Model Prototype · webmaster
AI supremacy assumptions
AI supremacy means an infinite time horizon over a fraction of task domains. The time-horizon anchor remains 11.6 hours at 50% success on January 1, 2026, with the existing 10–13 hour input interval.
The default starting supremacy is 5.5%. This is the site author's subjective judgement, not a measured result or an estimate independently established by the time-horizon data. It is an editable parameter. No earlier starting date or invented historical horizon is used.
Finite time-horizon observations do not uniquely identify how much of the task domain has infinite horizon. METR fits success against finite task duration; its measurements do not establish this infinite-horizon share. METR also describes sensitivity to modelling choices. Treating success percentages as ordered fractions of task domains is an additional assumption of this prototype.
How the starting curve works
The original curve is extended with an explicit success floor. Domains up to the chosen starting percentage already have infinite horizon. The remaining curve retains its logistic form in log time horizon and is fitted to the existing 50% and 80% starting horizons. Setting starting supremacy to zero recovers the original curve exactly.
In percentage units, let s be starting supremacy and p the selected task percentage. For p ≤ s, the horizon is infinite. Otherwise:
odds(p) = (100 − p) / (p − s) H(p) = H(50) × [odds(p) / odds(50)]^a a = log(H(80) / H(50)) / log(odds(80) / odds(50))
The existing calibration gives H(80) / H(50) = 0.18857874166344962. With the default 11.6-hour anchor, the 80% horizon remains about 2.19 hours. A finite 50% horizon requires starting supremacy below 50%, which is why the input has that bound.
Every percentage then spends the same algorithmic-and-compute progress under the same difficulty curve. A finite horizon becomes infinite when that progress exhausts its remaining difficulty integral. The 50% and 80% horizon trajectories, utility calibration, and full-supremacy timing retain the existing equations. The displayed mean averages supremacy over the current uncertainty ensemble. In the default absolute-progress prototype, this ensemble uses the Millennium-history update described below; the central estimate remains a separate reference.
Progress beyond supremacy
The additional readout counts the share of tasks where AI has progressed at least ΔP beyond the task's supremacy threshold. ΔP is the absolute-progress difference between 5.85% and 20% supremacy under the current central parameters: P(20%) − P(5.85%). With the default parameters it is approximately 2.15973 progress units. These are doubling units of the underlying absolute-progress scale, equivalent to a factor of 2ΔP ≈ 4.4683, rather than a percentage-point difference in task share.
At the selected date, each simulation subtracts ΔP from accumulated absolute progress and evaluates its supremacy share at that reduced progress. On the central trajectory, 20% supremacy corresponds to a margin-qualified share of 5.85%; the displayed readout instead averages across parameter draws. With no-ASI conditioning disabled, both task-share readouts average the ensemble selected by the prior/posterior buttons: original parameter draws in prior mode, or draws resampled from the trajectory posterior in posterior mode. With conditioning enabled, both average only the surviving draws used by the probability map. The checkbox changes which simulations count, without switching between a central estimate and an average. The central ΔP is used for every draw and is recalculated when parameters change.
The model does not specify how far beyond supremacy the starting 5.5% of tasks already were. Their thresholds are set to zero, so this readout conservatively credits only progress accumulated since January 1, 2026. It reports zero before that accumulated progress reaches ΔP, then includes the starting share. If a changed starting-supremacy parameter already covers both reference percentages, ΔP is zero.
Both task-share readouts also display a separate central estimate, evaluated at the selected date using the central parameter values. The central estimate changes with the date and parameters, but not with the conditioning checkbox: it is a fixed reference trajectory, not an average over the surviving simulations. The average beside it continues to follow the checkbox.
The task-distribution selection also shows its share of the outcomes below 100% supremacy. This is the probability in the selected bars below 100%, divided by the total probability in all bars below 100%. The separate exactly-100% bar is excluded from both numerator and denominator, even if the selection includes it. Selecting only that bar therefore gives 0% when other outcomes exist; if all probability is at 100%, the ratio is undefined and displays “Unavailable”. The calculation uses the displayed distribution, including its current conditioning.
The second task-share histogram, marked ΔP, shows the distribution of the margin-qualified share across the same parameter draws as the ΔP average. Each draw subtracts the same central ΔP before calculating its task share. The horizontal axis is a percentage of tasks, not progress units. Bins cover one percentage point each, giving 100 intervals below 100% plus a separate bar for exactly 100%. Both histograms follow the selected date, parameters, and conditioning; their bar selections are independent, and both provide the selected range's overall probability and share of below-100% outcomes.
The median beside each average uses the same simulations as that average, including the same conditioning filter. It is calculated from the exact task shares before histogram binning: the middle sorted value for an odd sample count, or the mean of the two middle values for an even count. For the ΔP median, each simulation first subtracts the common central progress margin. Empty or invalid samples display “Unavailable”. The central estimate remains a separate reference trajectory.
The comparison checkbox overlays a second date on both task-share histograms. Filled bars show the main selected date; dashed outlines show the comparison date, with a common vertical scale for both. Comparison defaults to the viewer's actual current day and can be changed with the date field. The “today” button restores the current-day default. Editing this date does not change the main selected date. Both distributions use the same current parameters, central ΔP, and conditioning on the actual current day; changing the comparison date does not move the conditioning observation. Selected ranges retain their main-date readouts and show an additional comparison probability.
Choosing the information used by the forecast
The prior/posterior buttons above the AI supremacy map control whether the September and current-day Millennium observations are used throughout the prototype. Posterior is selected by default and jointly updates trajectories and SD from the displayed history. Prior restores the original January 1 parameter distribution and the chosen SD prior, with no later problem evidence. The observed-count sequence beside these buttons always shows which history is active. Selecting a fixed SD does not switch the history mode.
The “condition on no ASI as of today” checkbox is a separate observation and works in either mode. Prior with this checkbox unchecked restores the original AI forecast. Prior with it checked uses only the additional absence of full supremacy through today. Posterior already excludes full supremacy through the current local day because this model’s unbounded progress would solve every remaining problem; the extra checkbox is redundant with the current-day problem evidence in this model.
The left date slider selects the date being evaluated; it does not change the information cutoff. An October 24 display in posterior mode still uses the the current local day observation. Choose prior to inspect the forecast without that later problem history. Under the default parameters, full-supremacy probability by October 24 is about 36.33% in prior mode with no additional no-ASI observation, and zero in posterior mode. The explicit history is shown at the top to make this hindsight visible.
In prior mode the Millennium count chart follows all six problems known open on January 1, and the gap chart predicts the time between the first two of those six solutions. It conditions their thresholds on exceeding January progress, without reweighting the January parameter prior by that already-known starting state. In posterior mode the count chart follows the five remaining after the current local day, and the gap is between the next two of those five. Thus the denominator changes from six to five when the later history is enabled. Both modes can use a shared uncertain SD or a manually fixed SD.
Time from all Millennium solutions to full AI supremacy
This histogram measures elapsed calendar days from the last remaining Millennium problem's modeled solution to 100% AI supremacy. Each of the same million simulations used for the other Millennium histograms supplies both milestones: the maximum of its problem thresholds and the point at which its absolute progress becomes infinite. This preserves their dependence on the same trajectory and shared problem-difficulty SD. It is not a subtraction of independently sampled dates or of median dates.
The chart follows the same prior/posterior, fixed-SD, central-trajectory, and no-ASI settings as the other Millennium charts. Posterior mode samples the five remaining problems; prior mode samples the six open at the January baseline. Problems already recorded as solved need no additional waiting time. Changing the inspected date does not move the observation cutoff.
Bars cover two-day intervals, with ≥210 days retaining the entire tail, including infinite waits when the model never reaches full supremacy. Those outcomes remain in the probability denominator and make the average infinite. A tail containing infinite waits cannot be expanded. The vertical scale adjusts to fit the largest bar. Hover or focus bars to inspect probabilities, select a range by dragging or Shift with arrow keys, and click a populated interval to subdivide it. Median and average describe the sampled elapsed times. Full supremacy can follow a solution extremely closely; the remaining time is calculated directly to avoid subtracting nearly equal dates. These are model-threshold forecasts, without an additional allowance for research, verification, publication, or prize adjudication.
Time between the first and second Millennium solutions
With posterior selected, the gap histogram uses the joint posterior described below. Each outcome selects one trajectory–SD pair, then draws five independent domains and normal offsets conditional on their solution thresholds exceeding that trajectory’s progress on the current local day. It subtracts the earliest of these five future solution times from the second earliest. The September solution is excluded: this is the gap between the next two solutions among the remaining problems. The historical likelihood enters once, when choosing the pair. The histogram uses 1,000,000 deterministic samples. Its summaries change with the current-day evidence cutoff. The simulation is checked against the independent identity ∫ P(exactly one of the remaining five solved at t) dt.
Bars are two-day intervals, with a separate ≥210-day bar retaining the entire remaining tail. Probabilities sum to 100%; the tail is not discarded or redistributed. The vertical scale is fixed at 20%, and taller bars can overflow. The median and average summarize the actual waiting-time gaps; the median gap is not the difference between the two median milestone dates. Short and long dashed lines mark the median and mean respectively. Hover or use the arrow keys on the bars to inspect their probabilities. Drag across bars in either direction, or use Shift with the arrow keys, Home, or End, to select a contiguous range. The readout adds all selected bars and shows both their share of the current level and their share of all outcomes. Selections persist when the pointer leaves the chart. Escape clears a selection before returning to a previous level; cancelled gestures restore the previous selection. A drag does not drill down. Click, tap, or press Enter or Space on a nonempty bar to split its interval into 20 smaller bars. This can be repeated; each level uses the original sampled gaps and normalizes probabilities to the outcomes inside that interval. The mean and median also refer to that interval. The selected bar shows both its percentage within the current zoomed interval and its overall percentage among all sampled outcomes. The overall percentage is calculated directly from the bar’s sample count, avoiding rounding across nested levels. The displayed P(interval) retains the current interval’s probability among all outcomes. Range buttons return to any ancestor; the arrow button, Escape, or Backspace returns one level. The ≥210-day tail can also be expanded, with its plotted endpoint beyond the largest sampled gap. Empty bins, single-sample bins, and bins whose samples are all identical cannot be expanded. Deep levels reflect finite Monte Carlo samples, not additional model precision. Parameter changes reset the view to the whole distribution; date playback preserves the selected level.
This distribution does not depend on the selected forecast date. It updates when model parameters, SD, trajectory mode, or the optional no-ASI observation change, using a separate worker so date playback can continue. The problem-history endpoint is January 1 in prior mode and the current local day in posterior mode. An open page checks for a date change every 30 seconds and when it becomes visible again. While recalculating, the previous histogram remains visible.
The monthly histogram uses the earliest solution time from the same samples and groups it by calendar month, including the year. Each probability is the number of outcomes in that month divided by the total sample count. In posterior mode it forecasts the next of the five remaining solutions after the current-day cutoff. In prior mode it instead shows the first of the six solutions from January. Hover, tap, or use the arrow keys to inspect a month; scroll horizontally to inspect longer date ranges.
Millennium Prize problem count forecast
This is a scenario forecast under the site author's assumptions, not a measured difficulty calibration. A problem's solution threshold is a randomly drawn task domain's supremacy threshold plus an independent normal offset with mean ΔP and standard deviation σ. ΔP retains its calibration to the central progress difference between 5.85% and 20% supremacy. Negative offsets remain possible; the normal is not truncated.
With posterior selected, the default “σ | N” mode jointly updates the uncertain model parameters θ and difficulty SD σ. It treats the existing parameter-confidence distribution and an independent half-normal SD distribution as a January 1 prior, conditional on all six problems already being open. Each parameter draw determines its own progress trajectory and domain-difficulty curve. The joint posterior is p(θ, σ | N) ∝ pJanuary(θ)pJanuary(σ)L(θ, σ). The common extra-difficulty mean ΔP is still calibrated to the central parameters. This remains a deliberately restricted model: the observations do not identify both an unrestricted progress law and an unrestricted difficulty distribution.
On January 1, 2026, all six problems are taken as known to be unsolved. At that point the prior for σ is half-normal with scale ΔP: σ ∼ |N(0, ΔP²)|. This is an explicit prior choice, not a uniquely neutral representation of ignorance. The later evidence consists of exactly one solution by September 5, 2026, and still exactly one by the current local day. September 5 is a count observation, not an exact first-threshold-crossing time. September 5 stays fixed. The later observation follows the viewer’s local calendar day, independently of the displayed forecast date. This assumes the remaining five problems are still open through the start of that day; it is not an automatic check of external solution announcements.
Let qθ,σ(t) be the probability that a single solution threshold is below the progress of trajectory θ at t. Write q₀, q₁, and q₂ for January 1, September 5, and the current local day. Conditional on all six being open on January 1, the likelihood is L(θ, σ) = 6(q₁ − q₀)(1 − q₂)⁵ / (1 − q₀)⁶. The denominator accounts for the January knowledge; multiplying in a separate January survival likelihood would count it again. Both later observations enter once. Trajectories that have already reached infinite progress by the current local day receive zero weight, regardless of SD.
For reference, with an October 26, 2026 evidence cutoff and the default joint settings, the SD posterior median is approximately 1.54 progress units, its mean is 1.79, and its central 80% interval is roughly 0.27–3.70. The prior median is approximately 1.46. These observations weakly identify SD: halving or doubling the chosen prior scale moves the posterior median to roughly 0.72 or 3.2. The SD chart compares the prior (dashed outlines, σ) with the posterior (filled bars, σ | N), including the tail beyond four prior scales. The default numerical update uses 10,000 prior trajectory draws with eight independent stratified half-normal SD draws per trajectory, weighted by the likelihood. The effective trajectory sample size is about 5,880. Checks with more trajectories, more SD samples, and another random seed support the displayed rounding; the numbers are approximations, not precision measurements.
With posterior selected, the general AI probability map, task-share histograms, averages, medians, horizon uncertainty bands, and full-supremacy timing use the same updated trajectory marginal. Selecting prior restores their original parameter draws; changing SD then affects the Millennium predictions without reweighting the AI trajectory prior. Those components consume equally weighted ensembles, so their 10,000 draws are deterministic systematic resamples of the weighted marginal; the horizon bands use 512 such draws. Millennium forecasts integrate the original joint weights directly. Resampling introduces a small numerical approximation but does not apply the evidence again. With the October 26, 2026 evidence cutoff, full-supremacy probability by January 1 falls from about 71% under the original parameter prior to about 45% after the update. Central estimate readouts remain references to the input parameters.
The separate “central trajectory” checkbox holds θ fixed across both Millennium and general probabilistic forecasts. With SD inference enabled this reproduces the conditional SD estimate (at the October 26, 2026 cutoff), median approximately 1.496 and 80% interval 0.262–3.649, using 2,048 midpoint cells across eight half-normal scales. It is a comparison, not the default inference. Unchecking it restores trajectory uncertainty and the joint update. The separate no-ASI checkbox then applies the same additional filter in either mode.
For a future date t, a remaining problem’s solution probability conditional on θ and σ is pθ,σ(t) = [qθ,σ(t) − qθ,σ(the current local day)] / [1 − qθ,σ(the current local day)]. The count distribution averages Binomial(5, pθ,σ(t)) over the joint posterior. One trajectory and one SD are shared by all five problems within an outcome. Keeping this dependence matters: the prediction is neither the binomial at the mean probability nor a forecast at the median SD. With the October 26, 2026 cutoff and default parameters, the chance of all five being solved by January 1, 2027 is about 45%, whereas fixing the central trajectory forces all five by that date.
In posterior mode the “σ | N” checkbox learns SD by default. In prior mode it is labeled “σ” and retains the SD prior without applying later problem evidence; its histogram shows that prior alone. Unchecking it fixes SD, while the prior/posterior buttons continue to control which history updates trajectory weights. The cartoon switch chooses ΔP/3 or an editable fixed SD, initially 1.5. Enter a finite nonnegative SD, then press Enter or leave the field to apply it and select manual mode. Zero gives a constant extra-difficulty offset. Blank, negative, or nonfinite entries preserve the last valid calculation; Escape cancels an uncommitted edit. Re-enabling inference retains the last manual value. These choices last for the current page visit.
With posterior selected, the selected-date chart counts the five remaining problems, with probabilities for 0 through 5 solved. Prior mode instead counts the six January problems, with probabilities for 0 through 6. Its median is the smallest count with cumulative probability at least 50%. The five squares show that count, not identified problems. In posterior mode, at or before the current local day the count is zero because these five are known to have remained open through that observation. The optional no-ASI-today checkbox filters the joint predictions on the additional absence of full supremacy through the viewer’s current day; it does not assert that no further Millennium problems have been solved. If no trajectory can explain the selected evidence, or the SD prior is invalid, the forecast is unavailable.
This forecasts reaching a modeled solution threshold; it does not model research effort, verification time, publication, prize adjudication, or which named problem is solved first. The separate comparison-date histogram control applies to the task-share histograms, not to this count forecast.
As checked on October 25, 2026, Clay's official list places five problems under “Unsolved”, Navier–Stokes under “Active”, and Poincaré under “Solved”. Clay's September 11 announcement describes the Navier–Stokes problem as apparently settled, with its evaluation process continuing. The five-problem scenario here uses the author's requested count; it does not equate active prize review with an awarded prize.
Probability by date, given no full supremacy today
The cumulative curve conditions the same trajectory draws selected by the prior/posterior buttons on the observation that 100% AI supremacy has not happened by today's date. Simulations reaching full supremacy on or before today are discarded; the surviving simulations retain equal weight. A forecast for a future date is the fraction of survivors reaching full supremacy by that date. Simulations reaching it later, or never, remain in the denominator.
If F(t) is the original probability of achievement by date t and d is today's date, the curve is zero for t ≤ d. For t > d it is [F(t) − F(d)] / [1 − F(d)]. This updates the timing forecast using the non-occurrence observation, without fitting a new progress law or shifting the time-horizon anchor.
Today's date comes from the viewer's local calendar, using the same convention as the main date control. The observation is through the start of that calendar date, at daily resolution. Moving the inspected date does not move the conditioning date. The curve refreshes when the calendar date changes, including while playback is paused. Hover or tap to inspect the curve; left and right arrow keys control the overall displayed date.
The post-adjustment histogram uses differences of this same conditional cumulative curve. Every interval is exactly 1/12 of a mean Gregorian year (365.2425 / 12 = 30.436875 days), matching the model's monthly-growth conversion, and the first begins today. Boundaries are evenly spaced rather than aligned to calendar months. Probabilities use the existing daily step curve at each boundary; the dates in the hover readout have daily resolution.
Only complete intervals are shown, so the final bar is never a shorter segment. The separate hatched bar begins at the end of the last complete interval and includes all subsequent probability, including any leftover part of the simulation window and outcomes that never achieve full supremacy. All bars sum to 100%, and their cumulative totals match the adjusted curve at every interval boundary. Both adjusted charts share the observation date and update together when parameters or the calendar date change.
If no sampled simulations survive the observation, the conditional forecast is undefined and both adjusted charts display “Unavailable”; they do not invent new probabilities. The same applies when today is beyond the available simulation range. With very few surviving draws, the estimate becomes coarse and sensitive to sampling. Conditioning updates the cumulative display and post-adjustment histogram; the tape and original monthly histogram retain the selected prior/posterior baseline for comparison. All displays remain conditional on the model family and input uncertainty assumptions.
Probability during the displayed day
The daily readout follows the date displayed at the top of the model. It gives the probability of first achieving full supremacy during that calendar day under the existing adjustment for non-occurrence through the start of the actual current date. If C is the adjusted cumulative curve and d is the displayed day, this probability is C(d + 1) − C(d). Selecting a future date does not assume that supremacy remains absent until that future date; it simply selects that day's share of the existing distribution. Days before the actual current date have zero probability under the non-occurrence observation. The readout is unavailable if the next day is outside the simulation range or if no sampled outcomes survive the observation.
Left and right arrow keys move the overall display backward or forward by exactly one calendar day and pause playback. Navigation starts from the displayed day, including a date being previewed by hovering over the main chart, and stops at the range boundaries. Focused text, date, select, and parameter inputs retain their normal arrow-key editing. The daily readout follows date edits, chart selection, keyboard navigation, and playback; the observation date and the adjusted chart distributions remain anchored to the actual current date.
When 100% AI supremacy is reached
The histogram below the tape shows the probability of first reaching 100% AI supremacy in each calendar interval. A bar from January 1 to February 1 counts achievements after January 1 and on or before February 1; the first bar also includes the starting date. Hover, tap, or focus a bar for its interval and probability. The dashed vertical line follows the date selected elsewhere on the page.
Bar heights are differences of the tape's daily cumulative probabilities, using the same parameter draws and definition of full supremacy as the readout. Full supremacy requires unbounded absolute progress; the earlier point when the 50% horizon becomes infinite is insufficient. Times are resolved to a day.
The separate hatched bar is the probability of not reaching 100% by the displayed end date. It includes both later achievement and simulations that never achieve it. All bars together sum to 100%; the chart does not rescale the outcomes inside the window to hide the remaining probability. Changing the model parameters updates the distribution.
What the uncertainty means
The bands and tape probabilities are conditional on the chosen starting supremacy and curve family. The 5.5% judgement has no empirically justified confidence interval, so the model does not invent one. All draws use the selected starting value; existing intervals still vary the horizon, growth, utility, and difficulty parameters. A 100% tape probability at a low threshold on January 1 is conditional on that starting assumption, not empirical certainty.
Try other starting values to assess sensitivity. A visually plausible curve, agreement with intuition, or narrow conditional intervals do not validate the starting judgement or the infinite-horizon extrapolation.