2. Evolution and Post-human Future -
Gregory Stock
• Where are the wonder drugs?
– Takes year for each clinical trial
– Approval process broken and much too slow and costly
– Approves only fixes to deficits, not enhancements
• Will Singularity lead to triumph of human values?
– More probably will lead to some form of end of humanity
– We are the old hotness – meat, blood, bone – not the
future
– “Consume my heart away; sick with desire and fastened to
a dying animal..” - Yeats
3. Evolution and Post-human future 2
• Evolution moves on
– Bio and complex non-bio is something new
– Non-bio intelligence is newer still
– Likely values among post-humans
• High levels of competition
• Cheap easy copies making death rather meaningless
• Uploading disengages humans from body
– But what of Moravec’s point about our minds being very wedded to our
evolution even in our metaphors and logic patterns?
– These beings will have little in common w/ humans
– “There will be a gradual elimination of all forms of beings
that we care about” – Bostrom
• For humans ready/willing/able to transcend, more of a
transformation, I think.
4. Evolution, Post-human Future #3
• Chance of preserving human values through
Singularity?
– Some super friendly near all powerful singleton
AGI may control and ensure it
– Thinks it is impossible
• “Emergent realm careening toward unknowable future
will go its own way regardless of our wishes.”
5. Evolution and Post-Human Future 4
• Signum – his company
– Targeting Alzheimers with goal of preserving enough of
brain at least to be worth freezing
– Noted that Alzheimers is helped by removing phosphate
buildups on proteins
– Molecule PP2A help this. Coffee activates this molecule
– Evidence coffee consumption decreases risk of Alzheimers
by 50%! Also adult diabetes.
– Caffeine is not the effectine agent. Sig1012 extract from
the coffee bean is
– Can move to human trials quickly because Sig1012 is an
approved food extract
6. The Mind and How To Build One –
Kurzweil
• Started off razzing critics
– Much of this and his talk was from or similar
material as @ Citizen Scientist
– Much ad lib (talk was teleconference)
– Critics include Doug Hofstader, Jaron Lanier and
Michael Anissimov (to a much smaller degree)
• Given 10**16 calc/sec for brain
– Henry Markram (Blue Brain) says this will be
achieved in 2018
7. The Mind, How To Build One 2
• Brain has Lisp nature?
– “..each cortical module is like a Lisp
statement..incredible hierarchy..”
– We have good and constantly improving ideas
how these modules work
– Says he believes a million or so lisp statements
could probably model the human brain (?!)
• Must have been talking with Minsky
8. AI Against Aging – Ben Goertzel
• AI applied to bioinformatics – CEO, Biomind
LLC
– Work in collaboration with Genescent
– Humans poor at understanding complex, high
component and relationship count systems
– This is where AI comes in:
• Searching for patterns and abstractions within large
genomic data sets
• Scanning relevant literature for patterns and
exploitable knowledge
9. AI Against Aging 2
• Why do we age and what to do about it?
– Hayflick limit
– Aubrey’s approach – fix all the main damage that occurs as
we age
• Many biologist skeptical esp. of unintended consequences of
things like plan to move mitochondrial DNA into the cell nucleus
– Antagonistic pleotropy
• Apparently changes/adaptations occur at many age points in our
development
• Unfortunately they stack on top of each other and interfere with
one another as more of them accumulate
• Our bodies literally try to run different age adaptations at once
10. AI Against Aging 3
• Genescent work
– Has bred flies that live 5.5x longer than usual
– Selective breeding like this would work in humans
if you did it for 5,000 – 10,000 years as it takes
hundreds of generations
– Long lived flies have a complex large array of
differences compared to regular flies. Requires
use of AI to mine the data for nuggets
– Looking for simple replicable critical factors
11. Extending Ourselves w/ Technology –
Steve Mann
• His eyecam is great!
– Everything he looked at was wirelessly broadcast and
displayed on the main screens
– Illustrated many points by drawing on a small paper pad
which he was looking at. The contents displayed on main
screen. Very natural and fluid
– Looks at audience and we see ourselves looking at him
looking at us
– He broadcasted and shared with world all his experiences
when out and and about for many years
– Has devised and worn wearable computers and
experienced mediated reality for over 30 years
12. Extending Ourselves w/ Tech 2
• Surveillance is a clear and present danger
• He originated Sousveillance
– Sur – from the top
• Authorities and such watching and controlling the people
– Sous – from the bottom
• People watching and controlling the authorities
• Wearable is better than ubiquitous
– More control over own data if on one’s person and only shared
as you wish
– Mediation of reality to remove unwanted stimuli, experience
and to augment reality
– Showed wearable chest camera like one MS now sells
13. Extending Ourselves w/ Tech 3
• Into new forms of interaction with tech and
environment
– Hydralophone
• Musical instrument that uses water through small holes that
the player closes and runs their fingers over to produce
complex wind instrument like sounds and chords
• Playing with these gives great tactile feedback and
experential shaping the water flow through each opening to
get different effects
• They have made these in many forms including large public
interactive sculptures and self play larger sculptures
• The model on hand was fun to play with
14. BCI Past and Future – Brian Litt
• Classification
– Open or closed loop (1 way or 2 way)
– Degree of invasiveness
• Generally the more invasive the finer the detail and control but
greater the risks
• BCI used today for
– Epilepsy
– Depression
– Obesity
– Parkinson’s
– Compensation for loss (hearing, vision, gait, artificial limb
control)
– Restore or repair (stroke, spinal cord trauma, peripheral nerve
injury)
15. BCI Past and Future 2
• Future BCI
– Augment : consciousness, memory, speed,
perception, cognitive processing
• Already controversial – olympics banned runner with
artificial lower leg as unfair to other runners
– Idea storage
– Transfer/sharing of knowledge, feelings, behavior
– Replay of experiences
– Direct brain recording
16. Machine Learning Rapidly Discovering How Brain
Works – DemisHassabis
• Nonbio approach to AI
– Symbolic AI is traditional way
• Formal logic, logic networks, lambda calculus, expert systems
– Flaws: brittle, time consuming, poor generalization, increasing cost of new knowledge in
some designs
• Bio approach to AI
– Use brain as blueprint
– If space of all possible designs yields only a few sparsely scattered
successes then good to start from a successful approach
– Problems
• 50 years from mapping entire brain
• That is not the same as understanding that part that makes for intelligence or
how it does so
• A human in a box (all of human brain) is not what we are looking for for AGI
17. Machine Learning 2
• A Third Way – System Neuroscience Approach
– Three levels of understanding brain systems (Marr)
• Computational – goals of the system
– Cognitive science and symbolic people want to focus here
• Algorithmic – how does system accomplish goals
– This area is largely overlooked in the main AGI argument
• Implementations – what is the physical realization
– Classic bio brain emulation people want to focus here
– So how do you find AGI relevant findings in 50,000
neuroscience papers a year?
• It takes at least 5 years of dedicated multi-disciplinary training to
come close to being good at this
18. Machine Learning 3
• So hybrid approad is to combine best of AGI and Neuroscience
– Some target areas
• Mirror neurons
• Model based vs model free systems
• Theory of mind
• Working memory
• Top down intention
• Concepts are key
– Three levels
• Symbolic – logic networks, symbolic systems
• Perceptual – HTM (Hawkins), HMAX (Poggio)
• Conceptual - ???
– Theory
» HC stores the memories of recent memories or episodes and replays those memories
during sleep at sped-up rate. gives high level neocortex samples to learn from memories
selected stochastically for replay. rewarded, emotional or salient memories are replayed
more; circumvents the statistics of the external environment and leads to abstraction.
19. Modifying Boundary between Life and
Death – Lance Becker
• Old notion of >4 minutes without oxygen is too late is
wrong
– Can resuscitate after 10, 20, 40 minutes – even an hour
• Lack of oxygen does note kill most cells directly
– They are fine for some time except build up electrons in
mitochondria and don’t regulate calcium as well
– Add oxygen at full normal values and they die
immediately? Why?
• The free electrons plus a lot of oxygen forms dangerous radicals
like crazy
• This destroys outright and/or triggers cell death response
20. Modifying Life/Death Boundary 2
• How can this be fixed?
– Cooling the body to slow down necrotic processes
• Standard cooling not fast enough. Invented slush machine for very
quickly (in minutes) bringing body temperature down
– Controlled slow reperfusion (reoxygentation) as heart is
restarted
• Gives system time to normalize
– Chemical cocktail to aid diffusing dangerous cellular conditions
as more oxygen is introduced
– This same process means that donate organs can be kept in
viable state much more easily and longer potentially solving
organ donor shortages
– Kit form being designed for use in ambulances and suitably
trained paramedics
21. Universal Measure of Intelligence –
Shane Legg
• He show an algorithmic method for determining
relative intelligence of AI systems
• Asks: Is computational intelligence going up as
Moore’s law goes up?
• How to approach the problem
– Internal properties of intelligence vs external
properties
• We don’t know and can’t say much about internal properties
• We can say a bit about external properties of intelligent
solutions
22. Univ. Measure of Intelligence 2
• What is definition of intelligence?
– He has collected over 80 distinct definitions
• “system that generates adaptive behavior for wide
variety of goals”
• “ability of system to act appropriately in uncertain
environment with appropriate being that which
increases probability of success”
• Summary: intelligence is the property of an agent that
interacts with its environment to successfully achieve
goals across a wide range of environments
23. Univ Measure of Intelligence 3
• General Formula for Intelligence
– Sum((2**-K(mu)) * V(pi, mu), All-Environments)
• K is complexity. As in Occam’s razor we won’t to disvalue
more complex solutions compared to simpler ones
• The agent is pi
• An environment instance is mu
• V(pi, mu) is success function for the agent in an environment
• So summing the weighted performance of the agent over all
environments possible for this agent gives us the measure of
the agents intelligence
– Of course in practice we cannot usually enumerate all
environments
24. Univ Measure of Intelligence 4
• Evaluating intelligence
– So use Monte Carlo approximation (random
sampling of generate environments)
– Actually running this has successfully classified
many AI systems correctly
• May be sensitive to perturbations in the environment
sample so must do many runs to converge to more
trustworthy value
Notas del editor
The Singularity Summit was mind blowing. There were a LOT of very interesting talks – far too many to do justice to today. But here are a few that I was most taken by. Even that subset leaves off some I really enjoyed.