1. Is the Singularity Near?
Nine key questions about the coming
Technological Singularity
?
Technology
David Wood,
Software Director (day job);
UKTA enthusiast, London
+ Many contributions
from the floor!
Varied expert and layperson
viewpoints welcome
20th Sept 2008 Time
2058
2. A friendly critique of
some of the ideas in
Ray Kurzweil’s
The Singularity is near:
When humans
transcend biology
3. The nine key questions
• Defining what we're talking about: What’s the relation
between the various different notions of the Singularity?
4. 1. Defining the Singularity
• “The ever-accelerating progress of
technology … gives the appearance of
approaching some essential singularity in
the history of the race beyond which human
affairs, as we know them, could not
continue” ?
Technology
– John von Neumann (1950s)
Time
2058
5. 1. Defining the Singularity
• “What is the Singularity? It’s a future period
during which the pace of technological
change will be so rapid, its impact so deep,
that human life will be irreversibly
transformed”
– Ray Kurzweil ?
Technology
Time
2058
6. 1. Defining the Singularity
• “When the first transhuman intelligence is
created and launches itself into recursive
self-improvement, a fundamental
discontinuity is likely to occur, the likes of
which I can’t even begin to predict”
– Michael Anissimov
Technology
Time
2058
7. 1. Defining the Singularity
• The advent of super-human general AI
• AI smarter than humans, for sufficiently many
purposes
– Including the ability to design and build new AI
• A trigger for recursive improvement
– A likely trigger for fast (and then even faster)
recursive improvement
• We could in a very short timespan have super-
super…-super-human general AI
• We are unlikely to be able to begin to conceive
what will happen next (Vernor Vinge)
8. Some related ideas (i)
• Exponential improvement in technology and/
or computing power and/or knowledge
• But exponential progress often stalls
– Speed of passenger airliners
– Heights of skyscrapers
– The human population
• Example of Kurzweil over-enthusiasm:
– “By the end of this decade, computers will
disappear as distinct physical objects, with
displays built into our eyeglasses, and
electronics woven in our clothing, providing full-
immersion visual virtual reality” (p. 105)
9. Some related ideas (i)
• Exponential improvement in technology and/
or computing power and/or knowledge
• This is an independent idea
– We could reach the Singularity slowly
– (Though the take-off, when it comes, would still
be spectacular)
– (Like reaching critical mass: possible slow
approach, but we still get a super-fast explosion)
• Pre-Singularity exponential improvements in
technology make the Singularity more likely,
but aren’t necessary to the argument
10. Some related ideas (ii)
• Computers drastically transform the
economy – doing more and more work in it
• This is an independent idea
– If/when the Singularity comes, the economic
transformation is likely to be very considerably
more drastic
• Society will change in completely
unpredictable ways, with existing structures
breaking down, like the Industrial Revolution
• I think we should concentrate on the more
specific idea of super-human general AI
11. AI singularity vs. nano-factory singularity
• A general purpose nano-factory can
manufacture goods better than humans can
• So it can manufacture an even better
general purpose nano-factory…
– Recursive improvement
• Economics would be transformed overnight
• An interesting topic!
– But probably not so scary/profound as the
advent of super-human general AI
• (The two developments could, however,
work in parallel…)
12. Eric Drexler, 1989
• “If you can build genuine AI, there are reasons to
believe that you can build things like neurons that
are a million times faster”
• “That leads to the conclusion that you can make
systems that think a million times faster than a
person”
• “With AI, these systems could do engineering
design”
• “Combining this with the capability of a system to
build something that is better than it, you have the
possibility for a very abrupt transition”
• “This situation may be more difficult to deal with
even than nano-technology, but it is much more
difficult to think about it constructively”
13. Hard vs. soft take-off
• AI smarter than humans, for sufficiently many
purposes
– Including the ability to design and build new AI
• A trigger for recursive improvement
– A likely trigger for fast (and then even faster) recursive
improvement
• We could in a very short timespan have super-
super…-super-human general AI: hard take-off
• This assumes there’s only one significant barrier to
building better AI – but there might be several
– Super-human general AI mightn’t automatically lead to
super-super…-super general AI
– In the latter case, we can talk about a “soft take-off”
15. The nine key questions
• Defining what we're talking about: What’s the relation
between the various different notions of the Singularity?
• Are there arguments in principle against the Singularity?
• What are the critical bottleneck determinants of
development towards the Singularity?
• What are the lessons we should learn from what’s been
called “the embarrassing history of AI”?
• Is the Singularity a plausible occurrence within (say) the
next 50 years?
• What’s the likeliest timescale for the Singularity?
• Should we be doing everything in our power to prevent
the Singularity from happening?
• Can we influence the Singularity to make its outcome
more likely to be good for humanity rather than
disastrous?
• What are the biggest uncertainties with the Singularity?
16. 2. Arguments in principle against
the possibility of the Singularity?
• “Computers will never be as intelligent as humans”
• “Progress with AI has never been as fast as AI
enthusiasts predicted”
• “There’s something mystical or vitalist about
human self-awareness and consciousness that
can never be captured in a computer”
• “Exponential progress always slows down after a
while – skyscraper heights, passenger airline
speed…”
• “An embodied super-AI will run out of resources or
energy before achieving anything God-like”
• “The likelihood is so small that discussing it is a
waste of our energy and a distraction…”
17. Could we keep the AI locked up?
• “An embodied super-AI will run out of resources or
energy before achieving anything God-like”
• Could we deliberately avoid connecting the super-
AI to the wider network?
• “AI-Box experiment” – Eliezer Yudkowsky
• The super-AI may well be an expert on human
psychology
– And would concoct extremely convincing reasons for
why it should be connected to the wider network
• In any case, we have to contemplate the
accidental route to the Singularity – when the AI
researchers get a better result than they expected!
18. 3. Critical bottleneck determinants?
• Improving the computing hardware:
– We’ve already had forty years of Moore’s Law
– That’s been driven by technology expertise coupled with
strong commercial incentives
– Experts generally expect Moore’s Law to last at least
another 10 years
– And there are good prospects for new types of computer
power even after that (3D chips)
• Human brain computational power is c. 100 TIPs
(100 x 10^12 instructions per second)
– Say this is an underestimate – use 10^16 TIPs
– Computers with this power will cost $1000 by c. 2020-25
– Assuming that sufficient financial incentives for
incrementally improved computing power remain in force
19. 3. Critical bottleneck determinants?
• Improving the computing hardware:
– We’ve already had forty years of Moore’s Law
– That’s been driven by technology expertise coupled with
strong commercial incentives
– Experts generally expect Moore’s Law to last at least
another 10 years
– And there are good prospects for new types of computer
power even after that (3D chips)
• Improving the software: This is my guess
– Wirth’s Law: “Software gets slower, faster than hardware
gets faster”
– Just calculating faster doesn’t make you wiser
• Understanding the human brain and human mind:
– May or may not turn out to be particularly useful
20. 4. Learnings from the history of AI?
• “Progress with AI has never been as fast as
AI enthusiasts predicted”
• => “Computers will never be as intelligent as
humans”
• Or, it will take an awfully long time
– So it’s mainly a distraction to discuss it
• Alan Turing predicted in 1950 that, by the
year 2000, machines would be able to fool
30% of human judges during a 5-minute test
• AI researchers need to be optimistic, to
ensure they get funded, but progress is poor
21. 4. Learnings from the history of AI?
• “Progress with AI has never been as fast as AI
enthusiasts predicted”
• Powerful algorithms eventually achieved for:
– Chess (but not yet Go)
– Algebra (including finding proofs to unsolved theorems)
– Real-time navigation
– Driving cars across deserts
– Quadrupeds that walk over rough terrain (“Big Dog”)
– John Koza’s Invention machine, which has won patents
– Composing music in the style of named composers
– Face recognition
– Language translation (Google)
– Playing twenty questions (Burger King)…
22. Possible unexpectedly fast progress?
• Technology improvements sometimes happen
outside the mainstream – happening somewhat
“under the radar” but with strong financial incentive
• “Adult entertainment” industry
• Computer games
– Incredible graphics (GPUs); Physics engines; “AIs”
• CAPTCHA defeaters (promoting spam)
• Search boxes that understand natural language
– Huge investment by Google, Microsoft…
• Virtual worlds: Ben Goertzel
– Bots there have fewer items of “accidental real-world
complexity” to worry about
• iPhone and other connected super-smart devices
23. 5. Is the Singularity plausible before
Hard to
(say) 2058? rule them
Vernor Vinge’s five possible routes to the Singularity: all out!
• The AI Scenario: We create superhuman artificial
intelligence (AI) in computers.
• The IA Scenario: We enhance human intelligence
through human-to-computer interfaces—that is, we
achieve intelligence amplification (IA).
• The Biomedical Scenario: We directly increase our
intelligence by improving the neurological operation of
our brains (smart drugs or otherwise)
• The Internet Scenario: Humanity, its networks,
computers, and databases become sufficiently effective
to be considered a superhuman being.
• The Digital Gaia Scenario: The network of embedded
microprocessors becomes sufficiently effective to be
considered a superhuman being.
24. 5. Is the Singularity plausible before
(say) 2058?
• Assuming there’s no major societal
breakdown
• Assuming there’s nothing mystical about
human-level intelligence
• Assuming at least some continuing
improvements in both hardware and
software
• … I see no reason to rule out the possibility
of the Singularity in this kind of timescale
• So it’s worth at least some attention from us!
25. “Maybe it needs 100 years”?
• 100 years of progress at the present rate of
achievement…
• Could be achieved in just 36 calendar years
• If the rate of achievement doubles every 10
years!
R(1 + 3.6)
Rate of
achievement
100.8R
R R (36 x 5.6)R/2
100R
2008 2108 2008 2044
Time
26. Example of accelerating rate of
achievement
• Sequencing the human genome
• Project started in 1990
• Cost of sequencing a base pair c. $10
• “It would take 1000 years to finish”
• Project forecast to finish in 15 years (2005)
• Actually finished in 2003
• Cost of sequencing a base pair c. $0.02
• HIV virus took 15 years to sequence
• SARS took 31 days
27. There’s more to acceleration than
Moore’s Law
• First computers were designed on paper and built
by hand
– Later computers benefited from computer-aided design
and computer-aided manufacture
– Even later computers will have even better computer-
aided design and manufacture
• Software creates and improves tools (including
compilers, debuggers, profilers, high-level
languages…) which in turn allows more complex
software to be created more quickly
• Technology reduces prices which allows better
technology to be used more widely, resulting in
more people improving the technology…
28. 6. Most likely timescale?
• I’m not qualified to say
• It depends on so many unknowns – and on
where society decides to invest effort
• “A new Manhattan project”?
• Perhaps already being carried out in secret
– In China / Singapore / …
– Inside Google / Microsoft / Apple / Nokia / …
– By DARPA…
• My guess: software will improve a great deal
in 20 years, with focused effort
• Add 10 years for contingency => 2038 ?!
29. 7. Oppose the Singularity with all our
effort? (Anti-Manhattan project?)
• Downside of bad Singularity is awful: disaster
• Much worse than Terminator movies…
30. 7. Oppose the Singularity with all our
effort? (Anti-Manhattan project?)
• Downside of bad Singularity is awful: disaster
• Much worse than Terminator movies…
• Upside of Singularity is debatable
– New cures, new medicines, new mind-trips
– Advice on geo-engineering and positive climate control
– Societal disruption
• “Better red than dead” – accept drawbacks of
vigorously controlling all experiments in AI, in order
to avoid the risk of a bad Singularity?
• But super-AI could improve “accidentally” or
“surreptitiously” – so we must create “friendly
super-AI” before “accidental super-AI” gets here
• “Moore’s Law of mad scientists” - Yudkowsky
31. 8. Can we influence the Singularity?
• Can we influence development towards the
Singularity to make its outcome more likely to be
good for humanity rather than disastrous?
• Or are we just passive spectators?
• We can contribute to the battle of ideas
– By debating and debugging them
– By communicating them – eg blogging / press
– By bringing them into the mainstream
– By influencing policy of think-tanks, governments,
universities, business researchers
• We can consider and possibly promote the ideas
of “beneficial super-AI” and “friendly AI”…
32. The project to create friendly AI
• AI that will in all circumstances maintain a respect
for humans, and will avoid harming humans
– Even through multiple generations of self-
reprogramming
– Even though the AI will have full access to all its source
code, and the ability to change any of it
– Even though the AI will observe many bad faults in
humans (and will be strongly tempted to “fix things”)
• The AI will not want to change this part of its
programming
– Like Gandhi would not want to take a pill to make
himself into a murderer
• (Depends on there being no serious bugs in this
part of the software!)
33. The project to merge humans with
super-AI
• AI that will in all circumstances maintain a respect
for humans, and will avoid harming humans
– Even through multiple generations of self-
reprogramming
– Even though the AI will have full access to all its source
code, and the ability to change any of it
– Even though the AI will observe many bad faults in
humans (and will be strongly tempted to “fix things”)…
• Alternatively, future humans may absorb and
directly interact with ever-improved AI systems
– “The man with three brains…”
– So super-AIs will find it much easier to respect these
future humans (“transhumans”)
• So we should study IA as much as we study AI
35. Summary (from Michael Anissimov’s
blog, Fri 19th Sept 2008) (1/2)
• Superintelligence is possible, it could have a huge
impact on the world, and our actions now may
influence the final outcome
• No fixed timeline
• No argument that all of history has been
predeterministically building up to this point
• No argument that technological progress is slowing
down, speeding up, moving sideways, or any other
such specific claims
• No particular attention given to pre-transhuman
intelligence technologies except insofar as they
influence when and how superintelligence is created
• Central focus on superintelligence as a distinct
technological milestone
• Acceptance of the point that deliberately designed AGI
may exist before neuromorphic AGI
36. Summary (from Michael Anissimov’s
blog, Fri 19th Sept 2008) (2/2)
• Acceptance of the fact that we might completely blow
ourselves up before the Singularity hits
• Acceptance of the fact that the first superintelligence
might not give a damn about us, and just decide to
rearrange our atoms into something more to its liking
• No magical rosy scenario where human upgrades and
AGI research coincidentally fuse seamlessly in a way
that happens to completely benefit mankind
• Acknowledgment of the Everest-sized challenge of
creating AGI that doesn’t eliminate us outright, rather
than hand-waving it over with “maintaining an open
free-market system…”
• Superintelligence is possible, it could have a huge
impact on the world, and our actions now may
influence the final outcome
37. 9. The biggest uncertainties with
the Singularity?
• Defining what we're talking about: What’s the relation
between the various different notions of the Singularity?
• Are there arguments in principle against the Singularity?
• What are the critical bottleneck determinants of
development towards the Singularity?
• What are the lessons we should learn from what’s been
called “the embarrassing history of AI”?
• Is the Singularity a plausible occurrence within (say) the
next 50 years?
• What’s the likeliest timescale for the Singularity?
• Should we be doing everything in our power to prevent
the Singularity from happening?
• Can we influence the Singularity to make its outcome
more likely to be good for humanity rather than
disastrous?
• What are the biggest uncertainties with the Singularity?