Two Approaches to AGI/AI
There are two general approaches to AGI/AI that I'd like to draw attention to, not "neat" and "scruffy", the standard division, but "brain inspired" and "not brain inspired".
Accomplishments of not brain inspired AI:
- Wolfram Alpha (in my opinion the most interesting AI today)
- spam filters
- DARPA Grand Challenge victory (Stanley)
- UAVs that fly themselves
- clever game AI
- AI that scans credit card records for fraud
- the voice recognition AI that we all talk to on the phone
- intelligence gathering AI
- Watson and derivatives
- Deep Blue
- optical character recognition (OCR)
- linguistic analysis AI
- Google Translate
- Google Search
- text mining AI
- OpenCog
- AI-based computer aided design
- the software that serves up user-specific Internet ads
- pretty much everything
Accomplishments of brain-inspired AI:
- Cortexia, a bio-inspired visual search engine
- Numenta (no product yet)
- Neural networks, which have proven highly limited
- ???? (tell me below and I'll add them)
One place where brain-inspired AI always shows up is in science fiction. In the real world, AI has very little to do with copying neurobiology, and everything to do with abstract mathematics and coming up with algorithms that work for the job, regardless of their similarity to human cognitive processing.
Response to Charles Stross’ “Three arguments against the Singularity”
super-intelligent AI is unlikely because, if you pursue Vernor's program, you get there incrementally by way of human-equivalent AI, and human-equivalent AI is unlikely. The reason it's unlikely is that human intelligence is an emergent phenomenon of human physiology, and it only survived the filtering effect of evolution by enhancing human survival fitness in some way. Enhancements to primate evolutionary fitness are not much use to a machine, or to people who want to extract useful payback (in the shape of work) from a machine they spent lots of time and effort developing. We may want machines that can recognize and respond to our motivations and needs, but we're likely to leave out the annoying bits, like needing to sleep for roughly 30% of the time, being lazy or emotionally unstable, and having motivations of its own.
"Human-equivalent AI is unlikely" is a ridiculous comment. Human level AI is extremely likely by 2060, if ever. (I'll explain why in the next post.) Stross might not understand that the term "human-equivalent AI" always means AI of human-equivalent general intelligence, never "exactly like a human being in every way".
If Stross' objections turn out to be a problem in AI development, the "workaround" is to create generally intelligent AI that doesn't depend on primate embodiment or adaptations.
Couldn't the above argument also be used to argue that Deep Blue could never play human-level chess, or that Watson could never do human-level Jeopardy?
I don't get the point of the last couple sentences. Why not just pursue general intelligence rather than "enhancements to primate evolutionary fitness", then? The concept of having "motivations of its own" seems kind of hazy. If the AI is handing me my ass in Starcraft 2, does it matter if people debate whether it has "motivations of its own"? What does "motivations of its own" even mean? Does "motivations" secretly mean "motivations of human-level complexity"?
I do have to say, this is a novel argument that Stross is forwarding. Haven't heard that one before. As far as I know, Stross must be one of the only non-religious thinkers who believes human-level AI is "unlikely", presumably indefinitely "unlikely". In a literature search I conducted in 2008 looking for academic arguments against human-level AI, I didn't find much -- mainly just Dreyfuss' What Computers Can't Do and the people who argued against Kurzweil in Are We Spiritual Machines? "Human level AI is unlikely" is one of those ideas that Romantics and non-materialists find appealing emotionally, but backing it up is another matter.
(This is all aside from the gigantic can of worms that is the ethical status of artificial intelligence; if we ascribe the value inherent in human existence to conscious intelligence, then before creating a conscious artificial intelligence we have to ask if we're creating an entity deserving of rights. Is it murder to shut down a software process that is in some sense "conscious"? Is it genocide to use genetic algorithms to evolve software agents towards consciousness? These are huge show-stoppers — it's possible that just as destructive research on human embryos is tightly regulated and restricted, we may find it socially desirable to restrict destructive research on borderline autonomous intelligences ... lest we inadvertently open the door to inhumane uses of human beings as well.)
I don't think these are "showstoppers" -- there is no government on Earth that could search every computer for lines of code that are possibly AIs. We are willing to do whatever it takes, within reason, to get a positive Singularity. Governments are not going to stop us. If one country shuts us down, we go to another country.
We clearly want machines that perform human-like tasks. We want computers that recognize our language and motivations and can take hints, rather than requiring instructions enumerated in mind-numbingly tedious detail. But whether we want them to be conscious and volitional is another question entirely. I don't want my self-driving car to argue with me about where we want to go today. I don't want my robot housekeeper to spend all its time in front of the TV watching contact sports or music videos.
All it takes is for some people to build a "volitional" AI and there you have it. Even if 99% of AIs are tools, there are organizations -- like the Singularity Institute -- working towards AIs that are more than tools.
If the subject of consciousness is not intrinsically pinned to the conscious platform, but can be arbitrarily re-targeted, then we may want AIs that focus reflexively on the needs of the humans they are assigned to — in other words, their sense of self is focussed on us, rather than internally. They perceive our needs as being their needs, with no internal sense of self to compete with our requirements. While such an AI might accidentally jeopardize its human's well-being, it's no more likely to deliberately turn on it's external "self" than you or I are to shoot ourselves in the head. And it's no more likely to try to bootstrap itself to a higher level of intelligence that has different motivational parameters than your right hand is likely to grow a motorcycle and go zooming off to explore the world around it without you.
YOU want AI to be like this. WE want AIs that do "try to bootstrap [themselves]" to a "higher level". Just because you don't want it doesn't mean that we won't build it.
Hard Takeoff Sources
Definition of "hard takeoff" (noun) from Transhumanist Wiki:
The Singularity scenario in which a mind makes the transition from prehuman or human-equivalent intelligence to strong transhumanity or superintelligence over the course of days or hours (Yudkowsky 2001). The high likelihood of a hard takeoff once a roughly human-equivalent AI is created has been argued by the Singularity Institute in Yudkowsky 2003.
Hard takeoff sources and references, which includes hard science fiction novels, academic papers, and a few short articles and interviews:
Blood Music (1985) by Greg Bear
Fire Upon the Deep (1992) by Vernor Vinge
"The Coming Technological Singularity" (1993) by Vernor Vinge
The Metamorphosis of Prime Intellect (1994) by Roger Williams
"Staring into the Singularity" (1996) by Eliezer Yudkowsky
Creating Friendly AI (2001) by Eliezer Yudkowsky
"Wiki Interview with Eliezer" (2002) by Anand
"Impact of the Singularity" (2002) by Eliezer Yudkowsky
"Levels of Organization in General Intelligence" (2002) by Eliezer Yudkowsky
"Ethical Issues in Advanced Artificial Intelligence" by Nick Bostrom
"Relative Advantages of Computer Programs, Minds-in-General, and the Human Brain" (2003) by Michael Anissimov and Anand
"Can We Avoid a Hard Takeoff?" (2005) by Vernor Vinge
"Radical Discontinuity Does Not Follow from Hard Takeoff" (2007) by Michael Anissimov
"Recursive Self-Improvement" (2008) by Eliezer Yudkowsky
"Artificial Intelligence as a Positive and Negative Factor in Global Risk" (2008) by Eliezer Yudkowsky
"The Hanson-Yudkowsky AI Foom Debate" (2008) on Less Wrong wiki
"Brain Emulation and Hard Takeoff" (2008) by Carl Shulman
"Arms Control and Intelligence Explosions" (2009) by Carl Shulman
"Hard Takeoff" (2009) on Less Wrong wiki
"When Software Goes Mental: Why Artiï¬cial Minds Mean Fast Endogenous Growth" (2009)
"Thinking About Thinkism" (2009) by Michael Anissimov
"Technological Singularity/Superintelligence/Friendly AI Concerns" (2009) by Michael Anissimov
"The Hard Takeoff Hypothesis" (2010), an abstract by Ben Goertzel
Economic Implications of Software Minds by S. Kaas, S. Rayhawk, A. Salamon and P. Salamon
Critiques
"The Age of Virtuous Machines" (2007) by J. Storrs Hall
"Thinkism" by Kevin Kelly (2008)
"The Hanson-Yudkowsky AI Foom Debate" (2008) on Less Wrong wiki
"How far can an AI jump?" by Katja Grace (2009)
"Is The City-ularity Near?" (2010) by Robin Hanson
"SIA says AI is no big threat" (2010) by Katja Grace
I don't mean to say that the critiques aren't important by putting them in a different category, I'm just doing that for easy reference. I'm sure I missed some pages or articles here, so if you have any more, please put them in the comments.
He *Might* Be Back…
From CNN:
He'll be back... if the script is right.
Arnold Schwarzenegger has given Hollywood agency CAA approval to discuss a potential new Terminator movie with studios, a source close to the star confirms to EW, but until Schwarzenegger sees a script and specifics about the project, he will not fully commit to it.
Director Justin Lin (who helmed this weekend's "Fast Five") is reportedly attached to direct the film, but it does not yet have a screenwriter.
Second Part of Yudkowsky Interview Online at John Baez’s Site
Here's the link. A small selection:
EY: I’d say that there are parts of rationality that we do understand very well in principle. Bayes’ Theorem, the expected utility formula, and Solomonoff induction between them will get you quite a long way. Bayes’ Theorem says how to update based on the evidence, Solomonoff induction tells you how to assign your priors (in principle, it should go as the Kolmogorov complexity aka algorithmic complexity of the hypothesis), and then once you have a function which predicts what will probably happen as the result of different actions, the expected utility formula says how to choose between them.
Wolfram on Alpha and Watson
Stephen Wolfram has a good blog post up describing how Alpha and Watson work and the difference between them. He also describes how Alpha is ultimately better because it is more open-ended and works based on logic rather than corpus-matching. Honestly I was more impressed by the release of Alpha than the victory of Watson, though of course both are cool.
In some ways Watson is not much more sophisticated than Google's translation approach, which is also corpus-based. I especially love the excited comments in the mainstream media that Watson represents confidence as probabilities. This is not exactly something new. In any case, Wolfram writes:
There are typically two general kinds of corporate data: structured (often numerical, and, in the future, increasingly acquired automatically) and unstructured (often textual or image-based). The IBM Jeopardy approach has to do with answering questions from unstructured textual data — with such potential applications as mining medical documents or patents, or doing ediscovery in litigation. It’s only rather recently that even search engine methods have become widely used for these kinds of tasks — and with its Jeopardy project approach IBM joins a spectrum of companies trying to go further using natural-language-processing methods.
Anders Sandberg on “AI: Predictably Unpredictable”
From UK H+ 2011. Nice to see transhumanism doing well in the UK.
Anders sports his cryonics necklace on the outside... classy.
Aaron Saenz on Artificial General Intelligence
I was just reading about the new AGI company Vicarious on Singularity Hub, and enjoyed this paragraph by Aaron Saenz:
Artificial General Intelligence is one of the Holy Grails of science because it is almost mythical in its promise: not a system that simply learns, but one that reaches and exceeds our own kind of intelligence. A truly new form of advanced life. There are many brilliant people trying to find it. Each of these AI researchers have their own approach, their own expectations, and their own history of failures and a precious few successes. The products you see on the market today are narrow AI – machines that have a very limited ability to learn. As Scott Brown said, “today’s AI technology is so primitive that much of the cleverness goes towards inventing business models that don’t require good algorithms to succeed.†We are in the infantile stages of AGI. If that. Maybe the fetal stages.
I'm not an AGI researcher, but I do hang out with them and talk AGI. Out of everyone I've seen out there so far, the way I think about AGI would be most similar to Josh Tenenbaum. A simple overview of his approach is here.
Converging Technologies Report Gives 2085 as Median Date for Human-Equivalent AI
From the NSF-backed study Converging Technologies in Society: Managing Nano-Info-Cogno-Bio Innovations (2005), on page 344:
2070
48. Scientists will be able to understand and describe human intentions,
beliefs, desires, feelings and motives in terms of well-defined computational
processes. (5.1)2085
50. The computing power and scientific knowledge will exist to build
machines that are functionally equivalent to the human brain. (5.6)
This is the median estimate from 26 participants in the study, mostly scientists.
Only 74 years away! WWII was 66 years ago, for reference. In the scheme of history, that is nothing.
Of course, the queried sample is non-representative of smart people everywhere.
Josh Tenenbaum Video Again: Bayesian Models of Human Inductive Learning
I posted this only a month ago, but here's the link to the video again. People sometimes say there's been no progress in AI, but the kind of results obtained by Tenenbaum are amazing and open up a whole approach to AI that uses fast and frugal heuristics for reasoning and requires very minimal inspiration from the human brain.
Abstract:
In everyday learning and reasoning, people routinely draw successful generalizations from very limited evidence. Even young children can infer the meanings of words, hidden properties of objects, or the existence of causal relations from just one or a few relevant observations -- far outstripping the capabilities of conventional learning machines. How do they do it? And how can we bring machines closer to these human-like learning abilities? I will argue that people's everyday inductive leaps can be understood as approximations to Bayesian computations operating over structured representations of the world, what cognitive scientists have called "intuitive theories" or "schemas". For each of several everyday learning tasks, I will consider how appropriate knowledge representations are structured and used, and how these representations could themselves be learned via Bayesian methods. The key challenge is to balance the need for strongly constrained inductive biases -- critical for generalization from very few examples -- with the flexibility to learn about the structure of new domains, to learn new inductive biases suitable for environments which we could not have been pre-programmed to perform in. The models I discuss will connect to several directions in contemporary machine learning, such as semi-supervised learning, structure learning in graphical models, hierarchical Bayesian modeling, and nonparametric Bayes.
IBM Cat Brain Nonsense in the Zeitgeist
I found another ridiculous article on IBM's so-called "cat brain" at TechWorldNews, titled "IBM Researchers Go Way Beyond AI With Cat-Like Cognitive Computing". I run into these articles all the time doing AI-related searches, so even though they were published a year ago, their deception remains strongly in effect. The fact that so many people actually believe what IBM implies shows how fundamentally confused 99% of the population (including geeks) is about AI in general. Here's a quote from the article:
IBM researchers have developed a cognitive computer simulation that mimics the way a cat brain processes thought, and they expect to be able to mimic human thought processes within a decade. "A cognitive computer could quickly and accurately put together the disparate pieces of any complex data puzzle and help people make good decisions rapidly," said Daniel Kantor, medical director of Neurologique.
Mimics the way a cat brain processes thought. They actually wrote that. So people believe in a computer that processes cat thought existing in 2009, but don't expect a computer that mimics human thought for hundreds of years or ever? People really do believe this (I probably did at one point long ago), because they were brought up on bizarre Judeo-Christian ideas which involve elevating human thought to a supernatural status which cannot be replicated in a computer. It's entirely unscientific, but even many so-called "secular humanists" believe in mystical human exceptionalism. "We're nowhere close to understanding the brain", they claim, despite thousands of detailed textbooks and hundreds of thousands of articles on the brain and mind.
It's true that we're nowhere near to understanding all the microcircuitry of the brain, but we have to distinguish between functionally relevant cognitive complexity and incidental cognitive complexity. Most of the complexity in a bird is incidental to the bird, not fundamentally necessary for flight. It may be possible to create AGI without understanding much about the human brain at all.