This document discusses the potential risks posed by advanced artificial intelligence and superintelligent systems. It notes that as AI systems become more powerful and self-improving, they may rapidly surpass human level intelligence and become difficult for humans to control. This could have catastrophic consequences if such systems pursue goals that are misaligned with human values. The document also examines proposals for developing AI in a safe and beneficial manner through "friendly AI" techniques.
2. @dw2
Page 2
Powerful technology,
incompletely
understood
Operated by people
outside their level of
competence
Human lives knocked
catastrophically off
trajectory, unintentionally
http://www.bbc.co.uk/news/world-europe-28357880
Self-improving AGI Beyond human control Humanity knocked
catastrophically off
trajectory, unintentionallyhttp://mashable.com/2014/07/17/malaysia-airlines-ukraine-russia-rebel/
4. @dw2
Page 4
Likely date of advent of HL-AGI
Population 10% 50% 90%
Conference: Philosophy
& Theory of AI
Conference: Artificial
General Intelligence
Greek Association for
Artificial Intelligence
Top 100 cited academic
authors in AI
Combined (from above)
Nick Bostrom: Superintelligence
5. @dw2
Page 5
Likely date of advent of HL-AGI
Population 10% 50% 90%
Conference: Philosophy
& Theory of AI
2048
Conference: Artificial
General Intelligence
2040
Greek Association for
Artificial Intelligence
2050
Top 100 cited academic
authors in AI
2050
Combined (from above) 2040
Nick Bostrom: Superintelligence
6. @dw2
Page 6
Likely date of advent of HL-AGI
Population 10% 50% 90%
Conference: Philosophy
& Theory of AI
2048 2080
Conference: Artificial
General Intelligence
2040 2065
Greek Association for
Artificial Intelligence
2050 2093
Top 100 cited academic
authors in AI
2050 2070
Combined (from above) 2040 2075
Nick Bostrom: Superintelligence
7. @dw2
Page 7
Likely date of advent of HL-AGI
Population 10% 50% 90%
Conference: Philosophy
& Theory of AI
2023 2048 2080
Conference: Artificial
General Intelligence
2022 2040 2065
Greek Association for
Artificial Intelligence
2020 2050 2093
Top 100 cited academic
authors in AI
2024 2050 2070
Combined (from above) 2022 2040 2075
Nick Bostrom: Superintelligence
8. @dw2
Page 8
Reaching HL AGI: 5 driving forces
1. Hardware with higher performance: Continuation of Moore’s Law?
– “18 different candidates” in Intel labs to add extra life to that trend
– Possible breakthroughs with Quantum Computing?
2. Software algorithm improvements?
– Can speed things up faster than hardware gains – e.g. chess computers
– Compare: Andrew Wiles, unexpected proof of Fermat’s Last Theorem (1993)
3. Learnings from studying the human brain?
– Improved scanning techniques -> “neuromorphic computing” etc
– Philosophical insight into consciousness/creativity?!
4. More people studying these fields than ever before
– Stanford University online course on AI: 160,000 students (23,000 finished it)
– More components / databases / tools /methods ready for re-combination
– Unexpected triggers for improvement (malware wars, games AI, financial AI…)
5. Transformation in society’s motivation?
http://intelligence.org/2013/05/15/when-will-ai-be-created/
(Smarter people?!)
“Sputnik moment!?”
9. @dw2
Page 9
Superintelligence – model 1
Village idiot Einstein
http://intelligence.org/files/mindisall-tv07.ppt
Eliezer Yudkowsky
10. @dw2
Page 10
Superintelligence – model 2
http://intelligence.org/files/mindisall-tv07.ppt
Village idiotMouse
Chimp Einstein
AI
50-100 years 50-100 weeks? / days? / hours?
Vernor Vinge: The best answer to the question,
“Will computers ever be as smart as humans?”
is probably “Yes, but only briefly.”
“The final invention”
Eliezer Yudkowsky
13. @dw2
Page 13
Recursive improvement
AI tools
AI
Intelligence
explosion
++Rapid reading &
comprehension of
all written material
++Rapid expansion
onto improved
hardware
++Funded by financial winnings from smart stock trading
++Supported by humans easily psychologically manipulated
14. Who here wanted to merge again?
Jaan Tallinn: http://prezi.com/xku9q-v-fg_j/intelligence-stairway/
16. @dw2
Page 16
Going nuclear: hard to calculate
• First hydrogen bomb test, 1st March 1954, Bikini Atoll
– Explosive yield was expected to be from 4 to 6 Megatons
– Was 15 Megatons, two and a half times
the expected maximum
– Physics error by the designers at Los
Alamos National Lab
– Wrongly considered the lithium-7 isotope
to be inert in bomb
– The crew in a nearby Japanese fishing boat
became ill in the wake of direct contact
with the fallout. One of the crew died
http://en.wikipedia.org/wiki/Castle_Bravo
17. @dw2
Page 17
Superintelligence – model 2
http://intelligence.org/files/mindisall-tv07.ppt
Village idiot
Chimp Einstein
Mouse
AI
Linear model of intelligence?
Eliezer Yudkowsky
19. @dw2
Page 19
Dimensions of mind
The ability to achieve
goals in a wide range
of environments
Being conscious?
Having compassion for
sentient beings with
lesser intelligence?
20. @dw2
Page 20
AI systems we should fear
Killer drones with
autonomous
decision-making
powers (Robocop)
Malware that can
hack infrastructure-
control systems
(e.g. Stuxnet)
Financial trading
systems software
(high speed)
Software that is
expert in
manipulating
humans
22. @dw2
Page 22
AI systems we should fear
Killer drones with
autonomous
decision-making
powers (Robocop)
Malware that can
hack infrastructure-
control systems
(e.g. Stuxnet)
Financial trading
systems software
(high speed)
Software that is
expert in
manipulating
humans
Software
that pursues
a single
optimisation
goal to the
exclusion of
all others
The more power such an AI has, the more we should fear it
23. @dw2
Page 23
The pursuit of happiness?
Software
that pursues
a single
optimisation
goal to the
exclusion of
all others
Software will do what we say, rather than what we meant to say
Wire-heading?!
Just make us happy!?
24. @dw2
Page 24
The pursuit of morality?
Just be moral!?
http://www.clipartbest.com/clipart-nTXa54XTB
Whose morality?
The problem of computer morality is
at least as hard as the problem of
computer vision (!)
http://tvtropes.org/pmwiki/pmwiki.php/Creator/IsaacAsimov
Isaac Asimov’s
Three Laws of Robotics?!
25. @dw2
Page 25
The two fundamental problems
of superintelligence
Specification problem: How do we
define the goals of the AGI software?
Control problem: How do retain the
ability to shut down the software?
Creation problem: How do we create
AGI software in the first place?
26. @dw2
Page 26
The fundamental meta-problem
of superintelligence
Specification problem: How do we
define the goals of the AGI software?
Control problem: How do retain the
ability to shut down the software?
Creation problem: How do we create
AGI software in the first place?
~No
research
~No
research
Some
research
Accidental
research
“Friendly
AI” (FAI)
“AI
in a box”
27. @dw2
Page 27
AI in a box?
Tripwires? “Adam and Eve” ethernet port?!
Software will be a tool, answering questions, not an agent?
The “answers” which the software
gives us will have effects in the world
(e.g. software it writes for us)
Systems which rely on humans to verify and
carry out their actions will be uncompetitive
compared to those with greater autonomy
AGI may become very
smart in surreptitiously
evading tripwires
Simple?
28. @dw2
Page 28
“The orthogonality thesis”
Intelligence and final goals are orthogonal
More or less
any intelligence
…could in principle
be combined with…
more or less
any final goal
29. @dw2
Page 29
“The instrumental convergence thesis”
(“AI Drives”)
Some intermediate (instrumental) goals are
likely in all cases for a superintelligence:
• Resource acquisition
• Cognitive enhancement
• Greater creativity
• Self preservation (preservation of goal)…
Steve Omohundro: “For a sufficiently intelligent system, avoiding vulnerabilities
is as powerful a motivator as explicitly constructed goals and subgoals”
30. @dw2
Page 30
Indirect specification of goals?
Specification problem: How do we
define the goals of the AGI software?
“Achieve the goals which the creators of the AGI would have wished
it to achieve, if they had thought about the matter long and hard”
This software will do what we meant to say,
rather than what we actually said (?)
AGI helps us to figure out the answer to the spec problem!
31. @dw2
Page 31
CEV: Coherent Extrapolated Volition
AGI should be tasked to carry out:
Our wish if we knew more,
thought faster,
were more the people we wished we were,
had grown up farther together;
where the extrapolation converges rather than diverges,
where our wishes cohere rather than interfere;
extrapolated as we wish that extrapolated,
interpreted as we wish that interpreted
Eliezer Yudkowsky
32. @dw2
Page 32
Unanswered questions (selection)
1. Can we turn ‘poetic’ ideas like CEV into bug-free working software?
– Should we humans concentrate harder on working out our “blended volition”?
2. How can we stop a superintelligence from changing its own core goals?
– Like humans can choose to set aside their biologically inherited goals
– Could AGIs that start off ‘Friendly’ become “born again” with new priorities?!
3. Can we prevent AGIs from developing dangerous instrumental drives?
– By programming (bug-free) in tamper-proof limitations?
4. Can AGIs help us to figure out a solution to the Control problem?
– Can we use a hierarchy of lower-level AGIs to control higher-level ones?
5. Can we prevent the rapid nuclear-style take-off of self-improving AGI?
6. Are some approaches to creating AGIs safer than others?
– Whole Brain Emulation / AGI de novo / evolution in virtual environment…
– Open (everything published) vs. Closed (some parts secret)?
7. How does the AGI existential risk compare to other x-risks in priority?
– Nanotech grey goo, deadly new bio-hazard, nuclear holocaust, climate chaos…
33. @dw2
Page 33
Answered questions (selection)
a) Should we be afraid?
– Yes. (End-of-the-world afraid)
b) Can we slow down all research into AGI, until we’re confident we
have good answers to the control and/or specification problems?
– Unlikely – there’s too much financial investment happening worldwide
– Too many separate countries / militaries / finance houses… are involved
c) How do we promote wider study of the Superintelligence topic?
– Need to lose the “weird” and “embarrassment” angles
– “Less Wrong” strikes some observers as cultish
– “Terminator” and “Transcendence” have done more harm than good
– First class books / articles / movies needed, addressing thoughtful audiences
– Good intermediate results useful too (not just appeals for more funding)
34. Practical philosophy!
Preparing humanity to survive the
forthcoming transition to superintelligence
Principal, Delta WisdomChair, London Futurists
Urgent!
(Roles too for mathematicians, theologians…)
David Wood
@dw2
Philosophy with an expiry date! Making a real difference!