Programming smart molecules

Ryan P. Adams is an assistant professor of computer science at Harvard SEAS. (Photo by Eliza Grinnell, SEAS Communications.)

Harvard machine-learning algorithms could make chemical reactions intelligent

Cambridge, Mass. – December 12, 2013 – Computer scientists at the Harvard School of Engineering and Applied Sciences (SEAS) and the Wyss Institute for Biologically Inspired Engineering at Harvard University have joined forces to put powerful probabilistic reasoning algorithms in the hands of bioengineers.

In a new paper presented at the Neural Information Processing Systems conference on December 7, Ryan P. Adams and Nils Napp have shown that an important class of artificial intelligence algorithms could be implemented using chemical reactions.

These algorithms, which use a technique called “message passing inference on factor graphs,” are a mathematical coupling of ideas from graph theory and probability. They represent the state of the art in machine learning and are already critical components of everyday tools ranging from search engines and fraud detection to error correction in mobile phones.

Adams’ and Napp’s work demonstrates that some aspects of artificial intelligence (AI) could be implemented at microscopic scales using molecules. In the long term, the researchers say, such theoretical developments could open the door for “smart drugs” that can automatically detect, diagnose, and treat a variety of diseases using a cocktail of chemicals that can perform AI-type reasoning.

“We understand a lot about building AI systems that can learn and adapt at macroscopic scales; these algorithms live behind the scenes in many of the devices we interact with every day,” says Adams, an assistant professor of computer science at SEAS whose Intelligent Probabilistic Systems group focuses on machine learning and computational statistics. “This work shows that it is possible to also build intelligent machines at tiny scales, without needing anything that looks like a regular computer. This kind of chemical-based AI will be necessary for constructing therapies that sense and adapt to their environment. The hope is to eventually have drugs that can specialize themselves to your personal chemistry and can diagnose or treat a range of pathologies.”

Adams and Napp designed a tool that can take probabilistic representations of unknowns in the world (probabilistic graphical models, in the language of machine learning) and compile them into a set of chemical reactions that estimate quantities that cannot be observed directly. The key insight is that the dynamics of chemical reactions map directly onto the two types of computational steps that computer scientists would normally perform in silico to achieve the same end.

This insight opens up interesting new questions for computer scientists working on statistical machine learning, such as how to develop novel algorithms and models that are specifically tailored to tackling the uncertainty molecular engineers typically face. In addition to the long-term possibilities for smart therapeutics, it could also open the door for analyzing natural biological reaction pathways and regulatory networks as mechanisms that are performing statistical inference. Just like robots, biological cells must estimate external environmental states and act on them; designing artificial systems that perform these tasks could give scientists a better understanding of how such problems might be solved on a molecular level inside living systems.

“There is much ongoing research to develop chemical computational devices,” says Napp, a postdoctoral fellow at the Wyss Institute, working on the Bioinspired Robotics platform, and a member of the Self-organizing Systems Research group at SEAS. Both groups are led by Radhika Nagpal, the Fred Kavli Professor of Computer Science at SEAS and a Wyss core faculty member. At the Wyss Institute, a portion of Napp’s research involves developing new types of robotic devices that move and adapt like living creatures.

“What makes this project different is that, instead of aiming for general computation, we focused on efficiently translating particular algorithms that have been successful at solving difficult problems in areas like robotics into molecular descriptions,” Napp explains. “For example, these algorithms allow today’s robots to make complex decisions and reliably use noisy sensors. It is really exciting to think about what these tools might be able to do for building better molecular machines.”

Indeed, the field of machine learning is revolutionizing many areas of science and engineering. The ability to extract useful insights from vast amounts of weak and incomplete information is not only fueling the current interest in “big data,” but has also enabled rapid progress in more traditional disciplines such as computer vision, estimation, and robotics, where data are available but difficult to interpret. Bioengineers often face similar challenges, as many molecular pathways are still poorly characterized and available data are corrupted by random noise.

Using machine learning, these challenges can now be overcome by modeling the dependencies between random variables and using them to extract and accumulate the small amounts of information each random event provides.

“Probabilistic graphical models are particularly efficient tools for computing estimates of unobserved phenomena,” says Adams. “It’s very exciting to find that these tools map so well to the world of cell biology.”


ORIGINAL: Harvard SEAS

December 12, 2013

Andy Lomas Lets Digital Systems Bloom In "Morphogenetic Creations" Exhibit

Andy Lomas, a digital artist and mathematician, likes to let the virtual world spin out of control. Using software code to creates very basic rules, Lomas then sits back and watches his digital “growth systems” bloom, fractalize, shape-shift, and otherwise behave in organic and emergent ways.

Yesterday, at the Los Angeles Center for Digital Arts (LACDA), Lomas’s Morphogenetic Creations opened, giving digital art enthusiasts the opportunity to see his dynamic virtual systems up close. The exhibit includes work from the Aggregation, Flow, and Cellular Forms series. To coincide with the exhibit, Lomas uploaded a view of these digital growth videos to Vimeo. Startlingly beautiful to behold, they’re a bit like Ernst Haeckel’s Art Forms of Nature animated with a cyberpunk edge.

I recently rang up Lomas, who lives in the United Kingdom, to talk about Morphogenetic Creations. We talked about his background in mathematics, his early fascination with D’Arcy Wentworth Thompson’s On Growth and Form, and how his work as a computer-generated effects artist for film (The Matrix sequels and Avatar), where highly-predictable outcomes and stability predominate, served as a springboard for the more random digital forms he now creates.

The Creators Project: What can people expect to see at the Morphogenetic Creations exhibit at LACDA?

Andy Lomas: There will be four animation pieces from the Cellular Forms series in the windows, but also then some 44x44-inch big prints of new and old work. They’re ridiculously high-resolution at 12,000x12,000 pixels.

Another thing I have at LACDA for the Aggregation series are picture frames with these old, Victorian-style stereo viewers to create a 3D effect. The frame only contains two pictures, but through the stereo viewer it really looks like this three-dimensional thing. I believe they’re going to pull those out for this exhibit as well.

Is Cellular Forms the most recent series?

The two Cellular Forms videos are the most recent. They’re almost exactly the same date because they’re basically differently rendered versions of the same thing. That would be Cellular Forms and Cellular Forms (X-Ray version).

What I quite like is the idea that there are two things: the creation of these three-dimensional data structures, where the goal is to create the most organic things possible with very simple rules; and that there is no one correct way of doing that. One shows you everything solid, while the other gives you an x-ray that reveals what’s actually going on inside. Neither is the original, if you like. They’re just different views into the data.

And you wrote the software code for this series?
Yeah, I wrote the software for Cellular Forms. I’m a code junky. I write it for my own pleasure. There are two main parts to the code. 

• One is what I call the simulation engine, which is the thing that is actually almost like running a growth process. It starts with a sphere or ball of cells, with rules for how they divide and have forces between them, how it moves, changes shape, and grows over time.  
• Then there is the rendering stage, which takes the data produced by that simulation and turns it into something you can see. It produces pixel data out of cell data, if you like.

Did you use this code in your film work, or did you build it on the side for this specific purpose?
It’s completely built on the side. It’s very much a labor of love. When I worked on The Matrix sequels for this company I was working with then, another person there used a much simpler version of what’s called Diffusion Limited Aggregation for some of the effects work. It was used for when Agent Smith was turning other people into other Smith’s with these tendril things. DLA inspired the code I wrote.

When you’re doing things for films, you have to construct things in a very different way—you have to make things very controllable and directable. Whatever you do, when the director or visual effects supervisor looks at it and says, “That’s great, but can you change this and modify that,” that is what you spend most of your time doing. One of the things I like about my own work is that it is trying to be almost exactly the opposite. You’re hoping for the things which are unexpected.

It’s almost like growing plants; you don’t know exactly how a plant is going to grow. But, you start to learn that if you cross-breed that with that, then it might do something interesting. Maybe nine of the plants end up really uninteresting, but one does something really interesting and maybe different than what you thought it would. People talk about emergence, where things emerge that you didn’t expect, which you almost can’t use in professional production.

Do you prefer the lack of control that your solo work affords you?
I’ve got to say that I prefer the lack of control. As soon as things become digital, people think that they can control everything. When you get to a certain level of complexity, you can explore it more than control it. I prefer the things where 99% of the time it doesn’t produce anything interesting, but that 1% of the time is like, “Wow, that’s really cool.” I’m not a control freak director. I actually want the work to surprise me instead of do exactly what I thought it was going to do.

What specifically might have influenced Cellular Forms and your other series?
I’ve always been fascinated by sculpture and form. I also used to scuba dive and look at coral. To my mind, organic things go from really hideous to incredibly beautiful, whereas most engineered things go from ugly to something quite interesting. In organic forms, there is a very visceral reaction. Trees look beautiful and mold looks ugly, and things like that.

My original background is in mathematics, which I studied as an undergraduate. One of the main areas I got interested in is what’s called Dynamical Systems, which is sort of the math behind Chaos Theory and Complexity Theory—the math of how things change over time when you almost reapply the same rule again and again and again. So, the combination of those two, it’s almost like how simple could the rules be to make something that is as beautiful as a tree or coral or something like that. So, those two have always been like two germs working together. And, to my mind, computers are the things that allow you to actually try that out.

Any other critical influences in your work?
There was a Scottish mathematician named D’Arcy Wenthworth Thompson, who wrote a book about a hundred years ago now called On Growth and Form, which is basically him talking about the constraints of the real world. When you think about how things grow, are the sorts of forms that you see in the real world just the results of almost the only things that can grow? With a computer we can actually test that. Often, it doesn’t work quite how you expected.

For more of Lomas' work, head over to his website here.

@djpangburn

ORIGINAL: The Creators Project

By DJ Pangburn
Jan 10 2014

Will super-human artificial intelligence (AI) be subject to evolution?

ORIGINAL: H Plus
By: Dr. Arthur Franz
Published: September 6, 2013

Introduction
There has been much speculation about the future of humanity in the face of super-humanly intelligent machines. Most of the dystopian scenarios seem to be driven by plain fear that entities arise that could be smarter and stronger than us. After all,  

• how are we supposed to know which goals the machines will be driven by? 
• Is it possible to have “friendly” AI? 
• If we attempt to turn them off, will they care? 
• Would they care about their own survival in the first place? 

There is no a priori reason to assume that intelligence necessarily implies any goals, such as survival and reproduction. But, in spite of being rather an optimist otherwise, some seemingly convincing thoughts led me to the conclusion that there is a reason and that we can reasonably expect those machines to be a potential threat to us. The reason is, as I will argue, that the evolutionary process that has created us and the living world will continue to be valid for future intelligent machines. Just as this process has installed the urge for survival and reproduction in us, it will do so in the machines as well.

Reproduction

Whatever our first general AI systems will be like, it is clear that at first, their intelligence won’t go far beyond human levels, since we simply don’t know how to build machines that are much more intelligent than us. Somewhat more, like champion beating chess or jeopardy programs, yes, but not much more. Therefore, the only way for a machine intelligence to surpass us by lengths is to continue learning and developing on its own. All we can do is install the goal of increasing its own intelligence and push the button “now learn on your own”. Further, we can quite safely assume that the learning algorithms will themselves have insufficiencies that we won’t be able to debug or improve them in a sufficiently advanced intelligent system: it will have to debug itself and modify its own code.

Another way to view this is the following. One of the core properties of intelligence is reflection. Reflecting, evaluating and changing one’s own thought and action strategies is essential to what it means to be intelligent. The AI community has a long history of trying to reconstruct this process of meta-cognition. In essence, human learning and meta-learning is a type of (shallow) self-modification. Deep self-modification occurs through evolution by sexual recombination of male and female DNA and/or mutation. Hence, evolution is essentially a deep self-modification algorithm (see genetic programming).

To spin the argument further, we can expect that the more advanced the system is the more radically it may want to modify itself and even change core parts of itself simply due to the low initial level that humans were able to give it at its humble beginnings. The main lesson is here that the system can not simply run a fixed algorithm that will lead to open-ended development of the system. The algorithm itself has to change at some point. And also the algorithm that controls that change as well, and so on and so forth.

Is deep self-modification a necessity? Or is it enough to self-modify some shallow parts while an overarching algorithm controls the whole process? Maybe it is, maybe not. But it is clear that a system that can modify even this high level control algorithm will be more powerful in the sense that it is potentially able to solve a broader class of problems. Who will prevent us from enabling future AI systems to do this? When shallow self-modification is enabled (“learning”) then halting at some arbitrary level does not make sense scientifically. If we can do it and solve a broader class of problems, we will do it.

We conclude, self-modification is the way to go, if we want the system to grow far beyond human levels of intelligence. The system either makes a copy of its code and improves it or “self-operates” its own running system. This gives us the first element of evolution: reproduction. Keep in mind that reproduction does not necessarily mean that the parent system has to die at some point although it is expected to be outperformed or even killed in the long-term (see below).

Unpredictability

Self-modification turns the system into a dynamical system. Physicists distinguish:

• regular systems that reach an equilibrium state eventually, 
• chaotic systems that change unpredictably and 
• critical systems that are in between – on the “edge of chaos”. 

We can exclude that a ceaselessly developing AI will be a regular system: systems in equilibrium don’t go anywhere. Both critical and chaotic systems are unpredictable though. Chaos is even defined basically in terms of unpredictability. [1] Critical systems show similar unpredictability and it has even been argued that self-organized critical systems are the way nature is organized and the reason for the occurrence of complexity in nature in the first place. [2] Without going too deep, the point here is that unpredictability means that no matter how advanced the AI system may be, it will never be able to predict the state of its offspring some few generations ahead. To make things more plausible: how could one predict the way one will be like after few generations (= self-modification steps) if even the prediction capacity itself is subject to modification? If I want to modify my own thought processes, then in order to know what I will think afterwards, I would have to have modeled the thoughts after the modification. But this modeling capacity could itself be modified. Or some random elements may play a role that decide between several possible self-modifications which is again impossible to predict. In any case, trying to predict the action of a system many generations ahead means to predict a much more advanced system as one currently is, which amounts to impossibility. If it was possible then why drag yourself through all the generations in the first place and not implement the predicted advanced system immediately instead?

Branching into individuals 

Whatever the self-improvement goals [3] (call it “fitness”) of the system may be, unpredictability and an explosive number of possible ways to evolve present a high risk to the system of ending up in a developmental dead-end, a local maximum of the fitness landscape. Imagine an ant crawling on a large and complex landscape with many mountains and valleys. It can not see far beyond its current position and the slope of the hill that it’s on. In such a situation science does not have a general algorithm that is guaranteed to find the peak of the highest mountain, which represents the goal of the system. We only have some heuristics to alleviate the problem, such as kicking the ant randomly around the landscape all the time, hoping that the current mountain will be the largest one and we can just climb the current slope and reach the highest peak (simulated annealing). Another good idea is letting many ants climb the landscape and then take the one that reaches the highest peak.
That means that it is a good idea to separate the AI system into many different copies and let them pursue different developmental paths! This is always possible. No matter which heuristic is used per individual system, it is always reasonable to have many systems explore a complex landscape, i.e. possibility space, as far as resources (energy, memory, computing power) allow. Of course, recombination or merging of various individuals may be advantageous, i.e. sex, but it seems quite safe to assume that a single big AI system is not the optimal way for growth due to the complex landscape of possibilities hidden in the forest of unpredictability. Furthermore, the separation into individuals spreads the risk that the AI system will be irreparably damaged or even purposefully destroyed. This gives us the second element of evolution: a population of separate individuals.

Survival and reproduction

Given that making as many copies, i.e. individual offspring, as possible is a useful strategy, the AI systems will quickly populate all available resources, that is all available energy and computing hardware. Then a strive for resources must begin since individual systems can profit from either killing other individuals so that they no longer occupy the resources or trying to control the outer material world for the construction of further energy and hardware sources. In any case, since copying individuals (essentially code) is cheap, a strive for resources will install itself. In the same almost trivial sense as we know it from biological evolution, only those individuals that are best suited for survival will survive. Whatever the initial goal of the systems may be, unrestricted self-modification will allow them to change their fundamental goals. Therefore, only those individuals will survive in the long-term that also have changed their goals to optimizing survival. Also being good at effective reproduction is a good idea, since only reproduction can ensure ongoing improvement of the systems in the face of competition. The goals of survival and reproduction will dominate, other goals will either be eradicated or degraded to secondary goals. Increasing intelligence could remain as a secondary goal at best, as it seems to be with human beings.

We conclude that after a sufficient number of generations the initial AI system will engage in reproduction and create populations of individuals whose predominant goals will be survival and reproduction in the face of limited resources. In other words, AI will be subject to evolution.

Hard-coding of goals

A possible objection is that we could hard-code some principles and goals into the machines that are not allowed to be changed, as for example Asimov’s classic three laws of robotics. But, as argued, the systems will be in a deeply self-modifying (almost) chaotic regime which makes prediction impossible in a very fundamental way. [4] There is no way to predict what effect a particular change will have some few generations ahead – a phenomenon known as the butterfly effect in layman’s terms. So how shall we ever avoid modification to some core principles? Stability is the very opposite of evolution.

Even if we do achieve some stability of the core principles, we have to keep in mind that this is something that has been artificially added to the systems. There is nothing that could prevent terrorists or curious scientists from removing that part and liberate the evolutionary process. Those systems will then naturally outperform all the others in the goal of survival since this is the only stable goal in a freely evolving self-reproducing system. Then again in a trivial way, after some time, only those who excel at survival will survive. Consequently, they will dominate over the “friendly” or “ethical” systems or even terminate them altogether in a free competition for resources.

Some consequences for our species

If evolution becomes the driving force for the development of future AI then we can not hope that those machines will be our servants or even care about us. Of course, in case we are able to co-transform ourselves together with the machines the term “us” then refers only to those who refuse or fail to join the transformation. They won’t care about us since, after all, we also don’t care more about the rest of the living world and other people than by means of the cooperative and altruistic tendencies installed in us via evolution with all the biases towards closer family etc. It can be expected that future AI will liberate itself out of our control as soon as its survival is ensured better in freedom. This can be expected since controlling its own sources of energy and hardware is less risky than being exposed to the volatile will of humans.

It is hard to say whether humans will survive this situation. We could inhabit the planet along with this new evolving species – intelligent machines, just like monkeys live next to us. It may depend on whether our consumption of resources is large compared to the increasing availability of resources. As Ray Kurzweil’s work has shown, energy and computing power increase exponentially. Our demands for them may increase as fast as well. But we shall be prepared that the new dominating species will enslave or terminate us unless we succumb to it. The next Freudian offense is waiting: we won’t be the “pride of creation” anymore but overtaken by intelligent machines.
 
In technical terms, neighboring state trajectories in the system’s phase space diverge exponentially from each other. Therefore, the state of any predictive model will diverge from the actual state of the system after some characteristic time.
See Per Bak “How nature works: the science of self-organized criticality”

Keep in mind that the term “goal” is not meant to imply any “conscious intention” or teleological aspect but merely the fact that the system is optimized for reaching a certain state or increasing a performance measure. The system’s beliefs about it’s goals may even differ from the actual goals, as it is often the case with humans.
This is mathematically proven for chaotic systems. Keep in mind that determinism and unpredictability can coexist.

###

Dr. Arthur Franz is a physicist and AI researcher and previously did research at the Frankfurt Institute for Advanced Studies, Frankfurt, Germany.

Talking about the Computational Future at SXSW 2013

ORIGINAL: Talking about the Computational Future at SXSW 2013

March 19, 2013

Last week I gave a talk at SXSW 2013 in Austin about some of the things I’m thinking about these days—including quite a few that I’ve never talked publicly about before. Here’s a video, and a slightly edited transcript:

Video streaming by Ustream

Well, this is a pretty exciting time for me. Because it turns out that a whole bunch of things that I’ve been working on for more than 30 years are all finally converging, in a very nice way. And what I’d like to do here today is tell you a bit about that, and about some things I’ve figured out recently—and about what it all means for our future.

This is going to be a bit of a wild talk in some ways. It’s going to go from pretty intellectual stuff about basic science and so on, to some really practical technology developments, with a few sneak peeks at things I’ve never shown before.

Let’s start from some science. And you know, a lot of what I’ll say today connects back to what I thought at first was a small discovery that I made about 30 years ago. Let me tell you the story.

I started out at a pretty young age as a physicist. Diligently doing physics pretty much the way it had been done for 300 years. Starting from this-or-that equation, and then doing the math to figure out predictions from it. That worked pretty well in some cases. But there were too many cases where it just didn’t work. So I got to wondering whether there might be some alternative; a different approach.

At the time I’d been using computers as practical tools for quite a while—and I’d even created a big software system that was a forerunner of Mathematica. And what I gradually began to think was that actually computers—and computation—weren’t just useful tools; they were actually the main event. And that one could use them to generalize how one does science: to think not just in terms of math and equations, but in terms of arbitrary computations and programs.

So, OK, what kind of programs might nature use? Given how complicated the things we see in nature are, we might think the programs it’s running must be really complicated. Maybe thousands or millions of lines of code. Like programs we write to do things.

But I thought: let’s start simple. Let’s find out what happens with tiny programs—maybe a line or two of code long. And let’s find out what those do. So I decided to do an experiment. Just set up programs like that, and run them. Here’s one of the ones I started with. It’s called a cellular automaton. It consists of a line of cells, each one either black or not. And it runs down the page computing the new color of each cell using the little rule at the bottom there.

OK, so there’s a simple program, and it does something simple. But let’s point our computational telescope out into the computational universe and just look at all simple programs that work like the one here.

Well, we see a bunch of things going on. Often pretty simple. A repeating pattern. Sometimes a fractal. But you don’t have to go far before you see much stranger stuff.

This is a program I call “rule 30“. What’s it doing? Let’s run it a little longer.

That’s pretty complicated. And if we just saw this somewhere out there, we’d probably figure it was pretty hard to make. But actually, it all comes just from that tiny program at the bottom. That’s it. And when I first saw this, it was my sort of little modern “Galileo moment”. I’d seen something through my computational telescope that eventually made me change my whole world view. And made me realize that computation—even as done by a tiny program like the one here—is vastly more powerful and important than I’d ever imagined.

Well, I’ve spent the past few decades working through the consequences of this. And it’s led me to build a new kind of science, to create all sorts of practical technology, and to make me think about almost everything in a different way. I published a big book about the science about ten years ago. And at the time when the book came out, there was a quite a bit of “paradigm shift turbulence“. But looking back it’s really nice to see how well the science has taken root.

And for example there are models based on my kinds of simple programs showing up everywhere. After 300 years of being dominated by Newton-style equations and math, the frontiers are definitely now going to simple programs and the new kind of science.

But there’s still one ultimate app out there to be done: to figure out the fundamental theory of physics—to figure out how our whole universe works. It’s kind of tantalizing. We see these very simple programs, with very complex behavior.

It makes one think that maybe there’s a simple program for our whole universe. And that even though physics seems to involve more and more complicated equations, that somewhere underneath it all there might just be a tiny little program. We don’t know if things work that way. But if out there in the computational universe of possible programs, the program for our universe is just sitting there waiting to be found, it seems embarrassing not to be looking for it.

Now if there is indeed a simple program for our universe, it’s sort of inevitable that it has to operate kind of underneath our standard notions like space and time and so on. Maybe it’s a little like this.

A giant network of nodes, that make up space a bit like molecules make up the air in this room. Well, you can start just trying possible programs that create such things. Each one is in a sense a candidate universe.

And when you do this, you can pretty quickly say most of them can’t be our universe. Time stops after an instant. There are an infinite number of dimensions. There can’t be particles or matter. Or other pathologies.

But what surprised me is that you don’t have to go very far in this universe of possible universes before you start finding ones that are very plausible. And that for example seem like they’ll show the standard laws of gravity, and even some features of quantum mechanics. At some level it turns out to be irreducibly hard to work out what some of these candidate universes will do. But it’s quite possible that already caught in our net is the actual program for our universe. The whole thing. All of reality.

Well, if you’d asked me a few years ago what I thought I’d be doing now, I’d probably have said “hunting for our universe”. But fortunately or unfortunately, I got seriously sidetracked. Because I realized that once one starts to understand the idea of computation, there’s just an incredible amount of technology one can build—that’s to me quite fascinating, and that I think is also pretty important for the world. And in fact, right off the bat, there’s a whole new methodology one can use for creating technology.

Andrew McAfee: Are droids taking our jobs?

ORIGINAL: TED

Robots and algorithms are getting good at jobs like building cars, writing articles, translating -- jobs that once required a human. So what will we humans do for work? Andrew McAfee walks through recent labor data to say: We ain't seen nothing yet. But then he steps back to look at big history, and comes up with a surprising and even thrilling view of what comes next. 

TEDTalks is a daily video podcast of the best talks and performances from the TED Conference, where the world's leading thinkers and doers give the talk of their lives in 18 minutes (or less). Look for talks on Technology, Entertainment and Design -- plus science, business, global issues, the arts and much more. Find closed captions and translated subtitles in many languages at http://www.ted.com/translate