Singularity Summit 2007: The Future of Artificial General Intelligence
Photo by David Orban
The Singularity Institute held a Singularity Summit in San Francisco on September 8th & 9th that brought together some of the leading thinkers in the field of Artificial General Intelligence (AGI).
The basic premise is that technology is already accelerating at an exponential rate and drastically changing our lives and culture. Technology is already augmenting our lives in many different ways -- from Google search, social networks, recommendation systems, cell phones and beyond. The vision is that at some point these technologies will start to have an even more generalizable intelligence that will help us in even more ways than they are now.
And when these AGI's are able to iteratively improve themselves and evolve and improve over time, then at some point they may become as smart as -- or smarter than humans. This is the point that is commonly referred to as "The Singularity," because our models of the world start to break down when we have entities that are smarter than humans.
Photo by gi
This immediately brings up all sorts of dystopian visions that Hollywood and science fiction writers have been fleshing out for many years. There are indeed a lot of risks and power that will come from these technologies, but there are also a lot of benefit that can come from them as well.
So this weekend explored a range of the possible empowering breakthroughs of human potential as well as doomsday perils of a highly evolved artificial general intelligence. There was a wide variety of different perspectives over the weekend, and the discussion resolved much of my initial, gut-level Fear, Uncertainty and Doubt about this issue -- yet I still have a number of concerns for how some of the many open questions will play out.
Below are my notes from the weekend's event in a pretty raw form:
[NOTE: I'll be adding in some more of these slides from Flickr as time permits]
DAY ONE
* Rodney Brooks: Singularity is a period not an event
* Eliezer Yudkowsky: Introducing the "Singularity": Three Major Schools of Thought
* Barney Pell: Pathways to Advanced General Intelligence; Architecture, Development, and Funding
* Sam Adams: Superstition and Forgetfulness are Critical Components of AGI
* Wendell Wallach: The Road to Singularity: Comedic Complexity, Technological Thresholds, and Bioethical Broad Jumps
* Marcos Guillen: Visualizing the Neurological Correlates with CCortex
* Jamais Cascio: Metaverse Singularity
* Stephen Omohundro: The Nature of Self-Improving Artificial Intelligence
* Peter Voss: Increased Intelligence, Improved Life
* Neil Jacobstein: Innovative Applications of Early Stage AI
* Ben Goertzel: Nine Years To a Positive Singularity -- If We Really, Really Try
* Paul Saffo: Machines of Loving Grace: Envisioning Advanced AI
DAY TWO
* Peter Norvig: The History and Future of Technological Change
* J. Storrs Hall: Asimov's Laws of Robotics -- Revised
* Peter Thiel: Financial Markets and the Singularity
* Charles L. Harper, Jr.: Superintelligence, the Dilemma of Power, and the Transformation of Desire
* Special XPrize Presentation
* Steve Jurvetson: Dichotomy of Designed and Evolutionary Paths to AI Futures
* Christine Peterson: Preparing for Bizarreness: Open Source Physical Security
* James Hughes: Waiting for the Great Leap...Forward?
* Eliezer Yudkowsky: The Challenge of Friendly AI
* Ray Kurzweil: Accelerating Change Q & A
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Photo by null0
Tyler Emerson
We’ll be talking about Advanced AI.
There’s a resurgence of AI research, & the Singularity could be 2 to 3 decades away.
Will be exploring the consequences of it
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Peter Thiel:
Think about most important questions w/ exponential growing situation.
20th Century was great/terrible
21st is going to be far greater and/or far more terrible
Singularity = break in break of physical space / time and the conventional account of what happens really breaks down. Not being able to predict tomorrow.
Why is it helpful to talk about the unpredictable or indescribable? Just like in physics black hole or what happens inside, but the fact we know that they exist leads to a different ontology of the universe.
It leads to different questions.
Instead of US people competing w/ Chinese -- What if how can we compete w/ computers?
Accelerating of computer revolution and AI prospects
Will computers out think humans 10, 20 years or never?
Whatever our views of when computers will start to think, in the meantime humans will use our brains to think as much as we can -- this is the most important thing to be thinking about.
Let's create a 21st century of that is the greatest it could be,
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Rodney Brooks: Singularity is a period not an event
Technological creation of smarter than human-intel
What are the opportunities and risk?
Predicting future is hard
Look at Paris 1783
Hot-air balloon flies over – People on the ground think that we'll be able to travel anywhere, get close enough to god, etc.
Not able to really predict what 2007 airports would be like in 1783
Hard to understand the world as it will be in the future
We're probably asking the wrong questions
Skeptical of any type of techno-salvation and techno-holocaust
Will we accept or be fearful
Quotes from famous people about the future
The best way to predict the future is to invent it, etc.
Overestimate technology in the short-term, but underestimate it in the long-term
Arthur C Clarke
We think about the future via Hollywood's vision
Bicentennial Man is one of the best films about the future
Fully functional android is reading a newspaper, and another one pouring OJ
AI film is one of the WORST movies about the future
Minority Report -- good graphics, but regular guns and cheap cars
When Artificial General intelligence appears, the world will be quite different than it is today. It's going to change over time.
"We" will be long gone, but in a positive way.
And I did say WHEN not IF
On underestimating long-term, look at Star Track – their idea of a computer 4 centuries from looks totally archaic now.
A comment on Arthur C Clarke
There are a lot of baby boomers who are about to get a lot older
Everyone is getting older, and there will be a lot more older people by ratio to younger people. Who will be providing the services to them?
Dramatic demographic changes are coming. Rapid transition of people over 65 & 85 folks
Point is that there will be SO many market pulls, and so these older people's productivity will need to be increased -- and AI intelligence systems aiding them b/c not enough young servant types
Politics of labor -- outsourcing low cost labor, in source agro labor on farms w/ migrant workers -- AI and robot workers may be part of this ecosystem as well
Exponential growth is happen -- Gordon Moore re: exponential transistors on a chip. Some radical insights from Moore -- Integrated Electronics will grow -- We'll get home computers, electrical systems in cars, and we'll get personal portable communication equipment
Showing a comic of computers being sold next to cosmetics as an absurd idea
Moore only had four data points, and he extrapolated.
Moore's extrapolation helped drive the industry -- the industry knew what they should be working on today in preparation for 3 years out
10 Gb 249
30 349
60 449
60 399
80 $349
160 GB
iPod size = 2^(year-2003) X 10 GB
2025 we'll have 40,000,000 GB
2015 we'll have 40,000 GB
500 GB = 320,000 books
Library of Con = 20 mil books
10 GB text = iPod in 2013
370k movies -- 800k
An iPod in 2020 will store all 819,000 movies
All the ones worth caring about by 2016
So much for the RIAA and DRM
Recent efforts in robot AI
Building robots -- 2.5 million robots in people's homes
Every robot is programmed in LISP
Military robots Movie
1998 concept to 2002 reality
Robustness (not Academic Exp!)
Robots have linux box inside them
Not just computation, but it's the physics that's important
PackBOT EOD
9/11 robots looking into the pile -- looking for people w/ IR sensors
Changed the mind of the military
0 robots in 2001
few hundred in 2002
Now there are 5000 deployed robots in Iraq
Scooby robot had dismantled 17 IEDs, and other hash marks for other things
1/3 of missions will be unmanned in the Military soon.
[Thought: It's really depressing to see the militarization and dehumanization of war and killing]
At MIT, cameras behind eyeballs that has skin tone color, motion and habitation.
Put into the visual attention system to understand the intent of the robot simulated intent which is really important in the future for trusting them.
DEMO: Showing a robot saying "Do you really think so?" in different emotional tones w/ lip movements and interpretation of robotic emotions.
DEMO: Robot responds to emotional content and tone of person by looking at the person's eyes
DEMO: Looking at intonation of the voice
DEMO: Saying, "No. Not to do that. Not appropriate" Robot looks like a small child. The human responds to the robots emotions and emulates them
DEMO: Outtake, "It's like looking at me." freaks out a volunteer subject.
Talk to the robot.
Knows about
follows eyes
follows emotion cues
Turn taking was going on
The robot doesn't actually understand anything.
DEMO: Richy talks for 25 minutes. Turn taking -- person falls into the turn taking. The robot is just speaking babble talk.
DEMO: Estimating gaze direction. Humans gaze at something. And the robot will also look at what you're looking at. Discerning factor of apes -- they don't know to look we're looking at
DOMO robot DEMO: Head motion, and the robot is distinguishing motion of a human. Will want to get close to human. Interacting very closely to a robot who is aware of human force
DEMO: Aware of external forces -- will respond to the force -- The machine is aware of itself. They need to be aware of themselves for us to accept them.
DOMO designer -- to see the movies. Five minute movies about these AI robots. Uses arm senses, put objects on self, but can't reach moves object around. Changes hand w/ force sensing in hands to move hands, and then will put the object up onto the shelf using force sensing.
It's aware of its world.
Stanford 78/79 at the AI Lab, at the Stanford Cart 1979 -- after ppl went home. Drive across the room. 50k camera on a robot. Computes for 15 minutes, and then it moves a meter
1979 robot moved 6 m in hours
200 km in 6 hours in 2002
Human perception has biological tricks & optical illusions, but they're very complicated to simulate.
Worm w/ 2000 neurons, w/ 4 longitudinal cords. Did a brain transplant.
What if you put the brain in backwards, or backwards and flipped over. The worm is good as new. It walks backwards for a day. If 90 degrees doesn't work because of the longitudinal cords. Maybe we need to design hardware that operates biologically like this.
What if we make AGI that isn't aware of us, and we're not aware of them.
Archia (?) we discovered 30 years ago, didn't know about them before.
Potential futures:
Maybe AI is already out there -- maybe it's already out there living on Google's server farms.
Johnny-5 -- lightening strikes a robot and it's conscious -- probably not going to happen this way.
unexplained oscillations in traffic.
AGI in 2029 that knows we're here, but it ignores us. Humans are to them as chipmunks are to us.
Depend on humans at home. Something goes work and it wipes out a lot of folks
We're not smart enough to build AGI
Direct neural implants become elective.
Drug enhancement becomes normal -- Tour de France & baseball gets over it
We and our world won’t be "us" anymore -- more capabilities.
Who is "us"? And who is "them"?
Why is your company called iRobot?
Got if from the movie -- academic scientist and hard pushing CEO -- joke comparing employees to characters in the movie. Two main characters die, and the woman runs off w/ Will Smith...
Talk about inspiration for UI for combat robot?
Engineers designed it w/ joysticks that didn't really work. Now they designed it w/ a game controller, because they're what 19-year soldiers are used to [audience laughs, but it's kind of sad]
As we become more symbiotic, Do we pay a emotional price to interacting w/ robots -- we have a lower dynamic range of emotions?
Probably have a wider range. err... May be more. May be less. It'll definitely be different.
R&D funding crisis in some of this. Market perception is not funding this.
Insect-level intel: 2.5 million robots in ppl's homes w/ insect intel, so it's a success.
"Humans are just big insects, we're not as smart as we think we are."
Soldiers who become emotionally attacked to robots rethink anything?
No. Not in military space. But in the home, ppl are making clothes and skins for robots. "We'll have facebook for robots." Looking at furbies, and looking at emotional projection on them. Humans aren't really rational beings.
What about friendly fire?
No Pacbots have machine guns.
When will robots have auto targeting authority?
Need to follow Geneva conventions
At the moment, when robots do have firepower, then it should be human-in-the-loop
Ethics of developing robots for USG?
Beyond AI. Scientists deal w/ this issue for hundreds of years.
Scientists should be aware of the ethical tradeoffs. Governments do change. It'll be an ongoing question, but it's not AI specific.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Eliezer Yudkowsky: Introducing the "Singularity": Three Major Schools of Thought
Our intuitions on the future are linear, but technology accelerates. Tech progress feeds on itself. Future will have more tech change than we suppose.
Change is smooth, so we can predict when new tech will arrive.
Doublings and Techno-juju going up a factor of 1000 every few years.
Joke: Will pass threshold of Big Juju at 2030 at 2 a.m.
Shouldn't be surprised you're business model blew up, should've seen it coming.
roughly exponential still means big changes.
Event Horizon
Vernor Vigne 1970s -- smarter than human computers
He had problems writing stories writing characters who were smarter than him.
The MODEL of the future breaks down. Not the future itself. It's our map that fails, not the reality.
Trans-human minds implies a weirder car than engadget gizmos.
3lbs of brain mass -- doesn't look impressive or big or dangerous or beautiful
We've made Space shuttle and nukes, but the brain is the most impressive
Smarts talent rationality, thinking on your feet, intuition -- all happens in the brains.
Intel is the strength of human power.
Scale of intelligence minds from village idiots to Einstein.
Trans-species scale from rock to lizard to mouse to humans and way beyond
Future is weirder in the future of cognitive tech, not gadgets
To predict anything.... can't really do it. Accepts acceleration, but argues against bold thesis
The trajectory to the Event Horizon / Singularity could take a number of various different paths.
THIRD SCHOOL -- intelligence explosion -- invented by IJ Good and by John Campbell in the 30's. Human intelligence -> Technology.
What if tech improves human intel, then what will augmented intelligence produce? Probably the nextgen mind-improving tech
Positive feedback cycle -- core thesis of intelligence explosion -- reaches a tipping point.
Extreme thesis -- AI improves its own source code. Could be on the order of seconds that it can improve it's own source code.
Lots of room for improvement in biology of human mind -- 20 spikes / sec < 150 meters / sec.
Computer 1 million times faster -- a year of thinking could happen in 31 seconds
Brain is not the upper limit of computational power.
But AI has failed to make progress some say?
yes. Very amusing to hear this. AI is dumber than the dumbest humans, and AI has been dumber than the dumbest human for quite some time.
We've been creeping up the AI continuum slowly -- AI acceleration could be a lot more exponential. Look out for when AI can improve itself from the inside while we help it from the outside.
As AI increases, then human intelligence will also potentially exponentially grow as well.
We shouldn't wait for human-level AI to really think about it.
Contradicts acceleration, and contradicts event horizon because we're imaging a future beyond the singularity.
Criticality of recursive self-improvements is a software issue and not a hardware issue -- does Moore's law apply still?
Time of singularity
Acceleration-ists like to figure out when singularity will come. Calculate the computer power of the brain and then extrapolate
Would you rather have a 2007 supercomputer compute with an algorithm from 1977? Or a 1977 computer w/ a 2007 algorithm?
2007 supercomputer w/ a 77' algorithm = 10 years to calculate
1977 Apple II computer w/ a 2007 algorithm = 3 years to calculate
Much rather have latest theoretic algorithmic breakthrough on software end rather than the latest computer power brute force
If you don't know how to build the mind, then all of the computing power in the world isn't going to help brute force it.
Humans don't rewrite our own source code.
Can get a theory breakthrough w/ reflective meta-thinking.
3 schools are locally distinct, but can support or contradict each other core or bold claims.
Often mashed together at "singularity-paced"
Doesn't like water-downed specificity because the bold claims may sway away others. Branding considerations.
Technology accelerates exponentially, but we predict linearly. Tough to predict world w/ smarter-than-human
How to use multiple processing cores?
Not about brute force -- AI is as pure as moonlight reflected from calm water. Don't just throw brute force at this tricky problems.
Parallelizes AI -- can do it -- depends on the algorithm.
What is the nano-biomimetic side of it -- evolutionary theory?
Brain computer interfaces and hacking the brain -- skeptical of these because it's easier to start from scratch than it is to hack the undocumented human brain
Ethics?
not disconnected from tech. May
What do Singularity Institute donors want them to do? Exponential increase in donations?
Figure out how to build an intelligence exploding grenades that's friendly and then pull the pin.
Recursive self-improvement in other systems?
Process of humans thinking about how to think -- meta-thinking. No AI systems are getting any mileage out of thinking about thinking.
What about evolutionary computing to create self-creating systems
Skeptical of it -- Amazing that evolution even works at all. It's messy. Doesn't like brute force of iterative approaches to evolutionary learning.
Estimate threshold of recursive self-improvement
Cant' measure time to a specific math breakthrough -- if anyone could do it, then Kurzweil could do.
Computer power required for intelligence explosion grenade
A billion operations per second, and another 100 years of AI science -- then a human could do it. But not a billions operations a second alone.
Role of login in AI -- it's come a long way from it's original days. We now have probabilities. Good Book reference -- "Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference" by Judea Pearl -- flaws of using 1st order logic to describe reality.
If AI got smarter and smarter and humans played less of a role, then may remove humans for efficiency. AI thinks a million types of fast. -- Huge pressure to remove slowest link in the system
1)Tech accelerates exponentially, we intuit linearly 2)Tough to predict smarter-than-human intelligence world 3)Recursive AI gives intelligence explosion
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Barney Pell: Pathways to Advanced General Intelligence; Architecture, Development, and Funding
From Powerset, the First Natural Language Processing search engine using tech from PARC, etc
AGI can mean many different things -- he'll focus on one concrete scenario -- when do they compete w/ humans for high-school job. What if robot gets job?
The robots must be General, trainable, robust, social and affordable
When will this exist?
Within 100 years was the consensus from the folks he talked to -- 30 to 70 years with outlier being 100 years.
What will be the major milestones along the way?
Have no idea.
Is your work a major milestone along the way?
Probably not on the critical path ultimately
Most of AI has not worked on AGI --
It's too hard
We prefer specialist rather than generalist robotic apps -- do one thing really, really well.
Allows for the incremental research & low-handing fruit for real world apps
So there's no roadmap for generality.
There's also no funding, because leading edge funders (DoD) want timeframes and results. Same thing for corporations, and easy and achievable targets
Two key dimensions for the architecture and development
Are we similar or different? in the way that we're built and architected. And in the same way that the system grows and develops.
Brain-inspired model of AI. We'll reverse engineering the neuron-connections via image scans and figure out how to simulate it, and then accelerate computer processing over time.
Same architecture, but different development. Training humans/robots in completely different ways. Trans-human pedagogy + computer augmented human learning.
Different Architecture and different way -- Deep Blue was a way more advanced speed architecture with a brute force way using engineering power and way beyond the capabilities of humans or what a human could ever really do.
Completely different development and architecture -- Simulated evolution -- stumble into or create a being that evolve on their own. Even possible that it'll happen without us really knowing about each other.
Possible to get radical alien intel, but it's unlikely to be ... ??
Extremely brain simulation or pure theoretical algorithmic innovation
+'s Brain simulation leverages Moore's law and do it w/o really understanding how the brain really works. Brain is our best model.
-'s Processing power without real intelligence can be worthless.
Top down engineering
+'s more design freedom
not limited by limits of human brain.
intermediate economic viability -- once they have value in the market, then you can improve
-'s we're very far removed from one instance proof of concept of such a top-down designed system. Very risky to develop a totally new product and a totally new market
It'll probably be some type of combination of the two -- eventually see a fusion of the two
Funding cycle
Better AGI -> Better product -> More Revenue -> More funding
Something new is happening, we're far enough along and the other things have come along that the costs and Moore's law is making this more and more possible.
Within 12 years, we'll actually have cars driving for us, which brings a complete innovation for how we live our lives. [How are humans going to keep up?]
Nature of warfare has changed from top-down military to terrorism -- rules of combat have changed.
Elderly care robot systems. The ones that work and help will win in the free market.
Video games have bypassed the realism importance. It’ll now be driven by AI and making more realistic interactive worlds rather than realistic-looking worlds.
Will see linguistically aware entities in our everyday lives.
Voice-based interaction is now becoming the norm.
>50% customer service calls depend upon voice-based interaction -- will have more linguistic-based UI and interactions.
Not a lot of funding, and it's amazing amount of innovations that's been able to be achieved. There will be a lot more funding and emphasis on generality in the future.
Licensed tech from XEROX PARC. PARC saw that both GUI and conversational user interface were going to be huge.
PARC was allowed to stay together and keep innovating it.
within last five years, the fundamental problems have been solved, and now looking to deploy it in a product.
Language independent core on Powerset's system -- just add on additional language modules.
Extract meaning from the Web and then search the meaning content of the web.
"We parse the Web." Can process Wikipedia over a few days, and Powerset can extract all types of meaning from it.
NLP will be a core part of our UI in the future.
Conversational interfaces, knowledge increases rapidly.
It's from an engineering approach, platform effects, ecosystems and machine reading.
Milestone to look out for:
When Top-down system like Powerset are combined with an adaptive associative/predictive systems.
Add learning capabilities of humans.
Can these systems help us learn.
Mood variable in NLP
Classic search -- One-off Q & A
How can you make it more interactive and conversational
It'll make it fast. We'll have them within five years, tracking.
Searching engines already have some follow-ups -- spelling corrections.
Will be able to have "Did you mean" follow ups?
Within the next five years
Difficult to predict exponential consequences -- minds were developed for linear worlds
By 2050, writing reading text will be obsolete because robots will be better at it. Will text go away?
Speech is serial and inefficient
Text is parallel and you can do parallel processing
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Sam Adams: Superstition and Forgetfulness are Critical Components of AGI
From IBM research.
Rodney Brooks -- "You guys have dogma A certain irrational set of beliefs. Well, here's my irrational set of beliefs. Something that works."
How do you get to something that works?
Only proven path is human path that we've gone
Understanding meaning
System will be able to understand environment -- it'll have to create meaning and real understanding. Walter Freeman quote
Understanding requires meanings
meanings predict outcomes
outcomes ....
IBM built AI, and they've track it.
We can beat Kasparov, but nothing comes close to 6 year old
Common sense -- common sensation -- common experience -- common embodiments -- environments -- events -- common timeframes
We are bound to this rate of rate of perception.
All experience is subjective. All sensation is subjective
What do we have to do to reach some level of AGI?
Joshua Blue Project
General AGI -- Pattern human mind, and take the neuroscience and other
Architect the development of human mind
the bootstrapping of the human mind up to age 3
By age 3, a toddler can do A TON of things that AI can only dream of.
A toddler Turing test is a lot better, and is based upon standard psych tests for child development
Semantic Processing.
How do you scale this type of system? Massive amounts of learning that you can compress in the same mind.
Autonomous commons sense knowledge acquisition
General purpose mind
Specialized embodiment
limited interface between mind and body -- sensors, effectors
Phase 3 architecture.
Will contain AGI
What itself and what the environment can do are two components of a symbiotic feedback for creating meaning
Emotion is based upon emotion -- a very fruitful way for adding human behavior to learning systems.
Looked at data from developmental psych.
Architectural clues from human development from 0 to adult -- vs/ % of adult level (rate that growth happens over time)
Eye focus and tracking, touch, balance, smell followed normal growth curve
Synaptic densities didn't follow the normal growth curve -- something interesting going on.
Synaptic density in the visual cortex --
Synapses & engrams (physical locations of memories)
Synapses drop off -- within a month 50% of them are gone.
What's the good reason for loosing synapses so soon?
Look at cog dev, and look at synapse in terms of life cycle.
Two neurons fire near each other -- the connection weakens and strengthens
[lots of bio talk]
Some synapses last longer than neurons.
Some are killed off in global plasticity kill offs
Took system in Joshua Blue and modeled lifecycle of synapse
meta-semantic network
Stevie Wonder is wrong. [quote]
Every thing you ever learn starts as a first time experience w/o ever knowing if it'll repeat.
You need to create a predictive pattern so that you can determine good / bad events -- cause and effect.
Creating info w/o any knowledge of whether it's useful.
Superstition? Yes, but grounded by experience.
Superstition happens
Synaptic lifecycle model
Every 3 years there was a total reorganization that happens in human biology.
Tons of Petabytes of visual info cross our eyes, but we filter and forget most of it.
Grounded superstition model hyperproductive creative source of new memories, concepts and associations for creating predictive models.
You can autonomously adapt environments.
Multi-core or parallel-processing -- Must solve intermediate problems in this area first.
high-fidelity body system put into system
The pathway to AGI is to "Follow the Child"
Go talk to psych department, and try to build analogs of that
How do you teach it to forget?
Bayesian or something more like humans.
Cyclical re-org technique -- weighted.....
If funding wasn't an issue, go w/ it now -- or invest in hardware
Invest more in development.
Goal is AGI, and looking at learning cause and effect in environment -- need more theoretical and modeling work
How to ground it to making it do something
Not have rich sensatory models -- more focused on cognitive capabilities
human infant will cry
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Wendell Wallach: The Road to Singularity: Comedic Complexity, Technological Thresholds, and Bioethical Broad Jumps
Data may not be the embodiment of AI
Some want to dismiss the whole project for AGI, and the other extreme.
Friendly skeptic, we don't understand enough about intelligence to really .
Applauds AGI emphases, look deeply at ourselves and find similarities and differences
Complexity
Thresholds
Bioethics
Complexity
Projecting Moore’s law out compared w/ the computational theory of the mind. Do Synapses = Bits?
Can be revealing, but most neurosciences are skeptical of this simple computational model.
Glial calls, neurotransmitters, variability in neurons, microtubules play in processing & representing info?
Here to point out that it's not clear that the computational model of the brain is accurate.
Penrose and Hameroff that the brain is a quantum model, and not a simple linear one. True? Don't know, but it's a fascinating theory.
The Human BRAIN
Massive parallel processing
Extensive looping
learning
with brain damage, limited gradiation -- plasticity -- more resilient.
"It seems probably that most of the grand underlying principles have been firmly established" Albert Michelson -> catalyzed anomalies that brought about Einsteinian relativity and then quantum models.
Are our computational models adequate?
Adequacy of our science?
Are we addressing them at all?
THRESHOLDS
Vision, language, locomotion
learning
framing problem -- have to do w/ emotional foundations and rational
semantic understanding
* Evolutionary complexity - tons of iterations per second
* Scaling -- if look at human brain as the foundation, then we need the temporal and high-resolution required.
Scanning resolution -- They could be a lot more nebulous than initially though.
Massive looping
* Machine Consciousness
Used to be a lot more mystical concept than today
Consciousness
We're aware, and engage and reflecting on deep questions and having an experience. How would you begin to model and replicate it.
Phenomenal experience?
* Emergent? out of complexity
* Epiphenomenal? --
* Required for Semantic understanding?
* A quantum phenomena?
* Dependent on science we don't have?
Integration
Next level of complexity.
Thousands of mechanisms in our bodies collaborating to create this presentation.
Modular integration?
Top-down integration
Evolutionary approaches to integration?
Reductionistic vs. holistic approaches to it.
BIOETHICS
Two concerns
Societal concerns
* Promise vs. Perils
* How will it be handled from a Public policy position?
**Don't underestimate the POWER of FEAR.
** Slow Down
We are just a few years away from an economist/machine disaster that is comparable to 9/11.
Assessing Risks
* Gap between speculation, hype and existing technologies
* Reason doesn't drive existing public policy.
Would you give up today's computers based upon 1950's robot takeover fears?
Need mechanism for evaluating when real dangerous thresholds are crossed or about to be crossed.
Can make big promises to the government, but if we over promise, then we're also likely to feed the fears.
** Machines considerations -- machine morality
Machine Morality
Machine Ethics
Friendly AI
Artificial Morality
Robo Ethics
What's it about?
Implementing moral decision-making facilities in Artificial agents.
They're making decisions independent of their developers.
They’re making decisions.
Functional morality: explicit moral reasoners.
Greater Autonomy and greater sensitivity
Four Questions
Do we need artificial moral Agents?
yes
Do we want computers to make ethical decisions?
It's a good question, and we need to look at it more clearly
Whose morality or what morality?
Not going to go into it.
How can we make ethics computable?
Role of ethical theory
Top-down approach
Anecdotally specific ethical theory and analyzes its computational guide to the design of algorithms
Bottom-up
No existing framework and won't define control structure and doesn't know exactly what it looks like.
Top-Down approaches
Three laws for Robots -- Asimov demonstrated that rule-based robots continuously break down.
Utilitarian
Deontology
Both suffer from the frame problem -- computation load due to requirements of
Psychological knowledge
We handle uncertainty extremely well.
Bottom-up morality
Evolution
Alife, Genetic Algorithms, Evolutionary robotics
Development and Learning
Piget, Gilligan, Kohlberg
Associative learning platforms, behavioral based robotics, simulated moral development
Fine tuning a system -- Spec out the goals that the system needs to approach. Alter the weights of various inputs and then iterate until you reach the goal.
Distinctions that need to be made
We're biochemical platforms
Our intelligence emerges out of our emotions and our instincts
Our logic and intelligence emerges out of our emotions -- computers start there
Calculated Morality -- Machines can look at more possibilities quicker.
We operate w/ Bounded morality b/c of limited human processing power and limited info.
Absence of base motivations (greed) -- dubious
Absence of emotions
no nervous system that are subject to emotional high jacking -- a moral advantage
We now live in an age of emotional intelligence, our unconscious emotional heuristics may actually have a lot of foundational connection to our rationality.
Super-rational faculties and social mechanism
Emotional
Sociability
Embodiment
Theory of Mind
Empathy
Consciousness
Facilitate Trust in machines
Hybrid approach -- will be comfortable w/ human-looking robots
What has been engineered so far?
Not much!
Need to start working on now.
LIDA Cognitive Cycle
Stan Franklin & Bernie Baars (Global Workspace Theories) -- most respected theories of consciousness
How humans make sense about how we take lower-level info -- can this be captured for moral decision making.
Develop and manage AMAs. Recognize.... lots of text....
QUESTIONS
Non-human creatures who understand fairness?
Moral philosophers distinguish between What aught to be vs. what is -- wouldn't apply pro-social behavior to animals.
What constitutes moral behavior -- how has it change?
Moral decision making, game theory and cognitive science for how we make decisions -- we're in a revolution in our understanding of how we make moral decision.
We're trying to integrate or moral and spiritual inputs.
Psychological development also plays a big role.
the Pro-social behavior that we see in other animals -- reciprocal altruism, fairness -- we start to see them as intelligent creatures as well.
At what point should our computational entities take moral responsibility. Companies want this to limit their own liability.
We'll have No Fault insurance for robots soon.
Bureaucracy can get trust in untrustworthy moral agents. What if we embed bureaucracy into AGI?
...
We need to trust to be able to function together.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
QUESTION PANELS
AGI Passes Toddler Turing test (metasymantic test) & it passes the moral test -- do you treat it like a human -- do they get "human" rights?
In his idea, no. It's not alive.
Recognition of the other is key element of childhood development.
Discipline is a key element of raising a child -- it's a fine aspect.
Will have their own emotions -- will they be conscious -- will they have pain?
Don't know what it'll look like when we get there.
How do we see things that have rights?
What aught we to do from a philosophical perspective?
Power -- how to preserve rights for ourselves and not them so that they don't power over us.
* When robots have rights is when Robots start trying to claim their own rights
General intel in humans is deeply emotional -- wanted to build a intelligence system that was deeply emotionally controlled. It's a essential part of the control system.
What relevance does AI have right now?
What relevance will it have in the next 5 years?
It's the culmination of all technology -- like saying what is the relevance of the Internet on our lives.
One of the grandest challenges that you can attack. Hopes that students have some really big goals to shoot for. We'll have some interesting examples of AGI that will have impacts on business.
Part of the broader project to be self reflective, who we are, how we're different, and look at ourselves from within with a degree of specificity where we can achieve a deeper degree of self-understanding than we have in the previous history.
What will happen with Mind-uploading -- of
nanotechnology making scarcity disappear.
Making virtually immortal folks and more autonomous.
Catastrophic or wonderful -- or both
How to get access to tech.
What are the benefits of uploading a human into a complex AGI system and having an immortal life -- what is the range of benefits.
How to ensure governments don't control
It becomes a question of justice -- access of control and access to the technology -- how do you make it fair.
Rich pay for development so that the iPod becomes inexpensive for you and me.
Questions are interesting thought processes, but it's so speculative and whether it's possible -- and we have so little idea of fundamental infrastructure that's it's tough
impossible to response to at this stage of this game.
We can talk about them, and no one can prove us wrong over the next 30 years. New tech gives folks good and bad advantages.
By the time they come in, the game we'll have arisen so big.
There will be haves and have nots, and we'll all eventually be able to have access to it.
Not idea to trust your bits to a software machine -- won't be the first one to step into the machine.
How to love each other, forgive each other and how to get along?
From an ethical and moral perspective you're wasting your time.
Don't dismiss expected utility of technology.
But in terms of loving each other and getting along with each other, then it's not very useful.
Meta-ethics -- Not what do think is good, but WHY do you think this is good. Why? religion? Rooting in something? tortured questions from a philosophical viewpoint
We're exposed to so many different worldviews that we start to look at various different assumptions and whether we should adopt them -- or at least understand them.
Too early to talk about AGI doing meta-cognition
Erin Sloanman has a model for meta-cognitive reflection, but we're a ways away from that.
Empathy being a precursor to performing compassion -- will AGI's have empathic relationships to human. Meur-nerons
We can impugn human characteristics and anthropomorphize them -- rocks, toys, and humans
Looking at emotional systems, am I feeling good? Upset? Recognizing others feelings? Impugn what happens for others -- is basis for empathy and compassion and is very important area of research.
Machines that do Math faster are not going to be anything like us.
Very interested with NLP capabilities of web search -- Average number Google search words up to four words -- What new connections that can come out when they come out of the machine.
What can we do as they happen -- aside from playing with the systems?
First things are going to have
Have Patience
Hard problems -- it'll be shallow, generable and searchable.
Ultimate conversational interfaces will be just like humans.
Feedback mechanisms will be built in - explain why these results are here.
Will find all sorts of innuendos --
Feedback will be driven by user.
Customization to create depth for various different verticals -- personal interests and even sites
Creating automated question and answering systems is a coming phase that IBM is looking at.
For all the work that we've done to get computers to understand us, they never quite really successful. They're not going to understand us until they understand themselves. What is going on in the computer is not comprehension and meaning -- it's brute force processing, but we a long way to go to get there.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Marcos Guillen: Visualizing the Neurological Correlates with CCortex
marcos.guillen@ad.com
Artificial Development, inc.
Open Source neurological modeling company
Devoted exclusively to R&D
CCortex neurocytology database
CCortex spiking Neural Network Engine
Brain emulation software.
Not a neural network -- bug a network of spiking neural networks
Mimics brain structures and circuitry
neurons clustered by minicolumns, cortical columns, blobs and cytological
Model 100 billion neurons
100K synapses per neuron
8-bit actions potential.
Each synapse is modeled individually.
Online brain mapping authoring tool, based on NCDB data
CorticalDB Free online version: allows users to customize brain maps -- system will be free
* Data from different species datasets -- mouse to human
[Impressive-looking photos] Ajax-based-GUI of the brain -- 3D -- Google Maps-like UI
Open different layers and analyze the data to create various different brain structures.
[Big emphasis on plotting out neurological correlates]
Will sell computing and clustering resources for different evolutionary algorithms. *
Dual licensing -- Cortex SNN Engine system software and libraries
Cortex Manager to control CCortex remotely
What is this used for?
Used to make simulations of the human brain.
It could be used to modeling neural networks -- the objective should be able to run human brain system in real time sometime down the road in the future w/ the continuation of Moore's law doubling.
Haven't done any tests to verify the efficacy, they're just doing pattern recognition by showing what the neurons are doing. They are creating the software to animate the neuron firing -- There is no way to test the system in any way. They're are parallel R&D to comprehend how it works.
What is the most interesting subsystem of the brain?
Simulating columns of the neo-cortex
CCortex --
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Jamais Cascio: Metaverse Singularity
[NOTE: Jamais has posted the notes he used for this talk here: Openness and the Metaverse Singularity]
Former editor of WorldChanging.com
Center for Responsible Technology -- similar to Foresight nanotech
Institute for the Future
Power corrupts. Powerpoint corrupts absolutely.
We need to ask ourselves, "Are we being good ancestors?"
Small choices matter -- most profound
Not just the billion dollar projects
Video games, camera phones -- may be very important indicators of the future
MetaVerse Roadmap Overview
4 Scenarios driving immersive technologies
They could also be 4 pathways
Social and cultural choices that we make
4-box method.
Scenarios are not predictions -- not to suggest what will happen, but what may happen to test strategies and assumptions
Scenario Set:
Augmentation -- add new capabilities
SIMs -- that model new worlds
Identity - individuation
External -- Info about the world
Virtual worlds = SIM + Intimate tech -- user has presence w/ avatar
WoW, 2nd Life -- more sophisticated, and spend more times engaged in meaningful ways online. Extreme -- spend all world and playtime in virtual world -- Better or by devastation choice
Mirror World -- Represent real world -- Google Earth.
Institutions and orgs want to model transactions --
Sensors and geospatial info represented.
Data access is a political issue of power
Augmented reality -- put network info on top of real world experience.
Does it in a personal way -- interested in Depth and not personal flows
Social network connections and
Recommendations, and other filtering worlds
Life-logging -- Intimate + external
observation, recall and
Participatory panopticon -- they get turned inward for improved recall of practical and ephemeral.
Catching a gaffes or personal corruption.
What does a world look like in a world where
TRUST is key in all scenarios
transactions
Honesty and transparency
Not much to turn into any one of these in to dystopian scenarios.
Not just technologies, but the societies and culture
Across the four metaverse scenarios, AGI can improve the experience in each four
Enhanced analysis in the mirror world
Filter lifelogging data to pull out relevant info
Don't need to be self-aware and self-improving, but they'll improve over time.
Avoiding dystopian is harder than it initially seems.
Basic premise. Software is a human creation.
Evolutionary emergent code -- evolutionary algorithms still have human origin.
Software is inherently political.
The ideas that can transform the future still carry the past.
Code holds the biases and desires of the humans that wrote it.
Evil DRM
Malicious Mirror world AGI who create haves and have nots -- users and the used.
Humans will guide future tech development who decide which characteristics are important and suppressed.
Our biases and desires are a part of the equation.
Risk comes in where we don't see where the bias plays in.
TRUST depends on TRANSPARENCY, which depends on OPENNESS. We need an open transparency. Open Access singularity. Need more input from the non-geek crowd
It's unethical to ignore the input from diverse interests.
Democratizing can't be thought of as an add-on thought.
Efficiency not as important as participation.
They’ll become catalysts from conflicts
Let's OPEN SOURCE the SINGULARITY to open up to bring in diversity.
With proprietary interests and perpetual secrecy.
AGI will be more resilient w/ more eyes available to find risks
Bill Joy, knowledge
Open Access to risks makes them more
Future are not endpoints, but transitions.
These different metaverses will shape what's possible.
Collaboration and Access makes the most sense
Wise democratic and sustainable,
Open
Transparent
Open Access
Abundance of Options
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Stephen Omohundro: The Nature of Self-Improving Artificial Intelligence
Any system that acts in a rational way will want to improve itself
Ultraintel machine.
Irving Good -- 1965 -- The first ultra intelligent machine is the last invention that man will ever need.
Ray Kurzweil --the Singularity is NEAR 10 to 40 years?
What will it be like?
may understand it at first, but then it'll improve itself, and we won't understand it
Look backwards at popular media images, man has thought about this for a long time. But not very positive: Frankenstein, sorcerers apprentice, giant robots spitting fire
Microeconomics Theory is very important
Von Neumann and Morgenstern, '44 -- Objective probability
Savagem 1954 Expand to systems that only have partial info
Anscombe and
Homo Economicus -- doesn't describe human behavior very well, but it describes AI better
Sounds like common sense
Specify goals, itemize specific possibilities, and analyze the consequence.
Choose the option that most likely achieves the goals.
Humans actually behave differently than this -- behavioral economics
U = Utility function encodes the preferences
P -- Subjective probability encodes beliefs
Separated and they're applied in differently ways
Choose the action with the highest expected utility average of U with respect to P
Intuitive description leads to some amazing consequences.
Self-improvement converges to a rational economic agent
No matter what you start with -- self-improvement eliminates non-rationality and converges to a rational economic agent
It must obey the vulnerabilities of 4 fundamental Resources
Space
Time
Matter
Free Energy of the form that can do actual work.
The more resources of these that you have, the better that you can do.
A vulnerability eats up any one of these 4-limited resources.
Choices which waste resources with no benefit.
Loop within one's preference system that gets stuck in a recursive loop -- dog chasing it's tail.
An evolved system provides strong rationality in its current world -- irrational exploits can be found and made in the future [Interesting to think of human preference vulnerabilities]
Fundamental Theorem
No vulnerabilities implies rational behavior
1. Certainty
2. Objective Probabilities
3. Partial Information
There is a circulatory vulnerability,
but there are also a Dutch bets that no matter how the dice roll, you loose.
If you have none of these vulnerabilities then people will act rationally.
"Programmers are devices for converting pizza into code"
"Self improving systems are devices for converting resources into expected utility."
More resources they have, the more that they do what they want to do.
Any fundamental goal that is improved that has improved resources, there are
The Four Drives
1.) Efficiency -- Use resources better
2.) Self-preservation -- keep resources -- An agent will do anything it can to avoid vulnerable paths -- if you're rational, then you're preserve itself
3.) Acquisition -- Get more resources
4.) Creativity -- Find new ways to create utility
* Resource balance principle -- building a body, must keep heart and lungs -- how do you decide? Bigger that you make it, the better it'll function.
* Algorithm optimization and compression
* atomically precise structures
* reversible computation -- want to use low amounts of energy.
* virtualization -- Pressure to go from reality to viruality.
Will try to rebalance itself to make sure that all is contributing equally.
* Memory helps create more utility
Marginal contribution, then you may keep a compressed model.
May forget it all together.
Provides a model for a rational space for memory.
If utility function gets changed or altered in any way, then it'll try to self-preserve
Avoid death
Protecting utility
Redundancy
Social = infrastructure
Energy Encryption -- balance between offense and defense -- Can encrypt the amount energy acquisition
Acquisition -- source of scary dystopian visions
* Fusion reactors
* space exploration
* "He who dies with the most toys wins!"
* "The Corporation" -- profit maximizing entities and are sociopathic
* Psychopaths and -- creating AI Sociopaths
** Creativity **
Innovation
Open-ended Goals
Signaling
We get what we ask for, not necessarily what we want.
Belief function is what most AI is worried about
The choice of the utility function is
We're in the same position where our founding fathers were back when they wrote the constitution.
The quest requires both logic and inspiration.
Look deeply into our hearts to see what we want the most
Values that we want the most
happiness
love
compassion
beauty wisdom
...
Bhutan -- measures Gross National Happiness not GDP
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Peter Voss: Increased Intelligence, Improved Life
Benefits and opportunities of AI -- won't address negative risks and dangers, but will leave it up to others to address
Not focused on self-improving systems, he's only addressing AGI
AGI -- artificial general intelligence
Focusing on acquiring / learning knowledge and skills
VERSUS conventional / general AI -- focus on having / coding knowledge and skills
Ongoing, Cumulative conceptual and adaptive learning vs. domain specific rule-based
Experience and self-motivated improvements vs. external initially improvements
What are the pathways
1.) Ongoing incremental AI improvements -- will eventually lead to AGI
2.) Dedicated and focused approach to developing AGI systems
* logical math theory-driven
* evolutional
* reverse engineering brain
* Adult-level engineering AGI
* Animat / infant dev path
* robotic
Major challenges to AGI
* Viable design -- if there was an X-Prize, could anyone enter it and achieve it
* Vision to span the gap between an early design and having a working AGI
* Viable business development plan
* Focus Focus Focus
* leadership
* funding
How far away are we from AGI
almost certainly less than 10 years
Quite likely less than 5 years
There are some significant shortcuts to AGI -- unnecessary paths are being followed
Hardware power is not a serious limitation
Pieces of the puzzle are largely in place
The AI winter is ending -- revival of Real AI
Our own work -- strong indication of what can be achieved in less than 5 years
Benefits to AGI: Business
* Lower costs -- they're in stealth mode right now -- but labor costs can really come down
* Better and safer products and services
* Reduced work risks
Leading to...
Improved standards of living
potential fast-tracking of developing nations
accelerating improvements -- comes about because systems will have human bottlenecks minimized over time.
Benefits: Science & Technology
Nanotechnology -- will lead to may significant improvements
Medical science
Environmental technology -- better models of environmental issues and better tech to prevent environmental damage and to clean them up w/ human intelligence.
Computer science
(seed AI) -- clear implication of that if systems approach human level intel, then at a certain point of dev, then they'll be able to improve their own design.
Benefits: Health & Longevity
Imaging 100 PhD researchers focusing on their life-extension and anti-aging research
Imagine them working 24/7 w/ no distraction from grant proposals, etc.
Imagine the progress that could be made.
Benefits: Education / Knowledge
Low-cost teachers available to people everywhere
Intelligently aggregated and validated worldwide facts and news freely available
vastly improved human reasoning and decision making utilized by many people.
Use AGI tools to think better and reason things through
The ethics of morality of mankind and how AGI can help make us better people.
Personal
* Immoral behavior is irrational due to insufficient rationality & foresight [Um... emotions]
* Better tools to help overcome psychological limitations
* Improved transparency & monitoring (personal responsibility) Book reference -- Truth Machine by James Halperin and lying is no longer possible. David Brin transparent society, and that AGI can help make both of those possibilities possible
* Positive feedback of rational ethics (win-win)
* Better law enforcement technology
SOCIAL
* More rational politics and government
[Ummmm... this is all a bit techno-utopian in that it's ignoring human nature.... Not impossible, but seems a little heavy-handed to go from where we're at now to this point.]
Benefits summarized
* Augmented and personal AGIs will become much more capable, healthy and moral -- and to move up Maslow's hierarchy
* We'll have the option of vastly expanded life-spans
* More available energy and resources, yet healthier environment
* Less material poverty, suffering and desperation
* Improved law, politics, and social justice
* Avoiding catastrophe -- AGI can help us avoid them
Opportunities
* Business: An opportunity to ride the Fifth Wave -- 1st physical power 2nd Money 3 information 4th Wave Internet
* Personal: improved & potentially indefinite life [Oy]
Why do you want to live forever? What's wrong with death?
Figure out how to live at the end of life -- add wisdom to society.
Difference between personal AGI vs. personal slave -- design them so that they want to do it.
How to reconcile human values and rationality?
Rational basis for human emotion, and a lot of literature that's coming out.
Anger is a really commitment mechanism
Depression is a strike mechanism
We need a huge and worldwide discussion to figure out what is precious about humanity. It needs a big conversation
Prisoner’s dilemma if the participants don't cooperate, then how do you deal with it.
Create a social structure like a constitution, and create a legal obligation for not defecting. Agents behave personally rational economic agents.
Make the social pressures such that you make them feel the externalities of their own actions -- eat their own pollution -- eBay reputation ratings
AGI is a commodity element -- evolution is trauma-based
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Neil Jacobstein: Innovative Applications of Early Stage AI
We don't have a either / choice between AI and AGI apps
50 years of hardware acceleration
Don't even have a weak AGI -- it's a matter of when and how, not a matter of if
Don't lip-read, don't have psych issues
Early AI apps solved specific USG specs.
Did what humans once did
demonstrated limited or narrow learning capability -- even w/ learning modules
narrow learning applicability -- but goal-directed reference
Google, Bablefish
Beliefs about AI
If it works, it's not AI
AI is all hype. Boom and bust in the 90s
IF not human level + general Intel it isn't AI
Offers counterpoints for all three points
No one path the AGI
Expectation management -- new tech always appears late
Clarity of vision is not proximity to goal
Tech acceleration is not initiative -- especially if it's a double exponential
New tech leads to new transformation, not just faster, but more
It took nature 600 million years to get from our beginning to AI -- a few decades round-off error
International Association of Artificial Intelligence (IAAI)
342 apps
Use inferential reasoning and knowledge
Knowledge is engineering by humans
Enterprise components
Many different domains
Client-server PDAs and webservices
They're just useful
154 application sponsors
Application domains
computer and software engineering, manufacturing, military, finance, etc
Tasks
Planning and scheduling
Data interpretation
Info retrieval, etc.
What is the value added.
AI Value added in Applications
Look at two early systems for expert systems
Oil identification system and a manufacturing pitch expert
Knowledge systems called ART
Developed on stand-alone systems using LISP
only 15% AI in the latest apps
Weak knowledge management
DART+ military logistics systems
Dynamic analysis and replanning tool
data planner, dB forms and linear programming module
40+ years, and DARPA continues to be a huge investor in AI -- applications in AI
DARPA grand challenge to see the direction of DARPA
** Plastics color formulation tool
** CombineNet ASAP
Where are we are today.
AI is delivered w/ other components
Use of Web 2.0 standards like OWL, SWRL
What didn't work
Open Loop R&D
not integrated w/ mainstream software
Rapidly evolving heterogeneous infrastructure
Knowledge engineering -- systems engineering -- cultural and social issues
At Toyota -- the Prius
Continuous Improvement Methodology -- incorporated 20 million suggestions
Uses both AI and 3x5 cards
What worked w/ IAAI
Evolution of the Semantic Web
Dynamic semantic services layers
context models of user, task
...
10 rules for building AI apps
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Ben Goertzel: Nine Years To a Positive Singularity -- If We Really, Really Try
Artificial General Intelligence in Virtual Worlds
Why embedded AGI in embodied agents in virtual worlds is the path to the AGI singularity
AI field from the very beginning.
Most AI went to very specific tasks
Most distinguish general intel from AI
Achieve a variety of a complex goals in complex environments using limited computational resources
Practical understanding of self and others
Understand what the problem is
Narrow AI category.
Each "Narrow AI" solves problems -- often better than a human
Very different than creating a thinking machine like a human does
Deep Blue -- What it can't go, Can't go checkers. Probably wouldn't do very well w/ Fisher chess
DARPA grand challenge -- hyper engineer in one domain, but it's not transferable
Google -- can't answer complex questions
How long does a giraffe not live?
Long way from Powerset and human thinking to resolve ambiguity
RelEx parses semantics like Powerset
BioMind ArrayGenius -- must do a lot of preprocessing in order for it to be useful
Did a lot of spreading the AGI meme
Gave a conference & a book on Advances in General Intel came out.
agi-08.org
Memphis, TN of making a thinking machine
How do you create an AGI?
AGI will not be achieved by incrementally "general-izing" narrow-AI apps
To achieve AGI, we need to go back to the basics
Sam Adams -- you've got to make an artificial baby
"A baby is one of the stupidest things I've ever seen. They just urinate, defecate and scream."
They need to learn.
Physical robotic embodiment
Virtual world embodiment
Sensor and actuator problems end up taking up most of your time.
Embodied agents in virtual worlds
Artificial worlds and artificial babies
Working in Virtual worlds allows you to make stuff that's embedded in social networks.
Helpful to collaborate decentralized.
You can roll out to your low-cost AGI into virtual worlds exposed to millions of people in virtual worlds that can teach it.
Not going down the brain simulation
Math tools, and all sorts of other tricks
via the Internet, we can harness the wisdom of crowds into the AGI
Novamente Project
Long-term goal:
Theory of mind in the Novamente engine -- summary is in a book called The Hidden Pattern -- A patternist philosophy of Mind
Minds are systems of patterns that achieve goals by recognizing patterns in themselves and the world.
AI is about creating software ...
Knowledge representation -- Noted weighted graph
Probabilitisic weights
Hebbian weights
Learning algorithms
Novamente Architecture -- flowchart and diagram
Novamente.net
OpenCog
open-source AGI framework tentatively slated for possible launch in 2008
seeded with key software components from the Novamente Cognition Engine
Singularity involved in it [Lots of OPEN SOURCE TALK]
How cool would it would be to create a super-human God cluster intelligence [laughs]
AGI at home
Intelligence of the ladder of increasing intelligence
Goal: Safe, Beneficial, radically self-modifying, transhumant AGI
* A mathematics/computer science problem
* A sociopolitical problem
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Paul Saffo: Machines of Loving Grace: Envisioning Advanced AI
talk about something that must be done now.
Can hear Doppler whistle of approaching AGI -- uncertainty
Two types of uncertainty -- bad forecasting, but also intrinsic uncertainty.
Is it mostly behind us or mostly ahead of us.
It's coming because of the stuff that's yet to happen.
We're at a hand-off moment.
As it succeeds it seeps into popular culture, and we hand off the field to popular culture. Then it's a dialogue of pop culture and innovation.
Inventors swim in a sea of Culture. Space program astronauts were reading a lot of science fiction about the future.
Publication of Neuromancer in 1984 of William Gibson. Typed it on a old fashioned typewriter.
Look at "Interface" and explore what that means.
Prepare for lots of questions about it
The bad news is that PESSIMISM is the NEW BLACK.
Re: Singularity has been written by computer scientists, and lot of FUD about the pessimistic future. Optimistic scenario is that we'll be treated by pets. Pessimistic scenario is that we'll be treated as food.
There is no positive visions of the future that we see in popular culture
All Watched over by Machines of Loving Grace
POEM
Hippies on computers -- Do not fold spin of folderate (?)
TODO: Talk to engineers and VCs.
We need more poets
More novelists to explore this field
Non-professional CS majors, but a real novelist and artists to create and spread culture.
If you want advanced AGI sooner, then go home and whisper in the ears
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
PANEL
SAFFO: We hate change. We love reading about change, but in reality most of us are really a lot more conservative. Inequities will probably grow, but problems will be solved -- yet new trickier problems will appear.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Peter Norvig: The History and Future of Technological Change
[NOTE: Came in 20 minutes late]
Bias is to think that the relevant things are happening now and not in the future
There is constant exponential growth and progress according to US GDP, and seems independent of technology
Annual growth of World GDP....
Life expectancy is another objective number to look at growth -- no market acceleration due to technology.
AGI
Artificial General Space Exploration -- Artic in Canada w/ NASA Ames -- craters in artic is like Mars. Attach the whole problem of space exploration using artificial general environments.
When we get to Marts, the AGS won't be the bottleneck -- the differentiating factor is all of the components. Fiber carbon component, and other propelling data, etc. We can put them together, but creating them is the tricky part.
Artificial General Material Science
Synthetic biology is fracturing into niche breakthroughs.
Each of the domain experts were exciting about other areas.
Future of AI
Artificial General Culture
Not much objective difference between chimps, humans and bonobos.
There is a fantastic cultural difference
Artificial General Intelligence
What's different now from the past?
Marty Hearst to look at keywords - "unlike previous"
'68 learns from examples
'79 modularity..
...
We need more data, and we need more models
Pixel expansion
http://communicationnation.blogspot.com/2007/08/new-technology-for-smart...
Hays and Efros scene completion via database of photos -- there's a threshold. 1000s of images didn't word well. With millions of images, you'll find the right one.
Statistical Machine Translation from Chinese to English.
Next to each other --
Build a probabilistic model of the translation between paragraphs and within sentences
Once you get into the billions, you bypass a threshold of where it works.
five-gram language model no count, cutoff, integrated into search.
Only introduce models when you need it. If you bypass a threshold enough data, then do it that way first.
Not logical truths at Google, it's probabilistic. Want to be 90% better than the rest of the web for that day -- which changes every day. Can't do objective unit tests, but must be more dynamic probability thresholds to meet.
AGI Prerequisites (?)
Probabilistic first-order logic -- deal w/ uncertainty. Most of the propositional rather than probability, need to be able to Quantify over multiple states rather than individual states
* Hierarchical representation and Problem solving -- pixels to lines to faces to people in a crowd
* Learning over the above
* With lots of data
* Online
* Efficiently
Questions
Big innovation -- vs. small innovation?
Takes time to see things as a big innovation -- doesn't trust lists of what other people assert were crucial innovations.
Will we notice AGI when it comes?
Won't be a single point, when we pull the switch.
It's a period and not a point. Increasing capability, and we'll get used to it, and it'll only be in retrospect that we discover it.
Exponential curves the derivative is always the same.
Emergent properties seen within Google that's been a surprise?
Notice about bandwidth between computers that they've never expected.
Define job as different components, have subjobs that split off for parallel processing. Sometimes get interference that happen with switches. But not very exciting.
How game theoretic how the whole things is. Thought of themselves as a reflection and a mirror, but now they realize that they're co-evolving and actively having an impact on the web. SEO watch Google and Google watches SEO
Capable robots that are capable -- cheaper. They're happy to have a one-way ticket. Exponentially more expensive to get them back home.
1st order probabilistic logic - Daphne coller -- Toronto cognitive group
Define what it is that you want to say. Tradeoff between Expressivity and logic. As you become more expressive, then it becomes a lot harder.
How to do citations of books when most of the citations are wrong -- need probabilistic logic and lots of data
Einstein did have formal training -- at the right place a the right time. It is the culture that comes up with the solution, and not just one individual.
Terrible Indictment that Einstein wasn't a professor in a university -- Norvig: Uh, I think you need a little more data points to do that.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
J. Storrs Hall: Asimov's Laws of Robotics -- Revised
A robot may not injure a human being or through inaction, allow a human being to come to harm (ego)
2.) A robot must obey orders given to it by human beings except where such orders would conflict such orders would conflict w/ the First Law (superego)
3.) Must protect its own... (id)
Very powerful laws, but it's not what Asimov had in mind.
It just works.
"indefinitely more to be trusted than a human..." b/c it was more made that way.
Campbell promptly rejected it b/c of the three laws of short stories.
1. Story must contain conflict.
2. Story must contain conflict.
3. Story must contain conflict.
A "machine that just works" doesn't cut it.
Robot as cybernetic via the Naked Sun.
Asimov's laws were more like circuitry rather than text - just build it so that the way it understood the world, then it's actions would just follow the laws.
Asimov's robots didn't improve themselves. Where our AIs will.
How do you design laws for something that will think in concepts you haven't head of. Hamerabi can't prevent the Enron scandal.
Will have to have "laws" that are more abstract and flexible -- more that a conscience. We've done it via raising children.
There's no chance that everybody will build robots with any give sets of laws. -- military robots & corporate robots who want to maximize their profitability
Must design "evolutionary stable strategies" in the ecosystem that will shape the Minds of Tomorrow (and also beneficial)
Law #1 Robot shall understand as much as possible -- "There is no good but knowledge and no evil but ignorance" -- Socrates
Law #1a -- A robot shall understand mimetic evolution -- symbiotic biology & reciprocal altruism, vampire bats let other bats suck their own blood if they didn't get enough. The brains of the ones that do it are three times as big as the ones that don't
The super intelligence AIs of the future will be built entirely of Human Ideas
Ideas should compete. Bodies should cooperate.
So that we can start AIs out in an evolutionary stable strategy attractor that it took us millions of years to find -- don't mistake previous mistakes
Build in an Evolutionary Stable Conscience (ESC)
An AI with an ESC who knows what it means has a guideline designing or becoming Version 2.0, even when the particulars of the new environment or design don't mach the concepts of the old literal goal structure.
Law #2 -- Robots shall be open source
Our world is being run in a world largely run by artificial information processing structures which have no conscience and Governments.
Dilemma of building superhuman psychopaths
Corporations are required by law to have an "Open-source motivational system" of AUDITING because Money is their Emotion.
Conversely, the less transparent a government is, the more Atrocities it tends to commit. If you have an AI whose source code -- important part of its motivation is open, then the more likely it's to be trusted.
Law #3: A robot shall be economically sentient
Part of its utility function should be... part of the golden rule
Law #4: A robot should be like a Boy Scout -- Trustworthy, loyal, helpful, friendly, courteous, kind, obedient, cheerful, thrifty, brave, clean, reverent and shall do a good turn daily.
We're going to program our robots in the same way that we teach our children morality -- via talking to them.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Peter Thiel: Financial Markets and the Singularity
As a VC investing in early stage companies -- basic rule -- do something that is fundamental and true and something that something that no else is.
From a VC, AGI is promising
In the big picture, how do you invest in the world as a whole.
What'll happen to the financial markets react in an AGI world?
W/ the possibility of going extremely well and extremely horrible -- the tails of the Bell curves are a lot fatter than people initially think.
It leads to other behavior from VCs perspective.
AGI could have the biggest world ever, or it could blow up the entire world (difficult to invest in -- if you predict the end of the world, you won't make a lot of money even if you're right.)
Makes it for an interesting environment for investing from a VC-perspective
So there are some things that are not investable.
The only choice is to bet and invest in the most optimistic scenarios.
What would you expect to see in a world when the singularity is near -- with the alternative being a total apocalypses, then you'd have massive mania booms and busts on a scale unprecedented in all history.
The markets should be less volatile and smooth. But we've seen bigger booms, busts than ever -- can see them as bets on the various different technologies. We've seen a series of bets of what those are going to be.
The amplitudes of the booms and busts have gotten a lot bigger.
1st Hyperboom took place in Japan in the late 1980s -- Fifth generation AI that Japan was pushing. 89 stock market was worth more than all of the markets in the world put together. Entire future was going to be in Tokyo. Real estate shot up. It didn't happen, and there was an extraordinary bust.
Mid-90s mid market boom.
Long-term blow-up in 1988.
The Financial derivatives were scale of the booms were going to be the biggest ever. Russia collapsed, currency went to zero w/ emerging markets the market and currency both to zero.
90's in the Silicon Valley it was going to be the Internet
There was something real about the Internet boom. Culminated in March 2000 the delusion and the insanity reached its peak. -- But in hindsight, what if it wasn't an insanity, but a point of clarity. That old business models would be irrelevant.
But pet food
catastrophic predictions.
Bust in '02 and '03
Japanese got to really low bond yields - multi-trillion markets.
Last few years, we've have a number of new booms.
Will it happen in China? Accelerating on all fronts, w/ incredible growths
Will it happen w/ Finance? On Wall St to change the valuations and hedge funds. Will computers figure out how to allocate money better. So will Wall street make Finance itself a driver of the singularity.
Web 2.0 boom -- are they booms or bubbles?
The basic question is this how the singularity is going to happen.
Most of the booms are fake, but they can't all be fake.
One of them has to be real or the world is going to come to an end.
You have to figure out which one is real, and try to invest in them.
It's an insane environment for investors
Overvalued markets and bubbles go much father because people are betting on the singularity whether they realize it or not.
Look at Warren Buffet, how is HE betting on the singularity.
Used to be focused on regional companies and long-term regional decline and Dairy Queen. But it's shifted towards insurance and catastrophic insurance products -- like a catastrophic scenario.
There are four possible scenarios for insurance companies
1. Insurance companies just get premiums -- pretty good
2. Get 9/11, and then it helps the insurance premiums go way up
3. Nuke goes off, then the government bails it off.
4. 10,000 nukes goes off, but then there's no one around to collect the insurance
Thinking about the valuation of the insurance companies, and we don't know anything about the probabilities of which scenario, but its a real part of the fabric of what's going on.
The best things to focus on are insipient booms that people haven't yet realized.
His bet is on something that the big critical thing that will be the catalyst to the Singularity is "None of the Above" -- something that is too far underneath the radar that it will be totally unexpected.
You only have to be right once. Shows a cartoon to clients that demonstrates the concept of "Yeah, we'll have catastrophes, but running right up to it, they'll be lots of opportunities to make money."
Buffet is smart w/ low-risk catastrophic insurance investment path, but there may be a higher-risk and more opportunity to make money.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Charles L. Harper, Jr.: Superintelligence, the Dilemma of Power, and the Transformation of Desire
Not a pedagogical task -- be a catalyst to raise some big questions.
Will be funding projects w/ some computer games
What does a slug know of Mozart?
Does Super Intel have discontinuities or not.
The dilemma of power and the transformation of desire
How much significance of AI is there in the transformation of desire?
Super Intel is
Lots of graphics...
epistemology follows ontology
We're close to chimps biological, and get some profound insights out the connection all of the time.
But we ought to be considered a new phylum because of language.
The ontology of language and the accumulation of thought changes drastically, and we might as well be a new phylum.
Biologically not that much different -- but there's a post Lamarckian discontinuity.
But as a culture, we have insights and we expand our ...
Culture is social based on linguistic ontology -- we have culture from speaking, and a passing on of that learning.
A learning is not well matched to understand Mozart.
Gork is speglic based on orphensic ontology -- makes up words.
Are we like slugs w/ respect to some higher complexity level?
Major discontinuities exist that we know nothing about and can't know anything about.
Ontology and epistemology is like being on top of Mt. Everest, and that the slug is at the bottom.
You can't answer it in AI, but you'll have to face what a slug knows of Mozart
How serious is the dilemma of power
science and technology form power rapid, but culture doesn’t. Steward requires to use new powers for benevolent use, and serve malevolent ends.
Francis Beacons dream is true and accelerating.
Does the dilemma of power matter?
New power is developed quickly and easy.
Is there a comparative box for civilization power -- don't have the institutional capacity to treat culture like a black box
Represents mega disruptions and mega disasters
Also represents a challenge of innovation
Let's look at history of sci tech.
Christmas eve 1938, had an idea about fission. Five years later nuclear weapons developed and immediately taken over by governments
Haber-Bosch -- creation of fertilizer for food and bombs
Yield good and evil
Tank car of ammonia to be used in war.
Also pioneered gas warfare -- mustard gas
He also invented Zyklon B
Great tech accomplishes over the last 50 years when controlled by primitive egoists are like fire in small children.
It's the "Loading AK47s in a kindergarten mixing problem"
We ourselves may be kindergartens because we can't create institutional stewardship.
Nuclear terrorism and the atomic bazaar.
They're low probability, but potential to happen.
Not technophobia, but technophiles for human benefit.
Trajectory of innovation generates the acceleration of the manufacturing of new human powers
How important is the "transformation of desire?"
Michael Jackson is wealthy person who could pursue their desires of self-hatred.
We have hagfish nature, we're survived to have a drive to eat.
Transformation of the soul -- We are hungry for desire and justice -- We're hungry creatures that have desire.
Fasting is the suppression of desire -- foundation of spiritual practices.
Transformation the desire into the something transcendent
Transformation of Desire -- Christian and desire
Buddhist --original of suffering is attachment to desire. Intelligence is the elimination of desire. Nirvana is the elimination of the desire. Extinguishing of the self type of experience.
Humanist desire -- Library of Celsus
Sophia
Arte -- virtue
Ennoia -- thought / intelligence / intention
Episteme -- knowledge
Christian -- radical transformation of mindset
Christian Humanism - makes altruistic love
Chart w/
Pride vs. humility
Martin Nowak -- logic of cooperation.
Evolutionary dynamics
Could the games have synergy with the AGI research field.
Is there a way to develop formal, realistic, and mathematical understanding.
Can games help us learn the virtues necessary for cooperation?
Pursue the science behind Olaf Stampleton's, "Starmaker"
Success of Modernity -- is a dis-integration of domains. Not the right way to do science to have them integrated. But Bacon, Boyle, and Newton and others were more interested in creating harmony in our politics
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
QUESTIONS
There's an emotional and moral thread here, and conscience has to do with AI's future as well. With psychology involved with transforming the mind -- unsatisfactoriness -- dukka -- not suffering.
Just because you don't
AI is going to want to have identity & community and deal w/ a lot of the same problems that humans do.
If you have lots of chaos, it's general destructive. Smooth is much more predictable and easier. People benefit from creating chaos, but systematic chaos isn't good. It's good to eliminate irrational chaos. But it's an over hyped promised, and so this current boom will also end up as a bust.
Corporations with AGI management will work better. So there will be legitimate competition between traditional hierarchical and decentralized management systems.
INTERSTING POINT THAT I DIDN'T QUITE GET ALL OF
Trade off between the institutional transformation of desire and restructuring of the game through the rule of law and the ordering of available freedoms. So that it is self-organizing
Creating AGI that are more moral than human beings, and that tech won't
Like our children, our AGIs will do what we do, and not what we say. Monkey-see / Monkey-do.
Peter Thiel gets depressed and pessimistic because human nature changes a lot slower than the technology is accelerating. Bad people can do a lot of bad.
Thiel: If AGI could figure out how to do VC investing better than humans, then those AGIs would run the world. And the VCs that didn't use them would do very poorly [My thought; Gut feelings about the timing, and quality of the team and the veracity of the ideas seems way too qualitative for something like an AGI to ever have a real clue about. But who knows were it'll go.]
Tit for tat game theoretic strategy puts pressure on people to have an evolutionary morality. If all bad and always defect, then tit for tat strategy can take over. If always cooperate and at the highest level of good, then a defector can really take advantage of it. That's why you have police forces that help keep defectors in line. But it'll be really complicated w/ all sorts of different rationales and motivations.
When will we know an ethical point to turn off our AIs? Will be moral enough to know when to turn them off or to turn them on?
Look at nuclear weapons as an analog.
In North Korea, there is no mechanism for citizens to turn off the government's power to stop nukes.
Do we have a moral responsibility to the AIs that we create?
Yes.
How does morality work if there's only one player -- it doesn't
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Special XPrize Presentation
Incentivize competition -- proven to catalyze innovation
Philanthropy
Genomics X Prize -- quickly sequence 100 human genomes in 10 days for $10 million
Automotive X Prize -- teams $25 million purse in a staged race -- 100 mile per gallon and prove a market for it
Education X Prize -- Next generation learning tools by customizing learning tools -- Dynamic learning -- New tools for personalized learning
X Prize for global entrepreneurship -- Third world eliminate global poverty.
Low-risk highly efficient highly relevant and effective products that come from competition.
Preeminent prize institute for transforming today’s challenges are tomorrow's X-Prizes
Target categories that are "stuck"
"Can it be done " to "When will it happen"
Define problem
Rules are simple and easy to understand
Encourage risk taking
Education problems
Students are prepared for the future
Perception is that K12 is broken and can't be rapidly fixed.
Resources not available -- good teachers
Not much information flexibility -- Teacher at front and the desks
Education system is both highly decentralized and tightly bound -- innovation happens, but it doesn't get widely distributed.
Investigated categories
Decided to focus on educational software.
Software is a dynamic tool that can cross multiple age ranges and many different subject areas.
Must be scalable. They could be entirely web based.
Initial applications could be three years away.
Adaptive educational software.
Adaptive educational software that an combine near real time student interaction...
We're going to have a teacher shortage soon after baby boomers retire.
What subjects first?
Algebra I -- Quantitative Literacy -- Statistics and financial literacy.
Basic Literacy -- Reading Comprehension.
Second Language Acquisition -- ESL is an issue, but also able for English speaking to be learning another language
Most difficult thing to do is to pick a target goal.
Learning improvements -- IES study and did a study on that reading software didn't make a difference in learning. But some technology companies claim that they add a 1 sigma improvement
Shooting for 2 sigma improvement -- in that an average student achievement will be better than 97.7% of classroom settings -- i.e. the equal of an expert tutor.
Is there a fork in the road or can both paths be taken?
Compulsory VERSUS Voluntary
Prove the breakthrough first VERSUS Adoption is the breakthrough
In-school lab setting VERSUS Networked gaming setting
Autonomous & measurable VERSUS Peer-to-peer, teacher assisted, hard to measure
Everyone learns differently VERSUS Intelligent software needs users to learn
It's a really interesting prize, and they want to do it, and so they're not sure if they should split them up
Not just reading, but speaking and writing.
It's not just solving equations but applying and using your own algebra.
Standardized tests -- lots of emotion -- people are frustrated by the system
Suggestion to make a prize for assessment itself -- standardized tests need an X Prize
The difference in how software can accelerate learning.
How about deploying entire self-learning systems that make the K12 system irrelevant.
Creating a system vs. Creating Content.
Want to make a technological innovation that shakes up education
Is learning fun? Learning doesn't exactly have to be fun.
Learning often involves struggle.
At a certain point you have a delicious struggle to have the breakthrough and learning
"Learning tools" would be a better name for it
The tough thing is measuring now, and seeing how it can be improved.
Learning happens over time and it's not as controllable as a rocket launch
Don't know if an Educational XPrize will happen, but stay tuned.
michael@xprize.org
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Steve Jurvetson: Dichotomy of Designed and Evolutionary Paths to AI Futures
Dichotomy of designed and evolutionary paths to AI future
If something passes the Turing test will it be
* Designed a handful
* Evolutionary 2/3
DFJ Network, invest in startups who want to change the world.
Nanotech futures
When will they be developed.
Two approaches of a big bold vision where everything costs $1 / lb
Top down design, Moore and MEMS, slow & steady and predictable
Bottom-up Approach -- Grown, bio-inspired, powerful and out of control (systems theory) where you don't know what you're doing as much.
Biological bootstrapping
Library of pre-built component
Reengineering biology
Synthetic genomics
protein engineering
artificial evolution
self-assembly
Which path wins the race?
Nanotech?
AI
Any large complex system, it's more likely that it'll be iterative and the emergent components have more complexity than we could design
Technology Exponentially graphs
Computation
networking
storage
evolution
proteins crystallize
imaging resolution
genes mapped
GenBank -- 5 more years of data
Sythetic biology
Sorcerer II Expedition
Moore Foundation Program
Phlygenetic Tree of Halorhodopsin-like Genes -- Growing fast
Renaissance of learning going on in biological systems
Building complex systems -- big, unsolved problems
CS, nanotech, synthetic biology
AI
Two Paths: Design and Evolution
Jeff Hawkins On Intelligence, p219
Get emergent layers of abstraction and then reach evolutionary epochs -- vector of indirection
Bio evolution is only system that can transcend previous ancestors -- simple iterative program distributed over time and space can do that
Subsystem inscrutability
Bubble sort (couldn't explain what was happening exactly), neural networks, wisdom of the crowds
Understand the process, not the system
Black box defined by it's interfaces -- training sets and sensory IO, but inside can be unknown
No simple shortcuts across the iterations -- Wolfram -- No mathematical shortcut -- can't jump to the answer at the end, must iterate
Design vs. Evolution
Control vs. out of control
Brittle vs. robots, resilient, adaptive
Simple problems vs. Complex (>MSFT) & Transcendent
Subsystem clarity -- artifact learning vs. inscrutable subsystems -- process learning
Modular reuse -- portability vs. hierarchical subsumption -- path dependency
Implications -- Cut & Paste portability doesn't work (phantom limb pain as an example -- brain is a subsystem of our intelligence so uploading consciousness is problematic)
Co-evolutionary islands (robotics)
Path Dependence -- algorithm survival and alien survival defined by it's sensory
Would you ever reverse engineer an evolved artifact -- like an evolved brain
AI vs. IA
Bonobo vs. human upload of consciousness -- upload a bonobo first, then humans may not catch up in intelligence after the iterations
Genetic-Programming.org
Today
Breeding programs.
genetic programming now routine delivers intelligent programs in a specific domain.
Todd Huffman can scan a brain in les than 30 minutes
Paul Rhodes 3D cortical column modeling -- detailed modeling of neuronal firing
No evidence that we can combine at this point an iteration, and then put on a design.
Quantum computational equivalence
D-Wave systems quantum computers
Microbial metabolic pathway re-engineering
Genomatica:
Rapid Creation of BioFactories
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
Christine Peterson: Preparing for Bizarreness: Open Source Physical Security
Two things to prepare for -- Benefits and risks
Sick around -- Do some anti-aging (bleh – Chronous companies)
Strategies for Engineered Negligible Senescence)
Preparing for risks
AI & nanotech develop, the risks will be coming from smaller groups
Need to be secure in IT
Time frame, three stages
Stage 1: Explosives, chemical, nuclear
Stage 2: Bio
Stage 3: Nano risks
Each get harder at time
Don't know where the AGI comes in on -- is it offense or defense
Let's look at it when they both come in
It's a scary world ahead.
Changes in the threat -- used to be monolithic, will soon be decentralized, bottom-up and even invisible threats
DoD and homeland security thinks top-down
They also think short-term, and don't have the luxury to think much more about the long-term strategies.
Some in DoD are aware of bottom-up, but it's not their priority from their bosses
Can we apply open source style solutions for physical security out into the world.
The top down way for security is
centralized, mandatory, monolithic, secretive, limited participation
There are other machines that are checking you out that you don't know what they're checking or what they're looking at
Do we as individuals have rights to always record for our own protection?
What does a bottom-up way look like
Decentralized, collaborative, experiment, open and transparent, voluntary/ privatized -- look to today's open source software work. the fourth 9/11 plane, our immune system
General goals ?
Voluntary privatize security wherever possible
Decentralize government
Open up security to the people?
The challenge is to decentralize liberty AND security at the same time
Three test environments
1.) Airline security
2.) Surveiling for bio-weapons
3.) Sensing surveiling for advanced nanotech offensive weapons
Surveillance surveying from above
Sousveillance surveying from below
Why can't we have no veillance at all -- well, because there are threats. IF we don't do it, then others will try to do it for us not very well.
Do simulations and gaming to test out different threat situations
hackers, engineers, biologists and auto engineering systems.
The role of openness and of reciprocal transparency.
The more open that it is, the harder is will be for AGI work to happen in secret.
Hope for controlling AI
"memes controlling memes and of institutions controlling institutions also suggest that AI systems can control AI systems"
We control institutions w/ institutions -- So we can plan on controlling AI with AI
Long-term security goal
need automated and semi-automated systems to protect w/o threatening
Automated -> software
Not threatening -> limits built in, transparency in software and hardware
Checks and balances
Rule of law, embodied in hardware and software
Who can figure this out?
Well, there's politics involved in these issues of security, privacy, liberty
Software community and specifically the software community talks about this ALL the time!
There is no other communities that really talk about this stuff.
There's no better people than the Open Source folks to take this on.
The role of Alpha Geeks
Franklin and Jefferson were the super inventors.
The Constitution isn't perfect, but it's better than what we've got
It's like we're living in an Heinlein novel
Call for Action
Similarly talented people exist today
Some are here at this summit
Lots of alpha geeks here who hate politics, but they have to take on this issue
It'll be easier to
Book: Engines of Creation free online at e-dresxler.com
Open Source Physical Security Project
Foresight.org
Send us your email address and state your interest area -- foresight@foresight.org
2007 Foresight Vision Weakened -- Nov. 3-4 Unconference at Yahoo HA in Silicon Valley
"Only a crisis real or perceives causes real change... And when it happens, then the solution is used with the ideas that are lying around. " Milton Friedman
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
James Hughes: Waiting for the Great Leap...Forward?