Super Intelligence (SI) is not only a question of algorithms. This article examines fast takeoff versus slow takeoff: comparing hypotheses about the rate of capability change without treating either transition speed as established fact. It preserves the locked Clementi-depth floor with mechanisms, worked cases, sensitivity analysis, diagnostics, failure modes and RFE closure.
Search Intent and Immediate Answer
The search question in fast takeoff versus slow takeoff is comparing hypotheses about the rate of capability change without treating either transition speed as established fact. A rigorous answer separates capability metric, feedback rate, bottleneck, deployment, adaptation and time. These variables can move at different speeds and can constrain one another. Super Intelligence (SI) forecasting becomes misleading when one fast-moving variable is extrapolated as though every supporting system must move with it.
First principles begin with throughput. Identify the resource or process entering the system, the transformation being attempted, and the slowest stage that determines completed output. If compute doubles but power, fabrication, verification or high-quality data does not, the realised capability gain may differ from the headline expectation. Bottleneck maps are more useful than single-variable stories.
A small-system example provides a controlled baseline. Measure how performance changes when one resource is increased while everything else is held constant. Then repeat at larger scale. If the gain diminishes, another constraint has become active. This is the basic logic behind analysing physical limits, learning bottlenecks and capability takeoff rather than assuming linear continuation.
Definition and Scope
A small-system example provides a controlled baseline. Measure how performance changes when one resource is increased while everything else is held constant. Then repeat at larger scale. If the gain diminishes, another constraint has become active. This is the basic logic behind analysing physical limits, learning bottlenecks and capability takeoff rather than assuming linear continuation.
Frontier scale changes the environment. Procurement, networking, cooling, data quality, evaluation and organisational coordination can become first-order constraints. A technique that works economically on a small experiment may become expensive or operationally fragile when multiplied across a large system. SI analysis should follow the complete stack from algorithm to deployed capability.
A changed-assumption test is essential for forecasts. If a timeline depends on compute growth, change the compute-growth assumption. If it depends on research automation, reduce the assumed automation rate. If it depends on synthetic data quality, model verification failure. A forecast that changes dramatically under modest assumption changes should be presented with correspondingly wide uncertainty.
First Principles
A changed-assumption test is essential for forecasts. If a timeline depends on compute growth, change the compute-growth assumption. If it depends on research automation, reduce the assumed automation rate. If it depends on synthetic data quality, model verification failure. A forecast that changes dramatically under modest assumption changes should be presented with correspondingly wide uncertainty.
Failed forecasts are evidence about models of the world, not merely embarrassing dates. Ask which assumption failed: capability scaling, economics, adoption, hardware, regulation or the definition of the milestone. Updating the causal model is more valuable than quietly moving the date. This is how forecasting becomes a learning process.
Measurement should match the variable. Power is measured differently from model loss; data diversity differently from token count; takeoff speed differently from calendar predictions. Define the numerator and denominator before comparing trends. Ambiguous metrics can create apparent acceleration simply because the measurement changed.
The System Boundary
Measurement should match the variable. Power is measured differently from model loss; data diversity differently from token count; takeoff speed differently from calendar predictions. Define the numerator and denominator before comparing trends. Ambiguous metrics can create apparent acceleration simply because the measurement changed.
Extrapolation is the dangerous step. An empirical relationship observed over one range may continue, bend or break outside it. Scaling laws, hardware trends and benchmark curves are useful precisely because they summarise observed regularities, but none is a physical guarantee of indefinite continuation. State the range and assumptions.
Uncertainty has structure. Some uncertainty comes from measurement noise; some from unknown future discoveries; some from disagreement about definitions; some from strategic or economic behaviour. Combining them into one confident date hides the reason for uncertainty. Good SI timelines show which assumption contributes most to the spread.
The Core Bottleneck
Uncertainty has structure. Some uncertainty comes from measurement noise; some from unknown future discoveries; some from disagreement about definitions; some from strategic or economic behaviour. Combining them into one confident date hides the reason for uncertainty. Good SI timelines show which assumption contributes most to the spread.
Feedback loops can accelerate progress when outputs become inputs to the next improvement cycle. But the loop contains latency. A software change may be tested in minutes; a new chip may require design and fabrication; an energy project may take years. Takeoff speed depends on which loop dominates and how tightly it couples to the rest of the system.
Economics determines which technically possible paths are actually pursued. Compute has a price, electricity has a price, expert supervision has a price and failed experiments consume resources. Falling unit costs can accelerate adoption, while rising marginal costs can slow it. Economic constraints should sit beside technical ones.
What People Commonly Assume
Economics determines which technically possible paths are actually pursued. Compute has a price, electricity has a price, expert supervision has a price and failed experiments consume resources. Falling unit costs can accelerate adoption, while rising marginal costs can slow it. Economic constraints should sit beside technical ones.
Supply chains make intelligence physical. Advanced chips depend on specialised equipment, materials, fabrication capacity and logistics. Data centres require construction, grid connections and cooling. These facts do not disprove rapid cognitive progress; they specify the channels through which digital progress reaches scale.
Current evidence supports rapid improvement in several AI capability domains and continuing investment in compute infrastructure. It does not establish a fixed SI date, a guaranteed fast takeoff or unlimited synthetic-data improvement. The evidence is compatible with multiple future trajectories, which is why conditional reasoning is preferable to inevitability language.
Worked Example: A Small System
Current evidence supports rapid improvement in several AI capability domains and continuing investment in compute infrastructure. It does not establish a fixed SI date, a guaranteed fast takeoff or unlimited synthetic-data improvement. The evidence is compatible with multiple future trajectories, which is why conditional reasoning is preferable to inevitability language.
Safety preparation depends on lead time. Faster capability change can reduce the time available for evaluation, workforce adaptation and institutional response. Slower change can create more opportunity for learning but does not guarantee good governance. Preparation should therefore focus on observable indicators and adaptable controls rather than one predicted date.
Education should teach forecast literacy. A student should identify the event being forecast, the definition used, the date of the forecast, the probability assigned, the population or expert group sampled, and the assumptions underneath it. A probability is not a promise and a median survey response is not a scheduled event.
Worked Example: Frontier Scale
Education should teach forecast literacy. A student should identify the event being forecast, the definition used, the date of the forecast, the probability assigned, the population or expert group sampled, and the assumptions underneath it. A probability is not a promise and a median survey response is not a scheduled event.
For organisations, convert forecasts into scenarios rather than a single plan. Define what action would be appropriate if capability advances slowly, moderately or rapidly. Identify indicators that would trigger movement between plans. This reduces dependence on guessing the exact year and increases readiness across plausible paths.
Progress has four stages: define the variable, explain the mechanism, model sensitivity to changed assumptions, and design an update rule when evidence arrives. This is the Clementi-depth progression from recognition to independent control. Readers should finish able to revise a forecast, not merely repeat one.
Worked Example: A Changed Assumption
Progress has four stages: define the variable, explain the mechanism, model sensitivity to changed assumptions, and design an update rule when evidence arrives. This is the Clementi-depth progression from recognition to independent control. Readers should finish able to revise a forecast, not merely repeat one.
RFE closes the loop. Receiver: who needs the forecast or resource analysis? Function: what decision must it support? Evidence: what observable indicator changes the decision? Exit: when should the model or plan be revised? Applied to fast takeoff versus slow takeoff, RFE makes uncertainty operational rather than rhetorical.
The search question in fast takeoff versus slow takeoff is comparing hypotheses about the rate of capability change without treating either transition speed as established fact. A rigorous answer separates capability metric, feedback rate, bottleneck, deployment, adaptation and time. These variables can move at different speeds and can constrain one another. Super Intelligence (SI) forecasting becomes misleading when one fast-moving variable is extrapolated as though every supporting system must move with it.
Worked Example: A Failed Forecast
The search question in fast takeoff versus slow takeoff is comparing hypotheses about the rate of capability change without treating either transition speed as established fact. A rigorous answer separates capability metric, feedback rate, bottleneck, deployment, adaptation and time. These variables can move at different speeds and can constrain one another. Super Intelligence (SI) forecasting becomes misleading when one fast-moving variable is extrapolated as though every supporting system must move with it.
First principles begin with throughput. Identify the resource or process entering the system, the transformation being attempted, and the slowest stage that determines completed output. If compute doubles but power, fabrication, verification or high-quality data does not, the realised capability gain may differ from the headline expectation. Bottleneck maps are more useful than single-variable stories.
A small-system example provides a controlled baseline. Measure how performance changes when one resource is increased while everything else is held constant. Then repeat at larger scale. If the gain diminishes, another constraint has become active. This is the basic logic behind analysing physical limits, learning bottlenecks and capability takeoff rather than assuming linear continuation.
How to Measure the Variable
A small-system example provides a controlled baseline. Measure how performance changes when one resource is increased while everything else is held constant. Then repeat at larger scale. If the gain diminishes, another constraint has become active. This is the basic logic behind analysing physical limits, learning bottlenecks and capability takeoff rather than assuming linear continuation.
Frontier scale changes the environment. Procurement, networking, cooling, data quality, evaluation and organisational coordination can become first-order constraints. A technique that works economically on a small experiment may become expensive or operationally fragile when multiplied across a large system. SI analysis should follow the complete stack from algorithm to deployed capability.
A changed-assumption test is essential for forecasts. If a timeline depends on compute growth, change the compute-growth assumption. If it depends on research automation, reduce the assumed automation rate. If it depends on synthetic data quality, model verification failure. A forecast that changes dramatically under modest assumption changes should be presented with correspondingly wide uncertainty.
What the Data Can Show
A changed-assumption test is essential for forecasts. If a timeline depends on compute growth, change the compute-growth assumption. If it depends on research automation, reduce the assumed automation rate. If it depends on synthetic data quality, model verification failure. A forecast that changes dramatically under modest assumption changes should be presented with correspondingly wide uncertainty.
Failed forecasts are evidence about models of the world, not merely embarrassing dates. Ask which assumption failed: capability scaling, economics, adoption, hardware, regulation or the definition of the milestone. Updating the causal model is more valuable than quietly moving the date. This is how forecasting becomes a learning process.
Measurement should match the variable. Power is measured differently from model loss; data diversity differently from token count; takeoff speed differently from calendar predictions. Define the numerator and denominator before comparing trends. Ambiguous metrics can create apparent acceleration simply because the measurement changed.
What the Data Cannot Show
Measurement should match the variable. Power is measured differently from model loss; data diversity differently from token count; takeoff speed differently from calendar predictions. Define the numerator and denominator before comparing trends. Ambiguous metrics can create apparent acceleration simply because the measurement changed.
Extrapolation is the dangerous step. An empirical relationship observed over one range may continue, bend or break outside it. Scaling laws, hardware trends and benchmark curves are useful precisely because they summarise observed regularities, but none is a physical guarantee of indefinite continuation. State the range and assumptions.
Uncertainty has structure. Some uncertainty comes from measurement noise; some from unknown future discoveries; some from disagreement about definitions; some from strategic or economic behaviour. Combining them into one confident date hides the reason for uncertainty. Good SI timelines show which assumption contributes most to the spread.
Generalisation and Extrapolation
Uncertainty has structure. Some uncertainty comes from measurement noise; some from unknown future discoveries; some from disagreement about definitions; some from strategic or economic behaviour. Combining them into one confident date hides the reason for uncertainty. Good SI timelines show which assumption contributes most to the spread.
Feedback loops can accelerate progress when outputs become inputs to the next improvement cycle. But the loop contains latency. A software change may be tested in minutes; a new chip may require design and fabrication; an energy project may take years. Takeoff speed depends on which loop dominates and how tightly it couples to the rest of the system.
Economics determines which technically possible paths are actually pursued. Compute has a price, electricity has a price, expert supervision has a price and failed experiments consume resources. Falling unit costs can accelerate adoption, while rising marginal costs can slow it. Economic constraints should sit beside technical ones.
Reliability and Uncertainty
Economics determines which technically possible paths are actually pursued. Compute has a price, electricity has a price, expert supervision has a price and failed experiments consume resources. Falling unit costs can accelerate adoption, while rising marginal costs can slow it. Economic constraints should sit beside technical ones.
Supply chains make intelligence physical. Advanced chips depend on specialised equipment, materials, fabrication capacity and logistics. Data centres require construction, grid connections and cooling. These facts do not disprove rapid cognitive progress; they specify the channels through which digital progress reaches scale.
Current evidence supports rapid improvement in several AI capability domains and continuing investment in compute infrastructure. It does not establish a fixed SI date, a guaranteed fast takeoff or unlimited synthetic-data improvement. The evidence is compatible with multiple future trajectories, which is why conditional reasoning is preferable to inevitability language.
Sensitivity to Assumptions
Current evidence supports rapid improvement in several AI capability domains and continuing investment in compute infrastructure. It does not establish a fixed SI date, a guaranteed fast takeoff or unlimited synthetic-data improvement. The evidence is compatible with multiple future trajectories, which is why conditional reasoning is preferable to inevitability language.
Safety preparation depends on lead time. Faster capability change can reduce the time available for evaluation, workforce adaptation and institutional response. Slower change can create more opportunity for learning but does not guarantee good governance. Preparation should therefore focus on observable indicators and adaptable controls rather than one predicted date.
Education should teach forecast literacy. A student should identify the event being forecast, the definition used, the date of the forecast, the probability assigned, the population or expert group sampled, and the assumptions underneath it. A probability is not a promise and a median survey response is not a scheduled event.
Feedback Loops
Education should teach forecast literacy. A student should identify the event being forecast, the definition used, the date of the forecast, the probability assigned, the population or expert group sampled, and the assumptions underneath it. A probability is not a promise and a median survey response is not a scheduled event.
For organisations, convert forecasts into scenarios rather than a single plan. Define what action would be appropriate if capability advances slowly, moderately or rapidly. Identify indicators that would trigger movement between plans. This reduces dependence on guessing the exact year and increases readiness across plausible paths.
Progress has four stages: define the variable, explain the mechanism, model sensitivity to changed assumptions, and design an update rule when evidence arrives. This is the Clementi-depth progression from recognition to independent control. Readers should finish able to revise a forecast, not merely repeat one.
Physical and Institutional Latency
Progress has four stages: define the variable, explain the mechanism, model sensitivity to changed assumptions, and design an update rule when evidence arrives. This is the Clementi-depth progression from recognition to independent control. Readers should finish able to revise a forecast, not merely repeat one.
RFE closes the loop. Receiver: who needs the forecast or resource analysis? Function: what decision must it support? Evidence: what observable indicator changes the decision? Exit: when should the model or plan be revised? Applied to fast takeoff versus slow takeoff, RFE makes uncertainty operational rather than rhetorical.
The search question in fast takeoff versus slow takeoff is comparing hypotheses about the rate of capability change without treating either transition speed as established fact. A rigorous answer separates capability metric, feedback rate, bottleneck, deployment, adaptation and time. These variables can move at different speeds and can constrain one another. Super Intelligence (SI) forecasting becomes misleading when one fast-moving variable is extrapolated as though every supporting system must move with it.
Economics and Cost
The search question in fast takeoff versus slow takeoff is comparing hypotheses about the rate of capability change without treating either transition speed as established fact. A rigorous answer separates capability metric, feedback rate, bottleneck, deployment, adaptation and time. These variables can move at different speeds and can constrain one another. Super Intelligence (SI) forecasting becomes misleading when one fast-moving variable is extrapolated as though every supporting system must move with it.
First principles begin with throughput. Identify the resource or process entering the system, the transformation being attempted, and the slowest stage that determines completed output. If compute doubles but power, fabrication, verification or high-quality data does not, the realised capability gain may differ from the headline expectation. Bottleneck maps are more useful than single-variable stories.
A small-system example provides a controlled baseline. Measure how performance changes when one resource is increased while everything else is held constant. Then repeat at larger scale. If the gain diminishes, another constraint has become active. This is the basic logic behind analysing physical limits, learning bottlenecks and capability takeoff rather than assuming linear continuation.
Supply Chains and Capacity
A small-system example provides a controlled baseline. Measure how performance changes when one resource is increased while everything else is held constant. Then repeat at larger scale. If the gain diminishes, another constraint has become active. This is the basic logic behind analysing physical limits, learning bottlenecks and capability takeoff rather than assuming linear continuation.
Frontier scale changes the environment. Procurement, networking, cooling, data quality, evaluation and organisational coordination can become first-order constraints. A technique that works economically on a small experiment may become expensive or operationally fragile when multiplied across a large system. SI analysis should follow the complete stack from algorithm to deployed capability.
A changed-assumption test is essential for forecasts. If a timeline depends on compute growth, change the compute-growth assumption. If it depends on research automation, reduce the assumed automation rate. If it depends on synthetic data quality, model verification failure. A forecast that changes dramatically under modest assumption changes should be presented with correspondingly wide uncertainty.
What Current Evidence Supports
A changed-assumption test is essential for forecasts. If a timeline depends on compute growth, change the compute-growth assumption. If it depends on research automation, reduce the assumed automation rate. If it depends on synthetic data quality, model verification failure. A forecast that changes dramatically under modest assumption changes should be presented with correspondingly wide uncertainty.
Failed forecasts are evidence about models of the world, not merely embarrassing dates. Ask which assumption failed: capability scaling, economics, adoption, hardware, regulation or the definition of the milestone. Updating the causal model is more valuable than quietly moving the date. This is how forecasting becomes a learning process.
Measurement should match the variable. Power is measured differently from model loss; data diversity differently from token count; takeoff speed differently from calendar predictions. Define the numerator and denominator before comparing trends. Ambiguous metrics can create apparent acceleration simply because the measurement changed.
What Current Evidence Does Not Establish
Measurement should match the variable. Power is measured differently from model loss; data diversity differently from token count; takeoff speed differently from calendar predictions. Define the numerator and denominator before comparing trends. Ambiguous metrics can create apparent acceleration simply because the measurement changed.
Extrapolation is the dangerous step. An empirical relationship observed over one range may continue, bend or break outside it. Scaling laws, hardware trends and benchmark curves are useful precisely because they summarise observed regularities, but none is a physical guarantee of indefinite continuation. State the range and assumptions.
Uncertainty has structure. Some uncertainty comes from measurement noise; some from unknown future discoveries; some from disagreement about definitions; some from strategic or economic behaviour. Combining them into one confident date hides the reason for uncertainty. Good SI timelines show which assumption contributes most to the spread.
Connection to Super Intelligence (SI)
Uncertainty has structure. Some uncertainty comes from measurement noise; some from unknown future discoveries; some from disagreement about definitions; some from strategic or economic behaviour. Combining them into one confident date hides the reason for uncertainty. Good SI timelines show which assumption contributes most to the spread.
Feedback loops can accelerate progress when outputs become inputs to the next improvement cycle. But the loop contains latency. A software change may be tested in minutes; a new chip may require design and fabrication; an energy project may take years. Takeoff speed depends on which loop dominates and how tightly it couples to the rest of the system.
Economics determines which technically possible paths are actually pursued. Compute has a price, electricity has a price, expert supervision has a price and failed experiments consume resources. Falling unit costs can accelerate adoption, while rising marginal costs can slow it. Economic constraints should sit beside technical ones.
Safety and Preparation
Economics determines which technically possible paths are actually pursued. Compute has a price, electricity has a price, expert supervision has a price and failed experiments consume resources. Falling unit costs can accelerate adoption, while rising marginal costs can slow it. Economic constraints should sit beside technical ones.
Supply chains make intelligence physical. Advanced chips depend on specialised equipment, materials, fabrication capacity and logistics. Data centres require construction, grid connections and cooling. These facts do not disprove rapid cognitive progress; they specify the channels through which digital progress reaches scale.
Current evidence supports rapid improvement in several AI capability domains and continuing investment in compute infrastructure. It does not establish a fixed SI date, a guaranteed fast takeoff or unlimited synthetic-data improvement. The evidence is compatible with multiple future trajectories, which is why conditional reasoning is preferable to inevitability language.
Governance and Coordination
Current evidence supports rapid improvement in several AI capability domains and continuing investment in compute infrastructure. It does not establish a fixed SI date, a guaranteed fast takeoff or unlimited synthetic-data improvement. The evidence is compatible with multiple future trajectories, which is why conditional reasoning is preferable to inevitability language.
Safety preparation depends on lead time. Faster capability change can reduce the time available for evaluation, workforce adaptation and institutional response. Slower change can create more opportunity for learning but does not guarantee good governance. Preparation should therefore focus on observable indicators and adaptable controls rather than one predicted date.
Education should teach forecast literacy. A student should identify the event being forecast, the definition used, the date of the forecast, the probability assigned, the population or expert group sampled, and the assumptions underneath it. A probability is not a promise and a median survey response is not a scheduled event.
Education: Reading Claims Carefully
Education should teach forecast literacy. A student should identify the event being forecast, the definition used, the date of the forecast, the probability assigned, the population or expert group sampled, and the assumptions underneath it. A probability is not a promise and a median survey response is not a scheduled event.
For organisations, convert forecasts into scenarios rather than a single plan. Define what action would be appropriate if capability advances slowly, moderately or rapidly. Identify indicators that would trigger movement between plans. This reduces dependence on guessing the exact year and increases readiness across plausible paths.
Progress has four stages: define the variable, explain the mechanism, model sensitivity to changed assumptions, and design an update rule when evidence arrives. This is the Clementi-depth progression from recognition to independent control. Readers should finish able to revise a forecast, not merely repeat one.
Student Diagnostic Checklist
Progress has four stages: define the variable, explain the mechanism, model sensitivity to changed assumptions, and design an update rule when evidence arrives. This is the Clementi-depth progression from recognition to independent control. Readers should finish able to revise a forecast, not merely repeat one.
RFE closes the loop. Receiver: who needs the forecast or resource analysis? Function: what decision must it support? Evidence: what observable indicator changes the decision? Exit: when should the model or plan be revised? Applied to fast takeoff versus slow takeoff, RFE makes uncertainty operational rather than rhetorical.
The search question in fast takeoff versus slow takeoff is comparing hypotheses about the rate of capability change without treating either transition speed as established fact. A rigorous answer separates capability metric, feedback rate, bottleneck, deployment, adaptation and time. These variables can move at different speeds and can constrain one another. Super Intelligence (SI) forecasting becomes misleading when one fast-moving variable is extrapolated as though every supporting system must move with it.
Organisation Diagnostic Checklist
The search question in fast takeoff versus slow takeoff is comparing hypotheses about the rate of capability change without treating either transition speed as established fact. A rigorous answer separates capability metric, feedback rate, bottleneck, deployment, adaptation and time. These variables can move at different speeds and can constrain one another. Super Intelligence (SI) forecasting becomes misleading when one fast-moving variable is extrapolated as though every supporting system must move with it.
First principles begin with throughput. Identify the resource or process entering the system, the transformation being attempted, and the slowest stage that determines completed output. If compute doubles but power, fabrication, verification or high-quality data does not, the realised capability gain may differ from the headline expectation. Bottleneck maps are more useful than single-variable stories.
A small-system example provides a controlled baseline. Measure how performance changes when one resource is increased while everything else is held constant. Then repeat at larger scale. If the gain diminishes, another constraint has become active. This is the basic logic behind analysing physical limits, learning bottlenecks and capability takeoff rather than assuming linear continuation.
Progress Ladder
A small-system example provides a controlled baseline. Measure how performance changes when one resource is increased while everything else is held constant. Then repeat at larger scale. If the gain diminishes, another constraint has become active. This is the basic logic behind analysing physical limits, learning bottlenecks and capability takeoff rather than assuming linear continuation.
Frontier scale changes the environment. Procurement, networking, cooling, data quality, evaluation and organisational coordination can become first-order constraints. A technique that works economically on a small experiment may become expensive or operationally fragile when multiplied across a large system. SI analysis should follow the complete stack from algorithm to deployed capability.
A changed-assumption test is essential for forecasts. If a timeline depends on compute growth, change the compute-growth assumption. If it depends on research automation, reduce the assumed automation rate. If it depends on synthetic data quality, model verification failure. A forecast that changes dramatically under modest assumption changes should be presented with correspondingly wide uncertainty.
Counterexamples
A changed-assumption test is essential for forecasts. If a timeline depends on compute growth, change the compute-growth assumption. If it depends on research automation, reduce the assumed automation rate. If it depends on synthetic data quality, model verification failure. A forecast that changes dramatically under modest assumption changes should be presented with correspondingly wide uncertainty.
Failed forecasts are evidence about models of the world, not merely embarrassing dates. Ask which assumption failed: capability scaling, economics, adoption, hardware, regulation or the definition of the milestone. Updating the causal model is more valuable than quietly moving the date. This is how forecasting becomes a learning process.
Measurement should match the variable. Power is measured differently from model loss; data diversity differently from token count; takeoff speed differently from calendar predictions. Define the numerator and denominator before comparing trends. Ambiguous metrics can create apparent acceleration simply because the measurement changed.
What Would Change the Conclusion?
Measurement should match the variable. Power is measured differently from model loss; data diversity differently from token count; takeoff speed differently from calendar predictions. Define the numerator and denominator before comparing trends. Ambiguous metrics can create apparent acceleration simply because the measurement changed.
Extrapolation is the dangerous step. An empirical relationship observed over one range may continue, bend or break outside it. Scaling laws, hardware trends and benchmark curves are useful precisely because they summarise observed regularities, but none is a physical guarantee of indefinite continuation. State the range and assumptions.
Uncertainty has structure. Some uncertainty comes from measurement noise; some from unknown future discoveries; some from disagreement about definitions; some from strategic or economic behaviour. Combining them into one confident date hides the reason for uncertainty. Good SI timelines show which assumption contributes most to the spread.
RFE Closure
Uncertainty has structure. Some uncertainty comes from measurement noise; some from unknown future discoveries; some from disagreement about definitions; some from strategic or economic behaviour. Combining them into one confident date hides the reason for uncertainty. Good SI timelines show which assumption contributes most to the spread.
Feedback loops can accelerate progress when outputs become inputs to the next improvement cycle. But the loop contains latency. A software change may be tested in minutes; a new chip may require design and fabrication; an energy project may take years. Takeoff speed depends on which loop dominates and how tightly it couples to the rest of the system.
Economics determines which technically possible paths are actually pursued. Compute has a price, electricity has a price, expert supervision has a price and failed experiments consume resources. Falling unit costs can accelerate adoption, while rising marginal costs can slow it. Economic constraints should sit beside technical ones.
Frequently Asked Questions
Economics determines which technically possible paths are actually pursued. Compute has a price, electricity has a price, expert supervision has a price and failed experiments consume resources. Falling unit costs can accelerate adoption, while rising marginal costs can slow it. Economic constraints should sit beside technical ones.
Supply chains make intelligence physical. Advanced chips depend on specialised equipment, materials, fabrication capacity and logistics. Data centres require construction, grid connections and cooling. These facts do not disprove rapid cognitive progress; they specify the channels through which digital progress reaches scale.
Current evidence supports rapid improvement in several AI capability domains and continuing investment in compute infrastructure. It does not establish a fixed SI date, a guaranteed fast takeoff or unlimited synthetic-data improvement. The evidence is compatible with multiple future trajectories, which is why conditional reasoning is preferable to inevitability language.
Continue the Super Intelligence (SI) Series
Current evidence supports rapid improvement in several AI capability domains and continuing investment in compute infrastructure. It does not establish a fixed SI date, a guaranteed fast takeoff or unlimited synthetic-data improvement. The evidence is compatible with multiple future trajectories, which is why conditional reasoning is preferable to inevitability language.
Safety preparation depends on lead time. Faster capability change can reduce the time available for evaluation, workforce adaptation and institutional response. Slower change can create more opportunity for learning but does not guarantee good governance. Preparation should therefore focus on observable indicators and adaptable controls rather than one predicted date.
Education should teach forecast literacy. A student should identify the event being forecast, the definition used, the date of the forecast, the probability assigned, the population or expert group sampled, and the assumptions underneath it. A probability is not a promise and a median survey response is not a scheduled event.
Capacity Map: Find the Real Limiting Resource
Build a capacity map by listing every resource required for one completed unit of progress. Include compute, high-quality data, electricity, cooling, specialist labour, hardware lead time, verification and deployment. Mark current utilisation and replacement time. The limiting resource is not necessarily the most visible one. When a constraint tightens, it can become the new rate limiter even if every other input continues improving.
A scenario matrix avoids false precision. Create slow, middle and fast cases with explicit assumptions rather than vague moods. For each case, state capability growth, infrastructure expansion, research automation, verification throughput and adoption speed. The purpose is not to choose a winner; it is to see which decisions are robust across several plausible futures and which depend on one narrow forecast.
Sensitivity analysis changes one assumption while holding others fixed. Double hardware efficiency, halve data quality, delay new fabrication capacity, or increase verification cost. Observe how much the conclusion moves. The assumptions with the largest effect deserve the most monitoring. A model that swings wildly when one uncertain input changes should not be presented with a narrow confidence band.
Calibration turns forecasting into a skill that can improve. Keep dated predictions, probabilities and definitions. When the date passes or evidence arrives, score what happened and inspect whether confidence matched accuracy. Forecasting communities use calibration because a person who says 70 percent should be right roughly seven times in ten across comparable forecasts. One dramatic success is not enough to establish forecasting skill.
Scenario Matrix: Slow, Middle and Fast Cases
A scenario matrix avoids false precision. Create slow, middle and fast cases with explicit assumptions rather than vague moods. For each case, state capability growth, infrastructure expansion, research automation, verification throughput and adoption speed. The purpose is not to choose a winner; it is to see which decisions are robust across several plausible futures and which depend on one narrow forecast.
Sensitivity analysis changes one assumption while holding others fixed. Double hardware efficiency, halve data quality, delay new fabrication capacity, or increase verification cost. Observe how much the conclusion moves. The assumptions with the largest effect deserve the most monitoring. A model that swings wildly when one uncertain input changes should not be presented with a narrow confidence band.
Calibration turns forecasting into a skill that can improve. Keep dated predictions, probabilities and definitions. When the date passes or evidence arrives, score what happened and inspect whether confidence matched accuracy. Forecasting communities use calibration because a person who says 70 percent should be right roughly seven times in ten across comparable forecasts. One dramatic success is not enough to establish forecasting skill.
Decision triggers replace obsession with one calendar year. An organisation might prepare additional evaluation capacity when models cross a measured long-horizon threshold, revise permissions when agent reliability reaches a specified level, or invest in infrastructure when utilisation crosses a capacity boundary. Indicators connect preparation to observable evidence and allow plans to move as reality changes.
Sensitivity Test: Which Assumption Moves the Result Most?
Sensitivity analysis changes one assumption while holding others fixed. Double hardware efficiency, halve data quality, delay new fabrication capacity, or increase verification cost. Observe how much the conclusion moves. The assumptions with the largest effect deserve the most monitoring. A model that swings wildly when one uncertain input changes should not be presented with a narrow confidence band.
Calibration turns forecasting into a skill that can improve. Keep dated predictions, probabilities and definitions. When the date passes or evidence arrives, score what happened and inspect whether confidence matched accuracy. Forecasting communities use calibration because a person who says 70 percent should be right roughly seven times in ten across comparable forecasts. One dramatic success is not enough to establish forecasting skill.
Decision triggers replace obsession with one calendar year. An organisation might prepare additional evaluation capacity when models cross a measured long-horizon threshold, revise permissions when agent reliability reaches a specified level, or invest in infrastructure when utilisation crosses a capacity boundary. Indicators connect preparation to observable evidence and allow plans to move as reality changes.
The workbook begins with one question: what decision are you trying to make? Then list the uncertain variables, assign ranges rather than single values, identify observable indicators, create at least three scenarios and define what evidence would trigger an update. Finally, record the costs of acting too early and too late. This converts an SI timeline from a prediction contest into a decision-support tool.
Forecast Calibration: Learn From Past Predictions
Calibration turns forecasting into a skill that can improve. Keep dated predictions, probabilities and definitions. When the date passes or evidence arrives, score what happened and inspect whether confidence matched accuracy. Forecasting communities use calibration because a person who says 70 percent should be right roughly seven times in ten across comparable forecasts. One dramatic success is not enough to establish forecasting skill.
Decision triggers replace obsession with one calendar year. An organisation might prepare additional evaluation capacity when models cross a measured long-horizon threshold, revise permissions when agent reliability reaches a specified level, or invest in infrastructure when utilisation crosses a capacity boundary. Indicators connect preparation to observable evidence and allow plans to move as reality changes.
The workbook begins with one question: what decision are you trying to make? Then list the uncertain variables, assign ranges rather than single values, identify observable indicators, create at least three scenarios and define what evidence would trigger an update. Finally, record the costs of acting too early and too late. This converts an SI timeline from a prediction contest into a decision-support tool.
Constraints and forecasts must remain revisable. A new algorithm can loosen a compute bottleneck; a power shortage can tighten an infrastructure bottleneck; better synthetic-data verification can change a learning constraint. The correct model is therefore a living map of rates, buffers and evidence. Super Intelligence (SI) preparation is strongest when it can adapt without pretending uncertainty has disappeared.
Decision Triggers: Act on Indicators, Not Dates
Decision triggers replace obsession with one calendar year. An organisation might prepare additional evaluation capacity when models cross a measured long-horizon threshold, revise permissions when agent reliability reaches a specified level, or invest in infrastructure when utilisation crosses a capacity boundary. Indicators connect preparation to observable evidence and allow plans to move as reality changes.
The workbook begins with one question: what decision are you trying to make? Then list the uncertain variables, assign ranges rather than single values, identify observable indicators, create at least three scenarios and define what evidence would trigger an update. Finally, record the costs of acting too early and too late. This converts an SI timeline from a prediction contest into a decision-support tool.
Constraints and forecasts must remain revisable. A new algorithm can loosen a compute bottleneck; a power shortage can tighten an infrastructure bottleneck; better synthetic-data verification can change a learning constraint. The correct model is therefore a living map of rates, buffers and evidence. Super Intelligence (SI) preparation is strongest when it can adapt without pretending uncertainty has disappeared.
Build a capacity map by listing every resource required for one completed unit of progress. Include compute, high-quality data, electricity, cooling, specialist labour, hardware lead time, verification and deployment. Mark current utilisation and replacement time. The limiting resource is not necessarily the most visible one. When a constraint tightens, it can become the new rate limiter even if every other input continues improving.
Practical Workbook: Build an Updateable SI Model
The workbook begins with one question: what decision are you trying to make? Then list the uncertain variables, assign ranges rather than single values, identify observable indicators, create at least three scenarios and define what evidence would trigger an update. Finally, record the costs of acting too early and too late. This converts an SI timeline from a prediction contest into a decision-support tool.
Constraints and forecasts must remain revisable. A new algorithm can loosen a compute bottleneck; a power shortage can tighten an infrastructure bottleneck; better synthetic-data verification can change a learning constraint. The correct model is therefore a living map of rates, buffers and evidence. Super Intelligence (SI) preparation is strongest when it can adapt without pretending uncertainty has disappeared.
Build a capacity map by listing every resource required for one completed unit of progress. Include compute, high-quality data, electricity, cooling, specialist labour, hardware lead time, verification and deployment. Mark current utilisation and replacement time. The limiting resource is not necessarily the most visible one. When a constraint tightens, it can become the new rate limiter even if every other input continues improving.
A scenario matrix avoids false precision. Create slow, middle and fast cases with explicit assumptions rather than vague moods. For each case, state capability growth, infrastructure expansion, research automation, verification throughput and adoption speed. The purpose is not to choose a winner; it is to see which decisions are robust across several plausible futures and which depend on one narrow forecast.
Final Synthesis: Constraints and Forecasts Must Stay Revisable
Constraints and forecasts must remain revisable. A new algorithm can loosen a compute bottleneck; a power shortage can tighten an infrastructure bottleneck; better synthetic-data verification can change a learning constraint. The correct model is therefore a living map of rates, buffers and evidence. Super Intelligence (SI) preparation is strongest when it can adapt without pretending uncertainty has disappeared.
Build a capacity map by listing every resource required for one completed unit of progress. Include compute, high-quality data, electricity, cooling, specialist labour, hardware lead time, verification and deployment. Mark current utilisation and replacement time. The limiting resource is not necessarily the most visible one. When a constraint tightens, it can become the new rate limiter even if every other input continues improving.
A scenario matrix avoids false precision. Create slow, middle and fast cases with explicit assumptions rather than vague moods. For each case, state capability growth, infrastructure expansion, research automation, verification throughput and adoption speed. The purpose is not to choose a winner; it is to see which decisions are robust across several plausible futures and which depend on one narrow forecast.
Sensitivity analysis changes one assumption while holding others fixed. Double hardware efficiency, halve data quality, delay new fabrication capacity, or increase verification cost. Observe how much the conclusion moves. The assumptions with the largest effect deserve the most monitoring. A model that swings wildly when one uncertain input changes should not be presented with a narrow confidence band.
Takeoff Speed Is About the Rate of Capability Change
Fast takeoff and slow takeoff are not competing definitions of Super Intelligence (SI). They are competing models of how quickly capability might change around an important threshold such as broad human-level performance. A fast-takeoff scenario compresses the transition into a short period. A slow-takeoff scenario spreads it across enough time for competitors, regulators, institutions and infrastructure to adapt.
The strongest analysis therefore begins by choosing a measurable capability variable. Is takeoff about benchmark performance, economically useful task length, AI research automation, scientific discovery rate, or some composite? Without a metric, “fast” and “slow” remain rhetorical.
Different Capability Layers Can Take Off at Different Speeds
Software reasoning may improve rapidly while physical robotics remains slower. Coding capability may accelerate before medicine because code can be tested automatically whereas medical claims need experiments and clinical evidence. Research automation may move faster than energy infrastructure because software loops can iterate in hours while grid expansion takes years.
An SI transition can therefore be fast in one layer and slow in another. The global outcome depends on which layers are necessary for the capabilities people care about.
METR’s Time-Horizon Work Gives One Concrete Rate Measure
METR measures the length of software-related tasks that frontier AI agents can complete at stated reliability levels, where task length is defined by how long skilled humans take. Its 2026 Time Horizon 1.1 measurements continue a multi-year trend of rapidly increasing task horizons. METR also stresses that these tasks are concentrated in software engineering, machine learning and cybersecurity and should not be interpreted as a measure of whole-job automation.
This is a useful example of disciplined takeoff measurement: define the capability, fit the trend, publish uncertainty and state the limits of the task distribution.
Longer Task Horizons Do Not Translate Linearly Into Automation
METR explicitly cautions that a 50% success horizon of a given duration does not mean every shorter task can be safely delegated. Some tasks require far higher reliability. Human intervention may remain necessary. Complex failures can consume more labour than simple failures.
For SI forecasting, this means task-horizon doubling should not be converted mechanically into “jobs automated per year.” Capability growth and economic substitution are related but distinct.
Recursive Self-Improvement Can Change the Takeoff Rate
If AI systems increasingly contribute to AI research, then capability growth can feed back into the process that creates future capability. Anthropic’s 2026 reporting on recursive self-improvement argues that AI is already accelerating parts of AI development while also stating that fully autonomous successor design has not yet been reached and is not inevitable.
The takeoff question therefore becomes empirical: how much of frontier R&D is automated, how productive is that automation, and does each generation meaningfully improve the research process for the next?
Fast Software Loops Can Still Hit Slow Physical Bottlenecks
A research agent can revise code quickly. It cannot instantly produce new HBM capacity, grid interconnections or fabrication plants. The IEA’s energy work shows that power infrastructure and data-centre construction operate on different time scales from model development.
This creates a natural brake on some fast-takeoff stories. Even if software capability accelerates, the deployment of additional compute may remain physically rate-limited.
Verification Can Be the Hidden Takeoff Bottleneck
Generating ideas is easier than proving they are improvements. In code and mathematics, verification can be automated. In science, policy and strategy, the result may require experiments, replication or human judgement. If proposal generation accelerates faster than validation, the system can produce a backlog of untrusted claims rather than a capability explosion.
The relevant rate is therefore verified improvement per unit time, not raw output per unit time.
Competition Can Make Slow Technical Change Feel Fast Socially
Even gradual capability improvement can create rapid adoption if multiple firms race to deploy similar systems. Conversely, a technically fast improvement can spread slowly if integration is expensive or access is restricted. Social takeoff and technical takeoff should therefore be modelled separately.
For organisations, the practical question is not only how fast frontier models improve but how fast those improvements arrive inside the workflow that matters.
Worked Example: A Six-Month Software Takeoff
Imagine coding agents become capable of completing substantially longer and more complex software projects within six months. AI labs use those agents to accelerate internal research tooling. Capability rises quickly in software and AI R&D, but hardware supply and enterprise deployment lag. This would be a fast takeoff in one cognitive domain without an instantaneous economy-wide transformation.
The scenario is more useful when each clock is stated explicitly.
Worked Example: A Five-Year Broad Takeoff
Now imagine broad general capability improves steadily over five years, while businesses redesign processes, universities change curricula, governments update rules and physical infrastructure expands in parallel. The technical rate is slower, but the total social transition may be more orderly because adaptation occurs during the climb.
Slow takeoff is not necessarily low impact. It means more of the transition happens while institutions have time to observe and respond.
What Evidence Would Move Us Toward a Faster-Takeoff Model?
Evidence would include rapidly rising AI contribution to AI R&D, shortening model-development cycles, increasing success on long-horizon tasks, strong transfer across domains, scalable automated verification and the ability to use existing hardware much more efficiently. A recursive loop that measurably shortens its own iteration time would be especially important.
The evidence should be dated and updated. Takeoff models are forecasts, not identities.
What Evidence Would Move Us Toward a Slower-Takeoff Model?
Evidence would include persistent reliability plateaus, weak transfer, rising marginal compute costs, evaluator bottlenecks, slow physical experimentation, infrastructure delays and domains where tacit knowledge or embodied interaction remains difficult to automate.
A good forecast changes when the dominant bottleneck changes.
RFE Closure: Measure the Rate, the Layer and the Bottleneck
The problem is treating “fast” and “slow” as stories rather than measurable models. The function of takeoff analysis is to estimate how quickly verified capability changes and how that rate propagates through research, deployment and society. The receiver is the decision-maker deciding how much adaptation time may be available.
The exit condition is to replace the takeoff model whenever observed capability growth, research automation or infrastructure constraints diverge materially from its assumptions. SI forecasting should be continuously recalibrated.