- Model Specs
- Domains
- Hardware
- Software
- Scaling
- Money
- Forecasting
- Predictions
- Academia Data
- Effects of AI
- Countries
Summary of AI Model Specs Progress By Domain
History of AI Model Specs Progress By Domain
[1] Light red is the predicted Singularity/TAI based on training FLOPs bounded by 50% conservative and 50% aggressive Biological Anchor forecasts
[2] Light blue on the is the max optimal parameters and datapoints based on the on the largest model in terms of training FLOPs (PaLM/Minerva), found using the Chinchilla optimal scaling laws
[3] Estimated Singularity/TAI model's FLOPs, Optimal Parameters, and Optimal Datapoints. Predicted to all lie inside light red area based on the Biological Anchor forecasts and Chinchilla optimal scaling laws
[4] Light gray is the forecasted top model training in FLOPs from aggressive, best guess, and conservative Biological Anchor forecasts
[5] Switch from aggressive, best guess, or conservative using the button on the top right of the graph
[6] To see more of the forecasts select "pan" on the top right of the graph
Sources
Forcasting Transformative AI Full
Forcasting Transformative AI Summary
FLOPs, Datapoints of Animals Estimated (Good Collection of Paper links, links at the bottom)
Summary of AI Model Domains Progress
Language Domain Progress by Benchmark
Big-Bench
STEM/Math/CS Domain Progress by Benchmark
Games Domain Progress by Benchmark
Vision
Image Generation
Video
Music
Driving
Medical: AlphaFold
History of AI Model Domains Progress
Language
Games
Vision
Image Generation
Video
Music
Driving
Coding
Sources
Summary of Computer Hardware Progress
CPU transistors
GPU transistors
Storage
RAM
History of Computer Hardware Progress
Summary of AI Systems Progress By Architecture
Top models are using different forms of Transformers: *show data*
Summary of AI Systems Progress By Algorithm Efficiency
History of AI Systems Progress By Architecture
History of AI Systems Progress By Algorithm Efficiency
Summary of AI Systems Scaling Progress By Domain
History of AI Systems Scaling Progress By Domain
AI Scaling Laws for Compute, Dataset size, Parameters
Emerging Skills as Parameter Size Scales
Summary of AI Systems Progress By Money
Doubling time of spending on compute | Annual growth rate (%) of real frontier GDP
Estimated ~$35 Million to train Alpha Zero in 10/19/2017
Estimated ~$9-$23 Million to train PaLM in 4/4/2022
Summary of Estimated Cost of Training Models
History of AI Systems Progress By Money
Compute cost for the most expensive training run | Doubling time of spending on compute | Annual growth rate (%) of real frontier GDP
History of Estimated Cost of Training Models
High Level AI ERAs Timeline
[1] Narrow AI -> [2] Weak AGI/Easy Turing -> [3] Strong AGI/Hard Turing/HLMI -> [4] Singularity/TAI -> [5] ASI
[2] Weak AGI and [3] Strong AGI/HLMI AI Forecasting
HLMI as machines that are capable, either individually or collectively, of performing almost all economically-relevant information-processing tasks that are performed by humans, or quickly (relative to humans) learning to perform such tasks
[4] Singularity/Transformative AI Forecasting
Transformative artificial intelligence (TAI) is artificial intelligence that constitutes a transformative development, i.e. a development at least as significant as the agricultural or industrial revolutions.
Aggressive Anchor Probabilities = Life: 10%, Short: 40%, Genome: 20%, Medium: 10%, Long: 10%, Evolve: 5%
Best Guess Anchor Probabilities = Life: 5%, Short: 20%, Genome: 10%, Medium: 30%, Long: 15%, Evolve: 10%
Conservative Anchor Probabilities = Life: 2%, Short: 8%, Genome: 5%, Medium: 15%, Long: 30%, Evolve: 20%
Countdown to the Singularity: Checklist of things that need/probably will happen before the singularity
[5] Artificial Superintelligence AI Forecasting
A superintelligence is a cognitive system whose intellectual performance across all relevant domains vastly exceeds that of any human.
AI Aftermath Scenarios
Sources
Future Predictions
https://rodneybrooks.com/predictions-scorecard-2022-january-01/
https://prod.hypermind.com/ngdp/en/showcase2/showcase.html?sc=JSAI
My Future Predictions
Pre-Singularity
1. I predict the intelligence and capabilities of AI agents will increase dramatically over the next few decades.
2. I predict more and more people will be replaced as the trained AI agents become more and more capable, eventually causing a large percentage of the population to be poor and unemployed or underemployed.
3. I predict wealth will be concentrated more and more into smaller and smaller number of hands and businesses that own these expensive trained AI agents will make tons of money. Giving them more and more power.
4. I predict the general population will fully own decreasing amount of things and will have to pay the rich to rent/use their things
5. I predict chaos, political tension, and tribalism will increase as more and more people have less and less wealth and are manipulated with more and more effective misinformation and propaganda. May eventually lead to a war or uprising of some sort. May lead to more countries to authoritarianism.
6. I predict the effectiveness of misinformation and propaganda will be orders of magnitude better than todays using trained language models and sentiment analysis. Directing the populations anger and frustrations to certain things for example. Discerning fact from fiction will be increasingly difficult.
7. I predict the companies that own and have access to computing power and data to train these large AI agents will dramatically increase in value and put many companies out of business that don't have AI agents.
8. I predict the scientists who are able to create, train, and implement these AI agents from the software side and hardware side will be some of the most valuable/most well paid people in society and will be strongly recruited by the military and government.
9. I predict the military/police will most likely get access the large AI models trained for offense and defense, especially as chaos increases and billionaires become targets. These agents will most likely be extremely effective. Not sure what form this will take.
10. I predict we will get weak AGI around 2025-2030. This is assuming no huge discontinuous jumps or large discoveries that completely shift the paradigm.
11. I predict we will get strong AGI or HLMI around 2035-2040. This is assuming no huge discontinuous jumps or large discoveries that completely shift the paradigm.
12. I predict based on the optimal scaling laws from the Chinchilla paper, companies will need an increasing amount of data to optimally train models. I predict that because of this companies will either directly or indirectly create products that extract as much data from people as possible.
13. Following #12, I predict that companies will heavily lobby for more lax data privacy laws
14. GDP will rise ____
15. Countries that are more democratic will fare better in general than those that are less due to proactive policies and more wealth distribution
Singularity
1. I predict due to the forecasted cost from Biological Anchors paper to train TAI/Singularity models (over $100 billion), it will be unlikley that a private or public company by itself will be the first to create a train TAI/Singularity models. It will most likely be a large government project spending around 1/2% to 2% of GDP working with top tech companies. This is assuming no huge discontinuous jumps or large discoveries that completely shift the paradigm.
2. I predict the first TAI/Singularity model will be from a militarized AI arms race between the top countries.
3. I predict the first TAI/Singularity model will be trained around 2035-2050. This is assuming no huge discontinuous jumps or large discoveries that completely shift the paradigm.
Post-Singularity
1. If artificial superintelligence is reached it will seem almost godlike to humans
2. I predict, if humans survive the singularity, the human race will no longer be the main ones in control. This is assuming ASI levels are reached
3.
Summary of Number of AI Papers by Field
*Show data/graphs*
Summary of AI Software Stared on Github
*Show data/graphs*
Summary of Number of People at Conferences by Conference
*Show data/graphs*
Summary of Number of Mentions in Legal Docs by Country
*Show data/graphs*
History of Number of AI Papers by Field
*Show data/graphs*
History of AI Software Stared on Github
*Show data/graphs*
History of Number of People at Conferences by Conference
*Show data/graphs*
History of Number of Mentions in Legal Docs by Country
*Show data/graphs*
Summary of The Effects of Automation and AI Models
History of The Effects of Automation and AI Models
Total Goverment Spending of Countries Compared
US Goverment Spending Overtime, Pie Chart or Flow Diagram, Per Person
Underemployment
Owning Things
Union Power
AI Fire Alarm
Number of Low wage Jobs Compared to High Wage Jobs
Prevalence of Propaganda
Prevalence of Radicalism/Tribalism/Extremism
Cost of Living
Amount of Dept of an Average Citizen
Suicide Rates, Alcohol/Drug Abuse, Depression Rates
Medium Income vs GDP
Wealth Distribution by Generation
Koomey's Law and Landaurs priciple to curve fit technological progress. It's nesessarily a logistical curve bc currently exponential and known upper bound ???
Capital income vs labor income
Don't inclue median graph needs to be capital
Graph of distibution of capital income over time
Capital income vs labor income
https://www.givingwhatwecan.org/blog/measuring-global-inequality-median-income-gdp-per-capita-and-the-gini-index
Democracy - worldpopulationreview.com
Median Income - https://fred.stlouisfed.org/series/MEPAINUSA672N
Gini Coefficient of Countries
Gini Coefficient vs GDP Per Capita Countries
Productivity vs Wages
Political Instability
Number of Well Paid Employees in Fortune 500 Companies
Summary of AI Model Specs Progress By Country
Summary of Detailed AI Score By Country
Summary of Hardware Progress By Country
History of AI Model Specs Progress By Country
History of Hardware Progress By Country
Countries Trends Since the Computer Age: AI systems training FLOPs vs Bad --|-- Good Environment
The singularity/TAI will probably amplify the environment around it. It will turn a bad environment to dystopian and a good environment to utopian
