You are on page 1of 1

Levels of defense in AI safety

Levels Global solutions Local solutions


Funding of Attracting Creation of AI safety
0 Preparation the AI safety guidlines and
the AI safety minds
reserch theory mindset

Prelimnary
measures Prevention
Destruction Promotion of Slowing AI
of the of AI labs
Restriction
AI safety idea
by law development
Strong AI everywhere and theory
creation

Limiting AI Creating Preventing self- Messy AIs and


scale and Manipulating
many AIs for improvement uploads
impact AI arms race
balansing? of an AI

Solving AI EthicsNet dataset Correct decesion Provably safe Human Human


1 alignment competition theory AI design emulation self-improving
problem
Safe
AI Installing CEV AI Vague values
correct constitution
design goals

Solving Taking over Creating AI First AI creators will


other AI the world police implement safe AI
problem theory

Safely interruptable Red button and


2 Control agents; Corrigibility other internal
Circuit-breakers
systems constrains
Catching
treacherous
turn Testing
Testing in Testing in Narrow AI
simulation small tasks to control AI

3 Information Disconection from Phisically remote


limiting interent place
AI boxing

Instruments
4 to affect Messaging to Utility monsters
mature AI future Young AI and landmines
Stopping Created by Alexey Turchin, 2017, CC4.0
See also other my maps on the topic:
AI failures modes
AI AI safety solution
and articles
Hopes to
in the wild survive AI Humans Humans as AI will
Message to any future AI
Catching treacherous turn MIRI strategy (as
in simulation slaves abandon Earth Global and local solutions of AI safety I understand it)
Human upload based AI Nanny marked in yellow

Time