves the model for takeoff periods from 9 to 36 months. Each length refits the safety yield and the speed of research, keeping the 75 percent probability at full speed and the 40 percent probability after a 12-month pause.

A short takeoff period packs the same 75 percent probability into fewer months, so each month at a capability level is more dangerous and speeding up saves more exposure. The value of a month is the hazard added by arriving with less safety knowledge minus the hazard removed by spending fewer months exposed. At 50 percent research the exposure part is 6.4 percent a month with a 9-month takeoff period and 2.6 percent with an 18-month takeoff period. The knowledge part barely changes, 11.0 against 10.8 percent. So with a 9-month takeoff period the value of a month peaks at 16 percent research and falls beyond it.

This breaks the paper's way of fitting κ. The baseline fits κ so that racing labs choose 50 percent research, and that needs the value of a month to rise through 50 percent. With a takeoff period shorter than about 13 months, no κ makes 50 percent a racing equilibrium. A lab whose rivals choose 50 percent does better by slowing almost to zero or by racing to the upper constraint.

The figure therefore holds κ at its baseline value, 0.075 per month, because κ describes how the market rewards a head start and need not change with the length of the takeoff period. Shorter takeoff periods then make labs race harder, because the value of a month is lower. Labs race at the 95 percent upper constraint with 9- and 12-month takeoff periods and at 33 percent research with a 36-month takeoff period. Removing the contest closes 91 to 92 percent of the racing–coordination gap at every length.

Four hurdles on the value-of-a-month curve, at the chosen takeoff period

Share of the racing–coordination gap closed by removing the contest

Top panel, four hurdles. The black curve is the value of a month at the takeoff length on the slider; the gray curve is the 18-month baseline. The hurdles and colors are those of Step 7, with κ held at its baseline value, and the thin vertical line marks 50 percent research. With short takeoff periods the curve stays below the racing hurdle, so racing labs use the 95 percent ceiling.

Bottom panel, share of the gap. Blue is the share of the gap between racing and a common rule that removing the contest closes; green is the average over both removal orders. Both hold κ at its baseline value, and the dots are the columns of Table 10. The shaded band marks the takeoff periods for which no κ makes 50 percent research a racing equilibrium.

Together: the length of the takeoff period changes how fast labs race, not how much of the gap the contest explains. The race settings are those of Step 6.

The math

Step 14 · Paper, Section 4.5 and Appendix A.5

If the hazard is lower, the contest explains less of the gap, and below a floor nothing

The paper's baseline is the AI Futures Project's 75 percent chance of catastrophe at full speed, a figure for a race with little safety effort. It reports 25 percent as the lower end of its range (Section 4.5), and this step lets the full-speed probability run from 5 to 90 percent. Each value refits the safety yield and the hazard scale, moves the pause target in the same proportion to the full-speed hazard, and refits κ so that three racing labs choose 50 percent research.

The refit absorbs most of the change. The value of a month falls in proportion to the cumulative hazard, so the hurdle a racing lab must meet falls with it, and the part of the race that needs the contest shrinks. With the contest removed first, it closes 86 percent of the gap between racing and a common rule at 50 percent, 69 percent at 25 percent, 58 percent at 20 percent and 35 percent at 15 percent. At every value on the slider, racing labs carry more risk than a common rule.

At the two ends of the paper's range the hurdle a racing lab must meet is 8.3 percent a month at 75 percent and 1.7 percent at 25 percent, 33 and 6.9 times the common rule's. The gap in catastrophe probability between racing and a common rule falls from 69 to 21 points. The contest closes 91 and 69 percent of it when removed first, 39 and 20 percent when removed second, and 65 and 44 percent on average. At 25 percent the one-sided counterfactuals swap order: without the contest labs use 5.2 percent research, and without the hazard externality 3.1 percent, against 0.33 and 3.1 percent at 75 percent.

There is a floor. Racing at 50 percent research needs the value of a month at that share to exceed the no-contest hurdle N r, 0.75 percent a month. Below a full-speed probability of about 11.8 percent it does not: the value of a month is 0.63 percent at 10 percent and 0.31 percent at 5 percent. Then no positive κ gives a 50 percent race, because discounting and the hazard externality alone already push labs past it, and the contest has no work to do. The figure draws the hurdles for κ = 0 there.

Four hurdles on the value-of-a-month curve, at the chosen full-speed probability

Share of the racing–coordination gap closed by removing the contest

Top panel, four hurdles. The black curve is the value of a month at the full-speed probability on the slider; the gray curve is the paper's 75 percent. The hurdles and colors are those of Step 7, with κ refitted so that labs race at 50 percent, and the thin vertical line marks 50 percent research. Where no positive κ does that, the hurdles are drawn for κ = 0, so the racing and no-contest hurdles coincide at N r. The implication is that a lower hazard lowers the curve in proportion, and below the floor the curve no longer reaches N r at 50 percent.

Bottom panel, share of the gap. Blue is the share of the gap between racing and a common rule that removing the contest closes; green is the average over both removal orders. The shaded band marks the full-speed probabilities with no 50 percent race. The implication is that the contest's share falls steadily as the hazard falls, and reaches zero at the floor.

Together: a lower hazard leaves the race faster than a common rule, but it makes the contest a smaller part of the reason. The race settings are those of Step 6.

The math

Step 15 · Paper, Appendix A and Figure 2

One parameter at a time, the contest still explains most of the gap

Many numbers in the calibration are judgments, so the paper varies each parameter in turn, refitting the safety yield, the hazard scale and κ each time. Removing the contest closes between 69 and 96 percent of the gap across these cases.

The share falls below ninety percent in five cases. Two of them lower the full-speed probability, to 50 percent (86 percent of the gap) and to 25 percent (69 percent). The others are five labs (83 percent), compute doubling every 20 months (89 percent), and a 50 percent pause target (89 percent, κ held). The average over both removal orders falls below one half only when the full-speed probability is 25 percent (44 percent).

Where no κ makes 50 percent research an equilibrium, at short takeoff periods, pause targets of 50 percent or more and a state-hazard weight of 1, κ keeps its baseline value and labs race at the 95 percent upper constraint, as in Step 13. Varying the safety share, the learning multiplier, or the reference research share with labs racing at it, leaves every number unchanged, so the figure leaves them out.

Share of the gap closed by removing the contest

Range over each parameter's sweep

Top panel, one parameter. Choose a parameter on the right. Blue is the share of the gap closed with the contest removed first and green the average over both removal orders, at each value the paper uses; the dashed line marks the baseline. Dots off the lines mark values where κ keeps its baseline value because no κ makes 50 percent research an equilibrium. At a 60 percent pause target the average is missing, because without the hazard externality the game has no symmetric equilibrium.

Bottom panel, ranges. For each parameter, the smallest and largest share over its sweep. These numbers come from the paper's Figure 2; they are computed once from the paper's model, not in the browser.

Part 5What the model leaves out

Step 16 · Paper, Section 6

What the model leaves out

The results are scenario calculations in an illustrative calibration. Four limits qualify them.

  1. Labs choose one research share for the whole takeoff period, safety effort is fixed inside the race, and customer services are left out of the labs' payoff. Dynamic strategies, observable leads, unequal allocations and service revenue could change both the speed of the race and its welfare effects.
  2. Compute grows on its own schedule, and sharing capability findings at home interacts with access by foreign rivals.
  3. Safety learning has diminishing returns in its share of compute but does not grow with total compute.
  4. The calibration uses illustrative targets, not identified causal estimates. Numerical best-response checks verify the reported equilibria without proving that they are the only ones.

How detection, response time and the cost of monitoring could sustain restraint between countries is the subject of the companion paper, Agreements Have a Clock: Compute Growth and the Verification of AI Restraint.