Here, you will find the distributions of results that come from sampling the parameters according to Tom Davidson's beliefs. The median training requirements for AGI are ~1e36 FLOP using 2022 algorithms.

Probability of full economic automation before 2100 : 89%

Probability of slow takeoff : 74%

Show

Quantile Powerful sub-AGI year AGI year 20% automation year 100% automation year 20% R&D automation year 100% R&D automation year Wake-up year 20-100% economic automation 20-100% R&D automation Powerful sub-AGI to AGI 2X to 10X cognitive output multiplier 5% to 20% GWP growth
0.01 2024.0 2025.2 2024.8 2025.7 2024.2 2025.6 2024.0 0.3 0.9 0.1 0.4 0.4
0.10 2026.9 2029.1 2027.7 2029.6 2026.8 2029.4 2026.6 0.8 1.6 0.4 0.7 0.7
0.20 2029.5 2032.2 2030.0 2032.7 2028.9 2032.4 2028.7 1.2 2.2 0.9 1.0 1.0
0.50 2038.8 2043.2 2038.3 2043.3 2036.7 2042.9 2036.6 2.9 4.3 2.4 2.1 2.4
0.80 2062.3 2070.6 2058.8 2070.3 2055.5 2069.7 2055.3 7.6 9.6 6.5 5.3 6.6
0.90 2095.0 ≥ 2100 2088.2 ≥ 2100 2082.2 ≥ 2100 2083.1 12.5 14.6 10.9 8.6 12.2
0.99 ≥ 2100 ≥ 2100 ≥ 2100 ≥ 2100 ≥ 2100 ≥ 2100 ≥ 2100 28.0 30.7 27.0 18.5 30.7
mean 2047.3 2051.4 2046.5 2051.5 2044.6 2051.2 2044.6 5.1 6.5 4.3 3.6 4.7

Show

Show

Timelines vs takeoff

"Years before full economic automation" tables

Percentile At time of full economic automation 1 year before 2 years before 5 years before 10 years before
1% 0.2 0.2 0.8 2.2 3.9
10% 0.7 0.3 1.2 3.3 6.7
20% 1.2 0.3 1.5 4.3 9.3
50% 3.7 0.7 2.4 8.4 21.1
80% 9.0 3.0 4.8 21.9 23.8
90% 10.9 6.3 7.3 23.7 25.3
99% 13.8 11.8 18.5 27.1 28.4
Percentile At time of full economic automation 1 year before 2 years before 5 years before 10 years before
1% 0.1 0.1 0.4 0.9 1.5
10% 0.1 0.2 0.6 1.4 2.2
20% 0.1 0.2 0.8 1.9 2.5
50% 0.2 0.3 1.2 2.7 3.1
80% 0.4 0.6 1.9 3.5 4.2
90% 0.5 0.8 2.3 4.1 4.8
99% 1.2 1.4 3.5 5.4 6.4
Percentile At time of full economic automation 1 year before 2 years before 5 years before 10 years before
1% 0.0 0.0 0.4 0.8 0.7
10% 0.0 0.1 0.6 1.0 1.0
20% 0.1 0.1 0.8 1.3 1.4
50% 0.1 0.2 1.2 2.4 2.9
80% 0.1 0.4 1.8 3.6 4.2
90% 0.3 0.5 2.3 4.3 5.3
99% N/A 1.2 4.1 8.2 10.9

N/A: at physical limit

Percentile At time of full economic automation 1 year before 2 years before 5 years before 10 years before
1% 3.6e+29 2.6e+26 3.5e+25 3.2e+23 8.0e+19
10% 5.2e+31 6.2e+28 5.5e+27 5.4e+25 9.2e+22
20% 9.6e+32 1.4e+30 1.5e+29 1.6e+27 3.0e+24
50% 1.1e+36 1.7e+33 2.5e+32 6.3e+30 6.5e+28
80% 8.6e+39 9.6e+36 1.9e+36 7.6e+34 1.7e+33
90% 1.2e+42 6.9e+38 1.6e+38 9.7e+36 3.4e+35
99% 1.1e+48 5.9e+42 1.6e+42 1.0e+41 4.2e+39
Percentile At time of full economic automation 1 year before 2 years before 5 years before 10 years before
1% 0.08% 0.01% 0.01% 0.00% 0.00%
10% 0.95% 0.18% 0.06% 0.02% 0.00%
20% 3.10% 0.68% 0.23% 0.06% 0.01%
50% 6.60% 4.76% 3.64% 0.79% 0.21%
80% 12.74% 10.85% 9.42% 6.03% 3.52%
90% 15.45% 13.99% 12.73% 9.81% 5.83%
99% 19.25% 18.72% 18.35% 16.50% 13.57%
Percentile At time of full economic automation 1 year before 2 years before 5 years before 10 years before
1% 0.05% 0.04% 0.03% 0.02% 0.01%
10% 0.22% 0.18% 0.15% 0.09% 0.04%
20% 0.48% 0.39% 0.32% 0.19% 0.08%
50% 2.86% 2.34% 1.92% 1.08% 0.44%
80% 7.25% 6.72% 6.30% 5.09% 3.73%
90% 11.22% 10.51% 9.81% 8.26% 6.52%
99% 23.46% 22.69% 21.80% 20.62% 18.40%

Inputs

Number of samples: 10000

Rank correlations: click here to view.

Input statistics: click here to view.

Show

Conservative Best guess Aggressive
AGI training requirements (FLOP with 2022 algorithms) sampled from Cotra's distribution (click here to view)
Effective FLOP gap (training) 1.000000e+08 1.000000e+04 1.000000e+01
Training goods vs R&D 1.000000e+00 3.000000e+00 3.000000e+01
AGI runtime requirements 3.153600e+27 5.256000e+23 3.153600e+22
Effective FLOP gap (runtime) 1.000000e+02 3.000000e+00 1.000000e+00
Runtime goods vs R&D 3.000000e+00 3.000000e+01 1.000000e+02
Trade-off efficiency 3.000000e+00 1.500000e+00 8.000000e-01
Maximum trade-off 1.000000e+00 3.000000e+01 1.000000e+03
Labour substitution goods -2.000000e+00 -5.000000e-01 -2.000000e-01
Labour substitution R&D -2.000000e+00 -5.000000e-01 -2.000000e-01
Substitution between capital and cognitive tasks for goods and services -2.000000e+00 -4.000000e-01 -2.000000e-01
Substitution between capital and cognitive tasks for hardware R&D -5.000000e-01 -2.500000e-01 -1.700000e-01
Research experiments substitution software -1.000000e-02
Efficiency of experiments for software R&D 4.000000e-01
Returns to hardware 3.000000e+00 5.200000e+00 7.000000e+00
Returns to software 8.000000e-01 1.250000e+00 5.000000e+00
Maximum hardware performance 1.000000e+23 1.000000e+26 1.000000e+29
Maximum software performance 1.000000e+08 1.000000e+12 1.000000e+16
R&D parallelization penalty 2.000000e-01 7.000000e-01 9.900000e-01
Hardware adoption delay 2.500000e+00 1.000000e+00 1.000000e-02
Growth rate fraction capital hardware R&D 1.000000e-02
Growth rate fraction labour hardware R&D 1.000000e-02
Growth rate fraction compute hardware R&D 1.000000e-02
Growth rate fraction labour software R&D 1.800000e-01
Growth rate fraction compute software R&D 1.800000e-01
Growth rate fraction GWP compute 7.000000e-02 1.900000e-01 3.700000e-01
Growth rate fraction compute training 9.312925e-01 5.475284e-01 1.637642e-01
Wake-up growth rate fraction capital hardware R&D 5.000000e-02 1.400000e-01 3.200000e-01
Wake-up growth rate fraction labour hardware R&D 5.000000e-02 1.400000e-01 3.200000e-01
Wake-up growth rate fraction compute hardware R&D 2.700000e-01 6.700000e-01 1.370000e+00
Wake-up growth rate fraction labour software R&D 1.200000e-01 2.200000e-01 3.700000e-01
Wake-up growth rate fraction of compute software R&D 2.700000e-01 6.700000e-01 1.370000e+00
Wake-up growth rate fraction of GWP buying compute 7.000000e-02 1.900000e-01 3.700000e-01
Wake-up growth rate fraction compute training AI models 5.500000e-01 1.100000e+00 4.500000e+00
Max fraction capital hardware R&D 1.000000e-02 3.000000e-02 1.000000e-01
Max fraction of labour dedicated to hardware R&D 1.000000e-02 3.000000e-02 1.000000e-01
Max fraction of compute dedicated to hardware R&D 5.000000e-02 2.000000e-01 3.500000e-01
Max fraction labour software R&D 1.000000e-02 3.000000e-02 1.000000e-01
Max fraction compute software R&D 5.000000e-02 2.000000e-01 3.500000e-01
Max fraction GWP compute 3.000000e-02 1.000000e-01 3.000000e-01
Max fraction compute training 3.000000e-02 1.000000e-01 2.000000e-01
Initial fraction capital hardware R&D 2.000000e-03
Initial fraction labour hardware R&D 2.000000e-03
Initial fraction compute hardware R&D 2.000000e-03
Initial fraction labour software R&D 2.000000e-04
Initial fraction compute software R&D 2.000000e-04
Initial biggest training run 3.000000e+24
Initial vs cumulative input - hardware R&D 4.700000e-02
Initial vs cumulative input - software R&D 2.000000e-01
Initial hardware production 1.000000e+27 1.000000e+28 1.000000e+29
Accumulated hardware vs initial hardware production 2.000000e+00
Initial market hardware performance 1.500000e+17
Initial GWP 8.500000e+13
Initial world labour force. 4.000000e+09
Initial cognitive share goods 3.000000e-01 5.000000e-01 6.500000e-01
Initial cognitive share in hardware R&D 7.000000e-01
Initial compute share goods 1.000000e-02
Initial compute share R&D 1.000000e-02
Initial experiment share software R&D 3.000000e-01
Wakeup trigger 1.000000e-02 6.000000e-02 2.000000e-01
Initial capital growth rate 2.750000e-02
Population growth rate 1.000000e-02
TFP growth rate 1.000000e-02
Compute depreciation rate 1.000000e-01 2.000000e-01 3.000000e-01
Money threshold training 2.000000e+10 4.000000e+09 4.000000e+08
Training requirements steepness (OOM) 0.000000e+00
Runtime requirements steepness (OOM) 0.000000e+00

Here, you will find the distributions of results that come from sampling the parameters according to Tom Davidson's beliefs but with an aggressive distribution for the amount of FLOP required to train an AGI. The median training requirements for AGI are ~1e31 FLOP using 2022 algorithms.

Probability of full economic automation before 2100 : 97%

Probability of slow takeoff : 49%

Show

Quantile Powerful sub-AGI year AGI year 20% automation year 100% automation year 20% R&D automation year 100% R&D automation year Wake-up year 20-100% economic automation 20-100% R&D automation Powerful sub-AGI to AGI 2X to 10X cognitive output multiplier 5% to 20% GWP growth
0.01 2023.1 2024.4 2024.3 2024.8 2023.2 2024.7 2023.7 0.2 1.1 0.2 0.5 0.3
0.10 2024.6 2026.3 2026.1 2027.0 2024.6 2026.8 2025.3 0.5 1.7 0.7 0.8 0.6
0.20 2025.8 2027.8 2027.3 2028.6 2025.6 2028.4 2026.5 0.7 2.2 1.1 1.0 0.8
0.50 2030.0 2032.9 2031.4 2033.7 2029.0 2033.2 2030.2 1.7 3.7 2.2 1.6 1.5
0.80 2038.9 2043.7 2039.9 2044.1 2036.4 2043.3 2038.1 3.9 6.4 4.6 2.9 3.3
0.90 2048.2 2054.7 2047.9 2054.9 2043.5 2053.7 2045.8 6.1 9.1 7.0 4.4 5.6
0.99 ≥ 2100 ≥ 2100 ≥ 2100 ≥ 2100 ≥ 2100 ≥ 2100 ≥ 2100 20.0 24.2 19.7 13.3 19.0
mean 2034.8 2038.1 2035.8 2038.7 2033.2 2038.2 2034.5 2.9 4.9 3.4 2.3 2.6

Show

Show

Timelines vs takeoff

"Years before full economic automation" tables

Percentile At time of full economic automation 1 year before 2 years before 5 years before 10 years before
1% 0.2 0.2 0.8 2.6 5.1
10% 0.7 0.3 1.2 4.5 12.1
20% 1.2 0.3 1.6 6.4 20.0
50% 3.9 0.7 2.9 19.9 21.5
80% 9.9 2.1 6.3 22.8 23.3
90% 11.6 4.9 10.4 23.9 24.5
99% 14.2 12.0 21.5 26.2 27.0
Percentile At time of full economic automation 1 year before 2 years before 5 years before 10 years before
1% 0.1 0.1 0.4 1.1 1.9
10% 0.1 0.2 0.6 1.9 2.3
20% 0.1 0.2 0.8 2.2 2.4
50% 0.2 0.3 1.2 2.7 3.0
80% 0.4 0.5 1.8 3.5 4.1
90% 0.5 0.7 2.1 4.0 4.6
99% 1.0 1.1 3.0 4.8 5.5
Percentile At time of full economic automation 1 year before 2 years before 5 years before 10 years before
1% 0.0 0.0 0.5 0.7 0.7
10% 0.0 0.1 0.6 1.0 1.0
20% 0.0 0.1 0.7 1.3 1.3
50% 0.1 0.2 1.0 2.3 2.8
80% 0.1 0.3 1.4 3.2 3.8
90% 0.3 0.5 1.7 3.6 4.2
99% 0.8 1.0 2.5 4.8 6.8
Percentile At time of full economic automation 1 year before 2 years before 5 years before 10 years before
1% 3.8e+28 4.9e+25 9.2e+24 9.1e+22 1.5e+19
10% 1.3e+30 1.2e+27 1.7e+26 3.0e+24 3.3e+21
20% 8.9e+30 1.1e+28 1.4e+27 2.1e+25 4.0e+22
50% 1.3e+33 1.4e+30 1.9e+29 4.8e+27 1.4e+25
80% 2.9e+36 1.4e+33 2.0e+32 5.8e+30 1.0e+29
90% 3.2e+38 1.2e+35 1.9e+34 6.2e+32 1.2e+31
99% 5.1e+43 7.7e+39 2.0e+39 1.3e+38 8.0e+36
Percentile At time of full economic automation 1 year before 2 years before 5 years before 10 years before
1% 0.02% 0.01% 0.00% 0.00% 0.00%
10% 0.16% 0.04% 0.02% 0.01% 0.00%
20% 0.47% 0.14% 0.06% 0.02% 0.00%
50% 3.40% 1.67% 0.74% 0.19% 0.03%
80% 7.14% 5.25% 4.34% 3.00% 0.73%
90% 10.44% 8.07% 6.71% 4.43% 3.12%
99% 16.89% 15.59% 14.16% 11.64% 8.56%
Percentile At time of full economic automation 1 year before 2 years before 5 years before 10 years before
1% 0.03% 0.03% 0.02% 0.01% 0.00%
10% 0.09% 0.08% 0.07% 0.04% 0.02%
20% 0.18% 0.15% 0.13% 0.08% 0.03%
50% 0.77% 0.66% 0.56% 0.33% 0.13%
80% 3.18% 2.91% 2.46% 1.44% 0.57%
90% 4.99% 4.61% 4.28% 3.45% 1.88%
99% 14.70% 13.82% 13.01% 10.90% 8.28%

Inputs

Number of samples: 10000

Rank correlations: click here to view.

Input statistics: click here to view.

Show

Conservative Best guess Aggressive
AGI training requirements (FLOP with 2022 algorithms) sampled from an aggressive distribution (click here to view)
Effective FLOP gap (training) 1.000000e+08 1.000000e+04 1.000000e+01
Training goods vs R&D 1.000000e+00 3.000000e+00 3.000000e+01
AGI runtime requirements 1.000000e+20 1.666670e+16 1.000000e+15
Effective FLOP gap (runtime) 1.000000e+02 1.000000e+01 1.000000e+00
Runtime goods vs R&D 3.000000e+01 1.000000e+02 3.000000e+02
Trade-off efficiency 3.000000e+00 1.500000e+00 8.000000e-01
Maximum trade-off 1.000000e+00 3.000000e+01 1.000000e+03
Labour substitution goods -2.000000e+00 -5.000000e-01 -2.000000e-01
Labour substitution R&D -2.000000e+00 -5.000000e-01 -2.000000e-01
Substitution between capital and cognitive tasks for goods and services -2.000000e+00 -4.000000e-01 -2.000000e-01
Substitution between capital and cognitive tasks for hardware R&D -5.000000e-01 -2.500000e-01 -1.700000e-01
Research experiments substitution software -1.000000e-02
Efficiency of experiments for software R&D 4.000000e-01
Returns to hardware 3.000000e+00 5.200000e+00 7.000000e+00
Returns to software 8.000000e-01 1.250000e+00 5.000000e+00
Maximum hardware performance 1.000000e+23 1.000000e+26 1.000000e+29
Maximum software performance 1.000000e+08 1.000000e+12 1.000000e+16
R&D parallelization penalty 2.000000e-01 7.000000e-01 9.900000e-01
Hardware adoption delay 2.500000e+00 1.000000e+00 1.000000e-02
Growth rate fraction capital hardware R&D 1.000000e-02
Growth rate fraction labour hardware R&D 1.000000e-02
Growth rate fraction compute hardware R&D 1.000000e-02
Growth rate fraction labour software R&D 1.800000e-01
Growth rate fraction compute software R&D 1.800000e-01
Growth rate fraction GWP compute 7.000000e-02 1.900000e-01 3.700000e-01
Growth rate fraction compute training 9.312925e-01 5.475284e-01 1.637642e-01
Wake-up growth rate fraction capital hardware R&D 5.000000e-02 1.400000e-01 3.200000e-01
Wake-up growth rate fraction labour hardware R&D 5.000000e-02 1.400000e-01 3.200000e-01
Wake-up growth rate fraction compute hardware R&D 2.700000e-01 6.700000e-01 1.370000e+00
Wake-up growth rate fraction labour software R&D 1.200000e-01 2.200000e-01 3.700000e-01
Wake-up growth rate fraction of compute software R&D 2.700000e-01 6.700000e-01 1.370000e+00
Wake-up growth rate fraction of GWP buying compute 7.000000e-02 1.900000e-01 3.700000e-01
Wake-up growth rate fraction compute training AI models 5.500000e-01 1.100000e+00 4.500000e+00
Max fraction capital hardware R&D 1.000000e-02 3.000000e-02 1.000000e-01
Max fraction of labour dedicated to hardware R&D 1.000000e-02 3.000000e-02 1.000000e-01
Max fraction of compute dedicated to hardware R&D 5.000000e-02 2.000000e-01 3.500000e-01
Max fraction labour software R&D 1.000000e-02 3.000000e-02 1.000000e-01
Max fraction compute software R&D 5.000000e-02 2.000000e-01 3.500000e-01
Max fraction GWP compute 3.000000e-02 1.000000e-01 3.000000e-01
Max fraction compute training 3.000000e-02 1.000000e-01 2.000000e-01
Initial fraction capital hardware R&D 2.000000e-03
Initial fraction labour hardware R&D 2.000000e-03
Initial fraction compute hardware R&D 2.000000e-03
Initial fraction labour software R&D 2.000000e-04
Initial fraction compute software R&D 2.000000e-04
Initial biggest training run 3.000000e+24
Initial vs cumulative input - hardware R&D 4.700000e-02
Initial vs cumulative input - software R&D 2.000000e-01
Initial hardware production 1.000000e+27 1.000000e+28 1.000000e+29
Accumulated hardware vs initial hardware production 2.000000e+00
Initial market hardware performance 1.500000e+17
Initial GWP 8.500000e+13
Initial world labour force. 4.000000e+09
Initial cognitive share goods 3.000000e-01 5.000000e-01 6.500000e-01
Initial cognitive share in hardware R&D 7.000000e-01
Initial compute share goods 1.000000e-02
Initial compute share R&D 1.000000e-02
Initial experiment share software R&D 3.000000e-01
Wakeup trigger 1.000000e-02 6.000000e-02 2.000000e-01
Initial capital growth rate 2.750000e-02
Population growth rate 1.000000e-02
TFP growth rate 1.000000e-02
Compute depreciation rate 1.000000e-01 2.000000e-01 3.000000e-01
Money threshold training 2.000000e+10 4.000000e+09 4.000000e+08
Training requirements steepness (OOM) 0.000000e+00
Runtime requirements steepness (OOM) 0.000000e+00

Here, you will find a sensitivity analysis we run, highlighting the most important parameters for pinning down the results of the model.

One-at-a-time

Show

Importance Variance reduction Powerful sub-AGI year AGI year Skew of time to AGI date 20% automation year 100% automation year 20% R&D automation year 100% R&D automation year Wake-up year 20-100% economic automation Skew of 20-100% economic automation 20-100% R&D automation Powerful sub-AGI to AGI 2X to 10X cognitive output multiplier 5% to 20% GWP growth Pattern of GWP doublings
Effective FLOP gap (training) 5.3 41% [2027.30, 2037.20, 2048.50] [2034.40, 2040.10, 2048.60] -2.8 [2028.10, 2036.20, 2047.80] [2034.30, 2040.00, 2048.60] [2026.30, 2034.70, 2046.90] [2034.20, 2039.80, 2048.30] [2025.20, 2034.40, 2047.30] [6.10, 3.70, 0.80] -0.5 [7.80, 5.00, 1.40] [7.10, 2.90, 0.10] [4.20, 2.90, 0.70] [7.40, 2.90, 0.70] [[21.99, 5.3, 2.7, 1.1], [21.99, 3.8, 1.5, 0.8], [21.99, 0.9, 0.2]]
AGI training requirements (FLOP with 2022 algorithms) 4.1 9% [2052.70, 2037.20, 2029.80] [2056.40, 2040.10, 2033.30] 9.5 [2049.00, 2036.20, 2030.20] [2056.30, 2040.00, 2033.40] [2048.30, 2034.70, 2028.50] [2056.20, 2039.80, 2033.00] [2047.10, 2034.40, 2028.60] [7.20, 3.70, 3.10] 2.9 [7.80, 5.00, 4.30] [3.70, 2.90, 3.50] [5.30, 2.90, 2.10] [3.70, 2.90, 2.70] [[21.99, 4.2, 3.3, 1.5], [21.99, 3.8, 1.5, 0.8], [21.99, 3.3, 1.1, 1.0]]
R&D parallelization penalty 4.0 8% [2037.50, 2037.20, 2037.00] [2043.80, 2040.10, 2039.30] 2.9 [2036.30, 2036.20, 2036.10] [2043.50, 2040.00, 2039.30] [2034.70, 2034.70, 2034.70] [2042.10, 2039.80, 2039.10] [2034.40, 2034.40, 2034.40] [7.10, 3.70, 3.10] 2.8 [7.30, 5.00, 4.30] [6.30, 2.90, 2.30] [3.60, 2.90, 2.70] [5.50, 2.90, 2.80] [[21.98, 4.4, 2.7, 1.4], [21.99, 3.8, 1.5, 0.8], [21.99, 3.5, 1.2]]
Returns to software 3.0 10% [2039.40, 2037.20, 2030.30] [2042.90, 2040.10, 2031.40] -5.9 [2038.10, 2036.20, 2029.60] [2042.80, 2040.00, 2031.30] [2036.50, 2034.70, 2028.80] [2042.40, 2039.80, 2031.20] [2036.20, 2034.40, 2028.50] [4.60, 3.70, 1.60] -1.2 [5.80, 5.00, 2.40] [3.50, 2.90, 1.10] [3.40, 2.90, 1.40] [4.50, 2.90, 1.50] [[22.3, 4.2, 1.8, 0.5], [21.99, 3.8, 1.5, 0.8], [19.76, 2.2, 0.5]]
Labour substitution R&D 2.2 5% [2037.50, 2037.20, 2036.90] [2041.50, 2040.10, 2039.10] 0.4 [2036.30, 2036.20, 2036.10] [2041.50, 2040.00, 2039.00] [2034.80, 2034.70, 2034.60] [2041.20, 2039.80, 2038.80] [2034.50, 2034.40, 2034.30] [5.10, 3.70, 2.90] 0.6 [6.30, 5.00, 4.10] [4.00, 2.90, 2.20] [3.90, 2.90, 2.10] [4.60, 2.90, 2.70] [[21.99, 4.2, 2.2, 0.6], [21.99, 3.8, 1.5, 0.8], [21.99, 3.4, 1.0, 0.7]]
Training goods vs R&D 2.1 4% [2037.40, 2037.20, 2036.60] [2041.00, 2040.10, 2038.40] -0.8 [2036.30, 2036.20, 2035.80] [2041.00, 2040.00, 2038.40] [2035.40, 2034.70, 2032.30] [2040.80, 2039.80, 2038.00] [2034.50, 2034.40, 2034.10] [4.60, 3.70, 2.50] -0.3 [5.30, 5.00, 5.50] [3.60, 2.90, 1.80] [3.40, 2.90, 2.20] [3.30, 2.90, 2.60] [[21.99, 4.0, 1.8, 0.7], [21.99, 3.8, 1.5, 0.8], [21.99, 3.2, 0.8, 1.0]]
Effective FLOP gap (runtime) 1.6 0% [2036.50, 2037.20, 2038.40] [2039.20, 2040.10, 2041.10] -0.1 [2034.50, 2036.20, 2038.00] [2039.20, 2040.00, 2041.10] [2033.50, 2034.70, 2036.00] [2039.00, 2039.80, 2040.80] [2032.50, 2034.40, 2036.40] [4.60, 3.70, 3.00] 0.2 [5.40, 5.00, 4.70] [2.70, 2.90, 2.70] [3.50, 2.90, 2.30] [4.70, 2.90, 2.40] [[21.99, 4.3, 1.9, 0.5], [21.99, 3.8, 1.5, 0.8], [21.99, 3.2, 1.1, 0.5]]
Maximum trade-off 1.6 2% [2039.10, 2037.20, 2037.10] [2044.20, 2040.10, 2039.70] 3.7 [2039.10, 2036.20, 2036.20] [2044.20, 2040.00, 2039.70] [2038.50, 2034.70, 2034.40] [2044.10, 2039.80, 2039.30] [2037.30, 2034.40, 2034.40] [5.00, 3.70, 3.40] 1.0 [5.50, 5.00, 4.80] [5.10, 2.90, 2.60] [3.60, 2.90, 2.50] [3.50, 2.90, 2.90] [[21.99, 3.6, 2.3, 0.9], [21.99, 3.8, 1.5, 0.8], [21.99, 3.7, 1.2, 0.9]]
Max fraction compute training 1.1 0% [2038.40, 2037.20, 2037.00] [2041.10, 2040.10, 2039.60] 0.5 [2036.50, 2036.20, 2036.20] [2041.10, 2040.00, 2039.60] [2034.70, 2034.70, 2034.70] [2040.90, 2039.80, 2039.30] [2034.40, 2034.40, 2034.40] [4.40, 3.70, 3.30] 0.3 [6.10, 5.00, 4.50] [2.70, 2.90, 2.60] [4.00, 2.90, 2.40] [4.60, 2.90, 2.90] [[21.99, 4.5, 1.7, 0.5], [21.99, 3.8, 1.5, 0.8], [21.99, 3.5, 1.3, 0.6]]
Returns to hardware 1.1 3% [2039.60, 2037.20, 2035.70] [2043.00, 2040.10, 2038.30] 1.1 [2038.40, 2036.20, 2034.80] [2043.00, 2040.00, 2038.30] [2036.80, 2034.70, 2033.40] [2042.70, 2039.80, 2038.00] [2036.40, 2034.40, 2033.10] [4.50, 3.70, 3.40] 0.5 [5.80, 5.00, 4.50] [3.40, 2.90, 2.60] [3.40, 2.90, 2.70] [4.40, 2.90, 2.60] [[22.01, 4.3, 1.8, 0.5], [21.99, 3.8, 1.5, 0.8], [21.97, 3.5, 1.3, 0.6]]
Initial hardware production 1.0 3% [2040.20, 2037.20, 2036.10] [2043.10, 2040.10, 2038.00] 0.9 [2038.90, 2036.20, 2034.80] [2043.10, 2040.00, 2038.00] [2035.80, 2034.70, 2033.80] [2042.80, 2039.80, 2037.70] [2035.90, 2034.40, 2033.00] [4.10, 3.70, 3.10] -0.2 [6.80, 5.00, 3.80] [2.90, 2.90, 1.90] [4.00, 2.90, 2.20] [5.30, 2.90, 3.00] [[21.99, 5.2, 1.6, 0.4], [21.99, 3.8, 1.5, 0.8], [21.99, 3.4, 1.2, 0.7]]
Hardware adoption delay 1.0 0% [2037.20, 2037.20, 2037.10] [2040.50, 2040.10, 2039.50] -0.2 [2036.10, 2036.20, 2036.20] [2040.40, 2040.00, 2039.50] [2034.60, 2034.70, 2034.70] [2040.10, 2039.80, 2039.30] [2034.30, 2034.40, 2034.40] [4.20, 3.70, 3.20] 0.0 [5.40, 5.00, 4.50] [3.30, 2.90, 2.40] [3.10, 2.90, 2.60] [2.90, 2.90, 2.80] [[21.99, 3.9, 1.8, 0.5], [21.99, 3.8, 1.5, 0.8], [21.99, 3.6, 1.2, 0.8]]
Growth rate fraction GWP compute 1.0 1% [2039.80, 2037.20, 2034.90] [2042.80, 2040.10, 2037.30] -0.1 [2038.40, 2036.20, 2033.90] [2042.80, 2040.00, 2037.30] [2036.60, 2034.70, 2032.70] [2042.50, 2039.80, 2037.00] [2036.40, 2034.40, 2032.30] [4.30, 3.70, 3.30] 0.2 [5.80, 5.00, 4.20] [3.00, 2.90, 2.40] [3.50, 2.90, 2.30] [4.20, 2.90, 2.70] [[22.01, 4.4, 1.6, 0.5], [21.99, 3.8, 1.5, 0.8], [21.96, 3.2, 1.4, 0.7]]
Wake-up growth rate fraction of GWP buying compute 0.9 0% [2037.40, 2037.20, 2036.90] [2040.60, 2040.10, 2039.50] -0.1 [2036.30, 2036.20, 2036.10] [2040.50, 2040.00, 2039.40] [2034.70, 2034.70, 2034.70] [2040.30, 2039.80, 2039.20] [2034.40, 2034.40, 2034.40] [4.10, 3.70, 3.20] -0.1 [5.50, 5.00, 4.40] [3.20, 2.90, 2.60] [3.20, 2.90, 2.50] [4.30, 2.90, 2.70] [[21.99, 4.1, 1.6, 0.5], [21.99, 3.8, 1.5, 0.8], [21.99, 3.4, 1.3, 0.7]]
Growth rate fraction compute training 0.9 0% [2036.90, 2037.20, 2039.00] [2039.80, 2040.10, 2041.10] -0.7 [2035.60, 2036.20, 2037.80] [2039.80, 2040.00, 2041.10] [2034.00, 2034.70, 2036.70] [2039.50, 2039.80, 2040.80] [2033.80, 2034.40, 2036.10] [4.10, 3.70, 3.20] -0.1 [5.40, 5.00, 4.00] [2.90, 2.90, 2.10] [3.30, 2.90, 2.20] [4.00, 2.90, 2.90] [[21.99, 4.1, 1.5, 0.6], [21.99, 3.8, 1.5, 0.8], [21.99, 3.3, 1.3, 0.7]]
Money threshold training 0.9 1% [2036.80, 2037.20, 2038.90] [2039.60, 2040.10, 2041.10] -0.5 [2035.40, 2036.20, 2037.80] [2039.60, 2040.00, 2041.00] [2033.30, 2034.70, 2036.70] [2039.30, 2039.80, 2040.70] [2033.30, 2034.40, 2036.00] [4.00, 3.70, 3.10] -0.3 [5.80, 5.00, 3.90] [2.80, 2.90, 2.20] [3.60, 2.90, 2.20] [4.30, 2.90, 3.00] [[21.99, 4.4, 1.5, 0.4], [21.99, 3.8, 1.5, 0.8], [21.99, 3.4, 1.3, 0.9]]
AGI runtime requirements 0.8 2% [2040.90, 2037.20, 2036.50] [2045.30, 2040.10, 2039.10] 4.2 [2041.80, 2036.20, 2034.50] [2045.50, 2040.00, 2039.00] [2040.00, 2034.70, 2033.60] [2045.10, 2039.80, 2038.90] [2040.00, 2034.40, 2032.70] [3.60, 3.70, 4.40] -0.6 [5.00, 5.00, 5.20] [4.40, 2.90, 2.60] [2.50, 2.90, 3.30] [2.90, 2.90, 3.00] [[21.99, 3.8, 1.3, 0.5], [21.99, 3.8, 1.5, 0.8], [21.99, 3.9, 1.8, 0.6]]
Labour substitution goods 0.6 0% [2037.20, 2037.20, 2037.10] [2040.20, 2040.10, 2039.90] -0.1 [2036.00, 2036.20, 2036.30] [2040.20, 2040.00, 2039.90] [2034.60, 2034.70, 2034.80] [2039.90, 2039.80, 2039.60] [2034.20, 2034.40, 2034.50] [4.10, 3.70, 3.50] 0.2 [5.20, 5.00, 4.70] [3.00, 2.90, 2.80] [3.10, 2.90, 2.80] [4.50, 2.90, 2.50] [[22.06, 4.8, 1.0, 0.9], [21.99, 3.8, 1.5, 0.8], [21.97, 3.1, 1.6, 0.7]]
Wake-up growth rate fraction compute training AI models 0.6 0% [2037.90, 2037.20, 2037.00] [2040.30, 2040.10, 2040.00] 0.1 [2036.70, 2036.20, 2035.60] [2040.30, 2040.00, 2039.90] [2034.80, 2034.70, 2034.60] [2040.00, 2039.80, 2039.70] [2034.40, 2034.40, 2034.40] [3.50, 3.70, 4.10] -0.2 [5.10, 5.00, 5.10] [2.40, 2.90, 3.00] [2.90, 2.90, 3.40] [3.90, 2.90, 1.70] [[21.99, 4.1, 1.4, 0.6], [21.99, 3.8, 1.5, 0.8], [21.99, 3.6, 1.5, 0.5]]
Wakeup trigger 0.6 1% [2036.70, 2037.20, 2038.90] [2039.40, 2040.10, 2041.10] -0.3 [2035.40, 2036.20, 2037.60] [2039.40, 2040.00, 2041.00] [2033.40, 2034.70, 2035.00] [2039.20, 2039.80, 2040.80] [2032.20, 2034.40, 2037.60] [3.80, 3.70, 3.20] -0.4 [5.70, 5.00, 5.60] [2.70, 2.90, 2.20] [3.60, 2.90, 2.60] [4.00, 2.90, 4.70] [[21.99, 5.1, 1.7, 0.4], [21.99, 3.8, 1.5, 0.8], [21.99, 2.2, 1.0]]
Initial cognitive share goods 0.5 2% [2037.20, 2037.20, 2037.10] [2040.30, 2040.10, 2039.90] 0.0 [2036.10, 2036.20, 2036.20] [2040.20, 2040.00, 2039.80] [2034.60, 2034.70, 2034.70] [2040.00, 2039.80, 2039.60] [2034.30, 2034.40, 2034.40] [4.00, 3.70, 3.50] 0.1 [5.30, 5.00, 4.80] [3.10, 2.90, 2.80] [3.10, 2.90, 2.80] [4.90, 2.90, 2.50] [[20.36, 5.0], [21.99, 3.8, 1.5, 0.8], [23.39, 3.2, 1.4, 0.6]]
Wake-up growth rate fraction labour software R&D 0.4 0% [2037.20, 2037.20, 2037.10] [2040.20, 2040.10, 2039.90] -0.1 [2036.20, 2036.20, 2036.20] [2040.20, 2040.00, 2039.80] [2034.70, 2034.70, 2034.70] [2039.90, 2039.80, 2039.60] [2034.40, 2034.40, 2034.40] [3.90, 3.70, 3.50] 0.0 [5.10, 5.00, 4.80] [3.00, 2.90, 2.80] [3.00, 2.90, 2.80] [3.00, 2.90, 2.90] [[21.99, 3.8, 1.6, 0.5], [21.99, 3.8, 1.5, 0.8], [21.99, 3.6, 1.4, 0.5]]
Wake-up growth rate fraction of compute software R&D 0.4 0% [2037.20, 2037.20, 2037.10] [2040.30, 2040.10, 2040.00] 0.1 [2036.20, 2036.20, 2036.20] [2040.30, 2040.00, 2039.90] [2034.70, 2034.70, 2034.70] [2039.90, 2039.80, 2039.70] [2034.40, 2034.40, 2034.40] [4.00, 3.70, 3.60] 0.2 [5.10, 5.00, 4.90] [3.10, 2.90, 2.90] [2.90, 2.90, 2.90] [2.90, 2.90, 2.90] [[21.99, 3.8, 1.5, 0.6], [21.99, 3.8, 1.5, 0.8], [21.99, 3.8, 1.4, 0.4]]
Trade-off efficiency 0.4 1% [2037.70, 2037.20, 2036.90] [2041.10, 2040.10, 2039.60] 0.5 [2037.10, 2036.20, 2035.30] [2041.00, 2040.00, 2039.60] [2035.90, 2034.70, 2034.10] [2040.80, 2039.80, 2039.30] [2035.40, 2034.40, 2033.40] [3.80, 3.70, 4.20] -0.4 [4.80, 5.00, 5.10] [3.40, 2.90, 2.70] [2.80, 2.90, 3.20] [3.00, 2.90, 3.80] [[21.99, 3.6, 1.6, 0.4], [21.99, 3.8, 1.5, 0.8], [21.99, 4.2, 1.5, 0.5]]
Maximum software performance 0.4 0% [2037.30, 2037.20, 2037.10] [2040.40, 2040.10, 2039.90] 0.1 [2036.30, 2036.20, 2036.10] [2040.40, 2040.00, 2039.80] [2034.80, 2034.70, 2034.60] [2040.10, 2039.80, 2039.60] [2034.50, 2034.40, 2034.30] [4.00, 3.70, 3.60] 0.2 [5.20, 5.00, 4.90] [3.10, 2.90, 2.80] [3.00, 2.90, 2.80] [2.90, 2.90, 2.80] [[21.99, 3.9, 1.5, 0.5], [21.99, 3.8, 1.5, 0.8], [21.99, 3.7, 1.5, 0.6]]
Substitution between capital and cognitive tasks for goods and services 0.3 0% [2037.20, 2037.20, 2037.20] [2040.40, 2040.10, 2040.00] 0.2 [2036.20, 2036.20, 2036.20] [2040.30, 2040.00, 2040.00] [2034.70, 2034.70, 2034.70] [2040.00, 2039.80, 2039.70] [2034.40, 2034.40, 2034.40] [4.00, 3.70, 3.70] 0.3 [5.20, 5.00, 4.90] [3.20, 2.90, 2.80] [3.00, 2.90, 2.90] [149.90, 2.90, 2.90] [[21.99, 5.9, 12.1, 10.8], [21.99, 3.8, 1.5, 0.8], [21.99, 3.7, 1.3, 0.4]]
Wake-up growth rate fraction labour hardware R&D 0.2 0% [2037.20, 2037.20, 2037.20] [2040.20, 2040.10, 2039.90] -0.1 [2036.20, 2036.20, 2036.20] [2040.10, 2040.00, 2039.90] [2034.70, 2034.70, 2034.70] [2039.80, 2039.80, 2039.60] [2034.40, 2034.40, 2034.40] [3.80, 3.70, 3.60] 0.0 [5.00, 5.00, 4.80] [3.00, 2.90, 2.70] [2.90, 2.90, 2.90] [2.90, 2.90, 2.90] [[21.99, 3.8, 1.5, 0.5], [21.99, 3.8, 1.5, 0.8], [21.99, 3.7, 1.4, 0.5]]
Maximum hardware performance 0.2 1% [2037.30, 2037.20, 2037.10] [2040.40, 2040.10, 2039.90] 0.1 [2036.30, 2036.20, 2036.10] [2040.30, 2040.00, 2039.90] [2034.80, 2034.70, 2034.60] [2040.00, 2039.80, 2039.60] [2034.50, 2034.40, 2034.30] [3.90, 3.70, 3.70] 0.2 [5.10, 5.00, 4.90] [3.10, 2.90, 2.80] [3.00, 2.90, 2.90] [2.90, 2.90, 2.80] [[21.99, 3.8, 1.6, 0.5], [21.99, 3.8, 1.5, 0.8], [21.99, 3.7, 1.5, 0.5]]
Runtime goods vs R&D 0.2 1% [2037.30, 2037.20, 2037.10] [2040.30, 2040.10, 2039.90] 0.0 [2036.20, 2036.20, 2036.10] [2040.20, 2040.00, 2039.90] [2034.80, 2034.70, 2034.70] [2039.90, 2039.80, 2039.60] [2034.40, 2034.40, 2034.30] [3.90, 3.70, 3.70] 0.2 [5.00, 5.00, 4.80] [3.00, 2.90, 2.80] [2.90, 2.90, 2.90] [2.90, 2.90, 2.80] [[21.99, 3.9, 1.5, 0.5], [21.99, 3.8, 1.5, 0.8], [21.99, 3.8, 1.4, 0.4]]
Wake-up growth rate fraction capital hardware R&D 0.2 1% [2037.20, 2037.20, 2037.20] [2040.10, 2040.10, 2040.00] -0.1 [2036.20, 2036.20, 2036.20] [2040.10, 2040.00, 2039.90] [2034.70, 2034.70, 2034.70] [2039.80, 2039.80, 2039.70] [2034.40, 2034.40, 2034.40] [3.80, 3.70, 3.60] 0.0 [5.00, 5.00, 4.90] [2.90, 2.90, 2.80] [3.00, 2.90, 2.80] [2.90, 2.90, 2.90] [[21.99, 3.8, 1.5, 0.6], [21.99, 3.8, 1.5, 0.8], [21.99, 3.7, 1.5, 0.8]]
Wake-up growth rate fraction compute hardware R&D 0.1 0% [2037.20, 2037.20, 2037.20] [2040.10, 2040.10, 2040.10] 0.0 [2036.20, 2036.20, 2036.20] [2040.10, 2040.00, 2040.00] [2034.70, 2034.70, 2034.70] [2039.90, 2039.80, 2039.70] [2034.40, 2034.40, 2034.40] [3.80, 3.70, 3.70] 0.1 [5.10, 5.00, 4.90] [2.90, 2.90, 2.90] [3.00, 2.90, 2.90] [2.90, 2.90, 3.00] [[21.99, 3.8, 1.5, 0.7], [21.99, 3.8, 1.5, 0.8], [21.99, 3.8, 1.4, 0.4]]
Max fraction compute software R&D 0.1 0% [2037.20, 2037.20, 2037.20] [2040.10, 2040.10, 2040.10] 0.0 [2036.20, 2036.20, 2036.20] [2040.10, 2040.00, 2040.00] [2034.70, 2034.70, 2034.70] [2039.80, 2039.80, 2039.80] [2034.40, 2034.40, 2034.40] [3.80, 3.70, 3.70] 0.1 [5.00, 5.00, 5.00] [2.90, 2.90, 2.90] [2.90, 2.90, 2.90] [2.90, 2.90, 2.90] [[21.99, 3.8, 1.5, 0.8], [21.99, 3.8, 1.5, 0.8], [21.99, 3.8, 1.5, 0.8]]
Substitution between capital and cognitive tasks for hardware R&D 0.1 0% [2037.20, 2037.20, 2037.20] [2040.10, 2040.10, 2040.10] 0.0 [2036.20, 2036.20, 2036.20] [2040.10, 2040.00, 2040.00] [2034.70, 2034.70, 2034.70] [2039.80, 2039.80, 2039.80] [2034.40, 2034.40, 2034.40] [3.80, 3.70, 3.70] 0.1 [5.00, 5.00, 5.00] [2.90, 2.90, 2.90] [3.10, 2.90, 2.90] [2.90, 2.90, 2.90] [[21.99, 3.8, 1.5, 0.6], [21.99, 3.8, 1.5, 0.8], [21.99, 3.8, 1.4, 0.5]]
Max fraction of labour dedicated to hardware R&D 0.0 1% [2037.20, 2037.20, 2037.20] [2040.10, 2040.10, 2040.10] 0.0 [2036.20, 2036.20, 2036.20] [2040.00, 2040.00, 2040.00] [2034.70, 2034.70, 2034.70] [2039.80, 2039.80, 2039.80] [2034.40, 2034.40, 2034.40] [3.70, 3.70, 3.70] 0.0 [5.00, 5.00, 5.00] [2.90, 2.90, 2.90] [2.90, 2.90, 2.90] [2.90, 2.90, 2.90] [[21.99, 3.8, 1.5, 0.8], [21.99, 3.8, 1.5, 0.8], [21.99, 3.8, 1.5, 0.8]]
Max fraction of compute dedicated to hardware R&D 0.0 0% [2037.20, 2037.20, 2037.20] [2040.10, 2040.10, 2040.10] 0.0 [2036.20, 2036.20, 2036.20] [2040.00, 2040.00, 2040.00] [2034.70, 2034.70, 2034.70] [2039.80, 2039.80, 2039.80] [2034.40, 2034.40, 2034.40] [3.70, 3.70, 3.70] 0.0 [5.00, 5.00, 5.00] [2.90, 2.90, 2.90] [2.90, 2.90, 2.90] [2.90, 2.90, 2.90] [[21.99, 3.8, 1.4, 0.5], [21.99, 3.8, 1.5, 0.8], [21.99, 3.8, 1.5, 0.8]]
Max fraction labour software R&D 0.0 1% [2037.20, 2037.20, 2037.20] [2040.10, 2040.10, 2040.10] 0.0 [2036.20, 2036.20, 2036.20] [2040.00, 2040.00, 2040.00] [2034.70, 2034.70, 2034.70] [2039.80, 2039.80, 2039.80] [2034.40, 2034.40, 2034.40] [3.70, 3.70, 3.70] 0.0 [5.00, 5.00, 5.00] [2.90, 2.90, 2.90] [2.90, 2.90, 2.90] [2.90, 2.90, 2.90] [[21.99, 3.8, 1.5, 0.8], [21.99, 3.8, 1.5, 0.8], [21.99, 3.8, 1.5, 0.8]]
Max fraction GWP compute 0.0 0% [2037.20, 2037.20, 2037.20] [2040.10, 2040.10, 2040.10] 0.0 [2036.20, 2036.20, 2036.20] [2040.00, 2040.00, 2040.00] [2034.70, 2034.70, 2034.70] [2039.80, 2039.80, 2039.80] [2034.40, 2034.40, 2034.40] [3.70, 3.70, 3.70] 0.0 [5.00, 5.00, 5.00] [2.90, 2.90, 2.90] [2.90, 2.90, 2.90] [2.90, 2.90, 2.90] [[21.99, 3.8, 1.5, 0.8], [21.99, 3.8, 1.5, 0.8], [21.99, 3.8, 1.5, 0.8]]
Compute depreciation rate 0.0 0% [2036.90, 2037.20, 2037.40] [2039.90, 2040.10, 2040.30] 0.0 [2036.00, 2036.20, 2036.40] [2039.80, 2040.00, 2040.20] [2034.60, 2034.70, 2034.90] [2039.60, 2039.80, 2040.00] [2034.20, 2034.40, 2034.60] [3.70, 3.70, 3.70] 0.0 [4.90, 5.00, 5.00] [3.00, 2.90, 2.90] [2.90, 2.90, 3.00] [2.90, 2.90, 2.90] [[21.62, 3.7, 1.5, 0.5], [21.99, 3.8, 1.5, 0.8], [22.43, 3.8, 1.5, 0.8]]
Max fraction capital hardware R&D 0.0 0% [2037.20, 2037.20, 2037.20] [2040.10, 2040.10, 2040.10] 0.0 [2036.20, 2036.20, 2036.20] [2040.00, 2040.00, 2040.00] [2034.70, 2034.70, 2034.70] [2039.80, 2039.80, 2039.80] [2034.40, 2034.40, 2034.40] [3.70, 3.70, 3.70] 0.0 [5.00, 5.00, 5.00] [2.90, 2.90, 2.90] [2.90, 2.90, 2.90] [2.90, 2.90, 2.90] [[21.99, 3.8, 1.5, 0.8], [21.99, 3.8, 1.5, 0.8], [21.99, 3.8, 1.5, 0.8]]
Totals Sum of agi_year skews: 12.40 Sum of full_automation_gns skews: 5.90

Inputs

Show

Conservative Best guess Aggressive
AGI training requirements (FLOP with 2022 algorithms) 1.000000e+40 1.000000e+36 1.000000e+33
Effective FLOP gap (training) 1.000000e+08 1.000000e+04 1.000000e+01
Training goods vs R&D 1.000000e+00 3.000000e+00 3.000000e+01
AGI runtime requirements 1.000000e+20 1.666670e+16 1.000000e+15
Effective FLOP gap (runtime) 1.000000e+02 1.000000e+01 1.000000e+00
Runtime goods vs R&D 3.000000e+01 1.000000e+02 3.000000e+02
Trade-off efficiency 3.000000e+00 1.500000e+00 8.000000e-01
Maximum trade-off 1.000000e+00 3.000000e+01 1.000000e+03
Labour substitution goods -2.000000e+00 -5.000000e-01 -2.000000e-01
Labour substitution R&D -2.000000e+00 -5.000000e-01 -2.000000e-01
Substitution between capital and cognitive tasks for goods and services -2.000000e+00 -4.000000e-01 -2.000000e-01
Substitution between capital and cognitive tasks for hardware R&D -5.000000e-01 -2.500000e-01 -1.700000e-01
Research experiments substitution software -1.000000e-02
Efficiency of experiments for software R&D 4.000000e-01
Returns to hardware 3.000000e+00 5.200000e+00 7.000000e+00
Returns to software 8.000000e-01 1.250000e+00 5.000000e+00
Maximum hardware performance 1.000000e+23 1.000000e+26 1.000000e+29
Maximum software performance 1.000000e+08 1.000000e+12 1.000000e+16
R&D parallelization penalty 2.000000e-01 7.000000e-01 9.900000e-01
Hardware adoption delay 2.500000e+00 1.000000e+00 1.000000e-02
Growth rate fraction capital hardware R&D 1.000000e-02
Growth rate fraction labour hardware R&D 1.000000e-02
Growth rate fraction compute hardware R&D 1.000000e-02
Growth rate fraction labour software R&D 1.800000e-01
Growth rate fraction compute software R&D 1.800000e-01
Growth rate fraction GWP compute 7.000000e-02 1.900000e-01 3.700000e-01
Growth rate fraction compute training 9.312925e-01 5.475284e-01 1.637642e-01
Wake-up growth rate fraction capital hardware R&D 5.000000e-02 1.400000e-01 3.200000e-01
Wake-up growth rate fraction labour hardware R&D 5.000000e-02 1.400000e-01 3.200000e-01
Wake-up growth rate fraction compute hardware R&D 2.700000e-01 6.700000e-01 1.370000e+00
Wake-up growth rate fraction labour software R&D 1.200000e-01 2.200000e-01 3.700000e-01
Wake-up growth rate fraction of compute software R&D 2.700000e-01 6.700000e-01 1.370000e+00
Wake-up growth rate fraction of GWP buying compute 7.000000e-02 1.900000e-01 3.700000e-01
Wake-up growth rate fraction compute training AI models 5.500000e-01 1.100000e+00 4.500000e+00
Max fraction capital hardware R&D 1.000000e-02 3.000000e-02 1.000000e-01
Max fraction of labour dedicated to hardware R&D 1.000000e-02 3.000000e-02 1.000000e-01
Max fraction of compute dedicated to hardware R&D 5.000000e-02 2.000000e-01 3.500000e-01
Max fraction labour software R&D 1.000000e-02 3.000000e-02 1.000000e-01
Max fraction compute software R&D 5.000000e-02 2.000000e-01 3.500000e-01
Max fraction GWP compute 3.000000e-02 1.000000e-01 3.000000e-01
Max fraction compute training 3.000000e-02 1.000000e-01 2.000000e-01
Initial fraction capital hardware R&D 2.000000e-03
Initial fraction labour hardware R&D 2.000000e-03
Initial fraction compute hardware R&D 2.000000e-03
Initial fraction labour software R&D 2.000000e-04
Initial fraction compute software R&D 2.000000e-04
Initial biggest training run 3.000000e+24
Initial vs cumulative input - hardware R&D 4.700000e-02
Initial vs cumulative input - software R&D 2.000000e-01
Initial hardware production 1.000000e+27 1.000000e+28 1.000000e+29
Accumulated hardware vs initial hardware production 2.000000e+00
Initial market hardware performance 1.500000e+17
Initial GWP 8.500000e+13
Initial world labour force. 4.000000e+09
Initial cognitive share goods 3.000000e-01 5.000000e-01 6.500000e-01
Initial cognitive share in hardware R&D 7.000000e-01
Initial compute share goods 1.000000e-02
Initial compute share R&D 1.000000e-02
Initial experiment share software R&D 3.000000e-01
Wakeup trigger 1.000000e-02 6.000000e-02 2.000000e-01
Initial capital growth rate 2.750000e-02
Population growth rate 1.000000e-02
TFP growth rate 1.000000e-02
Compute depreciation rate 1.000000e-01 2.000000e-01 3.000000e-01
Money threshold training 2.000000e+10 4.000000e+09 4.000000e+08
Training requirements steepness (OOM) 0.000000e+00
Runtime requirements steepness (OOM) 0.000000e+00

Here, you will find a comparative analysis of the model results conditional on beliefs that correspond to short, medium or long timelines until AGI.

Metrics

Show

FLOP to train AGI using 2022 algorithms Scenario Training FLOP gap Powerful sub-AGI year AGI year 20% automation year 100% automation year 20% R&D automation year 100% R&D automation year Wake-up year 20-100% economic automation 20-100% R&D automation Powerful sub-AGI to AGI 2X to 10X cognitive output multiplier 5% to 20% GWP growth Pattern of GWP doublings
1e30 Conservative 1.0e+04 2024.4 2039.0 2032.2 2049.5 2023.6 2040.8 2026.5 17.0 17.0 14.6 6.6 13.2 [21.4, 10.3, 4.8, 3.0, 2.6]
1e30 Best guess 1.0e+02 2028.3 2030.2 2029.1 2032.1 2027.7 2031.0 2027.4 2.9 3.3 1.9 1.4 3.0 [21.99, 3.4, 0.7, 1.4]
1e30 Aggressive 1.0e+01 2028.4 2028.5 2028.4 2028.6 2028.1 2028.5 2028.2 0.2 0.4 0.1 0.2 0.2 [19.75, 0.3]
1e36 Conservative 1.0e+07 2033.9 2065.5 2038.2 2065.5 2031.6 2064.5 2032.0 27.0 32.3 31.6 17.7 24.9 [22.32, 11.2, 7.6, 5.4, 4.1]
1e36 Best guess 1.0e+04 2039.1 2044.2 2039.1 2044.2 2038.5 2044.2 2037.3 5.0 5.6 5.1 3.7 3.7 [21.99, 3.9, 2.2, 0.8]
1e36 Aggressive 1.0e+02 2036.6 2037.0 2036.6 2037.0 2036.4 2037.0 2036.3 0.4 0.6 0.4 0.4 0.5 [19.75, 0.5, 0.2]
1e42 Conservative 1.0e+10 2048.0 2108.1 2058.2 2108.1 2046.1 2107.0 2050.4 49.5 60.3 60.1 30.2 42.9 [22.32, 12.6, 8.4, 7.5, 7.2]
1e42 Best guess 1.0e+06 2055.7 2072.1 2055.7 2072.1 2053.2 2072.0 2050.7 15.7 18.1 16.4 11.1 20.4 [21.99, 9.2, 6.1, 3.4, 2.0]
1e42 Aggressive 1.0e+03 2052.6 2053.5 2052.6 2053.5 2052.4 2053.4 2052.0 0.9 1.0 0.9 0.6 1.8 [19.75, 1.0, 0.4]

Compute increase over time

Model summary

Scenario

Compare to

FLOP to train AGI using 2022 algorithms Scenario Training FLOP gap Period Year Biggest training run Fraction of goods and services tasks automated Fraction of R&D tasks automated Hardware performance growth rate Hardware performance doubling time Software growth rate Software doubling time Investment on compute growth rate Investment on compute doubling time Fraction of compute invested in training growth rate Fraction of compute invested in training doubling time GWP growth rate GWP doubling time Capital growth rate Capital doubling time Labour growth rate Labour doubling time R&D TFP growth rate R&D TFP doubling time Software R&D input growth rate Software R&D input doubling time Cumulative software R&D input growth rate Cumulative software R&D input doubling time

Assumptions

Scenario

Compare to

Show