Stephen Wolfram, the British polymath known as the creator of Mathematica and Wolfram Alpha and the author of A New Kind of Science, will be attending the Singularity Summit 2009 next weekend. He will engage in a “Conversation on the Singularity” with Gregory Benford, at 3 PM on Saturday.
Compares the historical prediction accuracy of different functional forms for performance curves, based on a number of information technologies and energy technologies. His results suggest that taking into account production data may lead to better forecasts than relying on time alone.
1:40 pm Macroeconomics and Singularity, Peter Thiel, Clarium Capital Management
2:20 pm Venture Capitalist Panel: Peter Thiel, David Rose, Mark Gorenberg. Moderator: Robert Pisani
Opening talk: Shaping the intelligence explosion
How do you plan for a vast and unknown future? By seeking leverage: points from which you can see and address any threat, and from which you can see and reach out for any promise. Points that let you know what actions will bring what effects, and that let you know what effects you actually want.
Intelligence turns out to be the ultimate leverage. And while intelligence does not itself imply any specific goals, it turns out that one can build particular intelligent machines that have any particular goals stably built in — machines that can design smarter machines, that can design still smarter machines, while all the while optimizing for one fixed notion of “good”. This is what we mean by “shaping the intelligence explosion” — and it’s what the Singularity Institute is for.
Closing talk: How much it matters to know what matters: A back of the envelope calculation
How much time should you spend choosing your breakfast cereal? How about choosing your career, or your spouse? The mathematical concept of “value of information” describes the extent to which our decisions are likely to be better if we gather specific information before we act. In general, we can gain more of what we value if we focus research on decisions with high stakes and large but addressable unknowns.
The stakes around artificial intelligence are the entire world and its future. The unknowns are huge, and the avenues for reducing uncertainty are many. I’ll argue that makes such risk reduction one of history’s all-time best buys — better than purchasing Manhattan for beads or investing in Google at its inception. If your values are like most people’s, you may wish to take part.
We live in an era characterized by convergence of three complementary scientific disciplines that together offer the potential to yield a breakthrough in Artificial General Intelligence (AGI). Novel computational means that are both scalable and cost-efficient facilitate the infrastructure over which meaningful AGI technology can be realized. This includes modern reconfigurable computing devices as well as graphics processing units. Advances in the field of machine learning, particularly those made over the past two decades, offer profound insight into the paradigms governing decision making under uncertainty in the mammal brain. Lastly, deep machine learning models, as biologically-inspired cognitive architectures, deliver an unprecedented capacity to represent real-world information in a manner that facilitates robust state inference – a critical component in nearly all machine learning frameworks. I argue that the pieces of the puzzle are readily available, signifying a bright future for AGI progress over the next several years.
DNA is well-known as the genetic material of living organisms. Its most prominent feature is that it contains information that enables it to replicate itself. This information is contained in the well-known Watson-Crick base pairing interactions, adenine with thymine and guanine with cytosine. The double helical structure that results from this complementarity has become a cultural icon of our era. To produce species more diverse than the DNA double helix, we use the notion of reciprocal exchange, which leads to branched molecules. The topologies of these species are readily programmed through sequence selection; in many cases, it is also possible to program their structures. Branched species can be connected to one another using the same interactions that genetic engineers use to produce their constructs, cohesion by molecules tailed in complementary single-stranded overhangs, known as ‘sticky ends.’ Such sticky-ended cohesion is used to produce N-connected objects and lattices.
Structural DNA nanotechnology is based on using stable branched DNA motifs, like the 4-arm Holliday junction, or related structures, such as double crossover (DX), triple crossover (TX), and paranemic crossover (PX) motifs. We have been working since the early 1980’s to combine these DNA motifs to produce target species. From branched junctions, we have used ligation to construct DNA stick-polyhedra and topological targets, such as Borromean rings. Branched junctions with up to 12 arms have been produced. We have also built DNA nanotubes with lateral interactions.
Nanorobotics is a key area of application. PX DNA has been used to produce a robust 2-state sequence-dependent device that changes states by varied hybridization topology. We have used this device to make a translational device that prototypes the simplest features of the ribosome. A protein-activated device that can be used to measure the ability of the protein to do work, and bipedal walkers, both clocked and autonomous have been built. We have also assembled a robust 3-state device.
A central goal of DNA nanotechnology is the self-assembly of periodic matter. We have constructed 2-dimensional DNA arrays from many different motifs. We can produce specific designed patterns visible in the AFM. We can change the patterns by changing the components, and by modification after assembly. Recently, we have used DNA scaffolding to organize active DNA components, as well as other materials. Active DNA components include DNAzymes and DNA nanomechanical devices; both are active when incorporated in 2D DNA lattices. We have used pairs of PX-based devices to capture a variety of different targets. Multi-tile DNA arrays have also been used to organize gold nanoparticles in specific arrangements. We have self-assembled a 3D crystalline array and have solved its crystal structure.
The brain is three-dimensional and densely-wired with billions of heterogeneous computational primitives. Understanding how these elements work in real time to mediate behavior and consciousness, and how they are compromised in neural pathology, is a top priority. We have recently revealed methods for real-time optical activation and silencing of specific cell types in the brain, using naturally-occurring molecular sensitizers such as channelrhodopsin-2 and halorhodopsin. Building off of these molecular tools, we have created optical hardware and algorithms for systematically testing the contribution of brain regions, cell types, and circuit connections to behavioral functions, and to inputting information into the brain. We are also working on noninvasive methods of information delivery to the brain. We discuss the application of these technologies to the analysis of neural dynamics, as well as to translation for new treatments for human disease, and eventually towards augmentation of the human condition.