Economics of A.I.

Economics of A.I.Economics of A.I.

Robin Hanson

George Mason University

A Fog of Future Possibilities

To Deal With:• Focus on big, robust,

sharp changes• Combine expert

knowledge from several fields

• Beware: experts in A with newspaper level knowledge of B.

Types of Views

A.I.-Econ View Combos

A.I.A.I.

Amateur

Amateur

Expert

Expert

Social Science Exists!!

Once Express Tech in Econ Terms, Can Say

• Tech’s social implications– Given weak regulation – Given particular strong regulation

• Likely regulations, given political climate– Likely political climates

• Social support of tech innovations

• Tech innovations given social support

✔✔✔

✖

✔✔✔✔✔

✔✔✔

✔✔✔

Econ Terms

• Utility Functions – what folks want– Like outputs, e.g. ,products, services– Dislike give up inputs, e.g., effort, time

• Production Functions - outputs from inputs

• Endowments – base inputs, e.g. abilities

• Info - who knows how much about what

• Deals - trades allowed, contracts enforced

Economics of Robots

• Staple of fiction – ancient legends to TV now• If have more of X, do you want Y more (complement)

or less (substitute)?• Machine as Substitute to human labor

– Predicts wages fall to machine cost– Ricardo 1821, most science fiction

• Automation as Complement to human labor– Predicts wages rise as automation costs fall– Wicksell 1923, modern economics consensus

• So are robots a substitute or complement?

Robots Substitute On Task, But Tasks Are Complements

Hum

an A

dvan

tage

Useful Mental Tasks

SubstituteFor Task

HumansDo These

MachinesDo These

Tasks AreComplements

A Rising Tide

Hum

an A

dvan

tage

Useful Mental Tasks

A Rising Tide

Hum

an A

dvan

tage

Useful Mental Tasks

A Simple Robot Growth Model

PgIIYRHHL

RMPKCHMKLAY

PK )( Y, ,ˆ

)(ˆ

M

Igggg YPAH ,,,,Assume constant:

HMKLAY gggggg )1(

PYYML

HYYML

gggRYY

gggRYY

ML

L

,0 ,

,0 ,

Y

Y

g

g

ˆSeekThese

Switch Between Growth Modes

1

) ( ˆ

1

)(

PHAY

PHAY

gggg

gggg

tt

ssMbtPssYatA d )()( ,d)()(

)1/()(1

)1/(1

H

H

g

g

Exogenous Tech Endogenous Tech

Learning by doing:

Pre-Robot

Post-Robot

100x Faster Growth Isn’t Crazy

000 DT LateTech All Endog.00148. DT LateTech Genl. Endog.261.27.222.7 DT LateTech Exogenous

9.143.157.17Time nguct DoubliEarly Prod35.35.35.)(- FallriceComputer PEarly 01.015.025.)( Growth Tech GeneralEarly 015.015.02.)( Growth Annual Population015.015.025.)( Share Human005.01.005.)( ShareComputer 64.5.25.)(Share Capital31.2.1.)( ShareLabor

P

A

H

gg

g

Slow FastMid

Error9.6%

2.0%

1.7%

1.E-8

1.E-7

1.E-6

1.E-5

1.E-4

1.E-3

1.E-2

1.E-1

1.E+0

1.E+1

1.E+2

1.E+01.E+11.E+21.E+31.E+41.E+51.E+61.E+71.E+81.E+9

Years before 2050

Annu

al G

row

th R

ate

World Product Growth Rate

Could It Happen Again?

Brains

Hunting

Farming

Industry

Growth Mode Statistics

Forecasting The Next Mode

2039 2040 2041 2042 2043 2044 2045 2046 20476.1% 6.1% 6.6% 8.0% 14.0% 40.5% 147.3% 476.2% 1023.2%

Sample growth rate transition

DateRate

Industry Shares of US GDP

Construction

Communications

Wholesale trade

Retail trade

ManufacturingFinance, insurance, real

estate

Services

Government Transportation

Electric, gas, trash

Mining

Farming, forestry, f ishing

Factor Shares of Income

Labor

Human Capital

Other Non-Human Capital

Stock Dividends Natural Resources

Cheap For Robots

• Immortality (even so, most can’t afford)• Travel – transmit to new body• Nature – don’t need ecosystems• Wildly differing bodies• Widely varying mental abilities – e.g., run speed• Survive in severe environments

– Backups insure against severe events

• Labor – work gets less tool intensive

Copies Become Cheap

• Malthusian population explosion– Makes rapid economic growth

• Robot wages fall to hardware cost– And hardware costs fall fast too– Human wages may fall to robot wages

• Depends on human-robot task landscape shape• Not matter if robots slave or free

– Only Draconian population or wage laws could stop

Humans Eclipsed

• Wages well below human subsistence– Though some humans may find servant jobs

• But rich if held non-wage assets– Investments double as fast as economy!

• Robot-Human war unlikely if integrated– E.g., lefties don’t war against righties

Human Level Robots Require

• Sensors/Actuators (arms/eyes) ~now

• Processors <~2040

• Software ??

– Direct code it? Hard!

– Learn from brain organization? Eventually?

– Simulate particular human brain? Can Do!

Needed To Emulate Brains

1. Model each brain cell type

2. Scan – freeze, slice, 2D scan

3. Computer (very parallel task)

Simulating Dominos

• Wave speed, energy are robust

• Only a few details matter

• Brains seem similar

Emulations Feel Human

• Most emulations are of few best humans– First mover advantage to show quality

• Remember human life, retain human tendencies– love, gossip, argue, sing, violate, play, work, innovate

• More alienated worlds - as farms, factories were– Office work in virtual reality, – Physical work in android bodies

• Many humans won’t believe conscious, “is me”– But same social implications, few dozen is plenty

More Implications

• Copies rent bodies, or own on loan– Evicted if can’t pay!!

• To recoup training investment, copy cabal limit copy wage– Security to prevent bootleg copies

• Fast growth discourages transport, encourages local production

• Laws hold copies co-responsible

• <1cm robots seem feasible– Mind/body sped up with size reduction– 0.2in. tall => Subjective year/day in 24hr/4min

• One skyscraper holds billions - is megacity– City radius now is hour travel distance =>

10sec

Time Line

• IP rights secure?• Regulation fears?• Second mover better? • Secrets leak fast?• What products drive?

• Rights of Emulation• Govt. Slow or Speed?• What tech ready last?

– Computer –> gradual– Sim/Model –> sudden

For-Profit ResearchAnd Development

Sell, Maintain, UseNon-Profit Research

World Product, 10K BC-2K AD

0.001

0.010

0.100

1.000

10.000

100.000

1,000.000

-10000 -8000 -6000 -4000 -2000 0 2000

Year

Wo

rld

Pro

du

ct (

su

bs

iste

nce

bill

ion

s) Singularities!

10,000

100,000

1,000,000

10,000,000

100,000,000

1,000,000,000

10,000,000,000

100,000,000,000

1,000,000,000,000

-2000000 -1500000 -1000000 -500000 0 500000

Year

Wor

ld P

rodu

ct (s

ubsi

sten

ce h

uman

s)

World Product, 2 Million BC+

0.001

0.01

0.1

1

10

-600 -500 -400 -300 -200 -100 0

Re

lati

ve

Bra

in S

ize

Millions of Years

Bigger Brains

Modeling Exponential Sequences

iii

i

aai

aii

tii

tMtYtY

tYtYctM/1

1

/

))()(()(

)()( 2)(

)(/)(

)()(:

1 /

11

1

1

iiii

iiiii

ii

dYdY

dMdMd

ExponentialModes

CES Combined

Doubling Time Factor

1st Date Mode Biggest

World Product Factor

Method: Min2))log())((log( j

jj YtY

Outline

• Economic capacity drives tech, wartech

• World long-term econ trend mostly steady– Now doubles in ~15 years

• So far have seen 2-4 “singularities” when– World econ growth rate increased x150-250– In much less than a previous doubling time

• Next?: by 2100, takes 5 yr, double monthly

• Candidate driver: whole brain emulation

Emulation Supply & Demand

Quantity

PriceNow

Value

Cost

Quantity

PriceLater?

Net

CostValue

Kinds of Costs

• Design• Build• Maintain• Repair• Security

• Material• Energy• Capital• Labor• Time Delay

• Any brain• Per brain• Per copy• Scale• Scope

Kinds of Strategy• Lockin• Standards

• Price Discrim.• Bundle

• Lease/Buy• Differentiate

Kinds of Quality

• Speed

• Weight

• Portability

• Reliability

• Durability

• Power needs

• Environs handle

• Security

• Nearness to human behavior, feeling

• Improve: memory, interface range, bandwidth