ToDo: Artificial intelligence
As commented in the previous post, there has been some hype about artificial intelligence (AI) recently. Some fear a new AI winter, some see in AI an existential risk, and some think that AI will evolve to artificial general intelligence and then artificial superintelligence in the next decades, becoming indeterminately more intelligent than the whole humanity, leading humanity to immortality.
There are two groups of factors to consider in these speculations about the future. One group refers to the economy and the other group refers to the progress of AI.
AI may have a disruptive effect on economy. Regardless of the likelihood of such an event, IMHO, if a job can be properly done by a machine, then people should probably do something more important. The shift in the skills and jobs needed could be a problem to solve, perhaps hard times for many, but not an existential risk. Similarly, economic circumstances may reduce the investment on AI, causing an AI winter. The likelihood of such an event is beyond my current knowledge and the scope of this blog post.
On the side of the AI progress there are no reasons to believe that AI will cause neither our extinction nor our immortality in the foreseeable future. The reason is that the kind of AI and research on it that could lead to those outcomes has not started yet, and not even its beginning is foreseeable at this moment.
Computers can be programmed further than other machines to display some behaviour that is not constant, e.g being a function of the input and some internal state. AI can modify this function automatically, e.g. machine learning, and in some cases, like the again popular neural networks, any function can be approximated (they are universal approximators). This certainly gives a lot of “freedom” to computers, being able to do anything that is computable, with many references to “universal” stuff. However there are strong bounds not in their intelligence per se, but in their universe.
For example, for a self-driving car pedestrians may be labeled as “person”, but they are just moving obstacles to avoid, the knowledge representation used for the car may include everything that can have an address or GPS coordinates, and every object (including people) with which it must not collide. This model could include further features like eye tracking to consider potential changes in the direction of the movement of people. However, this is not an input of the system, considering such fine grained details in such a complex scenario would require a far greater computational power with current techniques, so nobody does it, AFAIK. For the AI in the car there is no difference between people and dummies that move: colliding with them must be avoided. The machine learning in self-driving cars may learn any function, but all the functions that it can learn are enclosed in some space and nothing beyond that space exists for the AI.
There is a second example, also from Google, of an AI that plays games. Old 2D Atari games are fairly straightforward, it is a universe of pixels and it is possible to do reasonably well without some deep knowledge about what is on the screen and its meaning. Technically it should be feasible to move on to 3D games and then to robots, learning to move, including learning to walk like a person, similarly to what BigDog does. These are fairly unconstrained spaces but they still define the boundaries for the AI, and then there are the boundaries on the goals and the means. Moreover, each of these steps is going to take a lot of work and time and even if they are successfully achieved they would mean no progress towards consciousness or a self-improving AI.
I am aware of genetic programming and neural Turing machines. However, in general applications AI does not self-improve by modifying its goals, inputs or means. AI tries to find the best function in some space of functions (potentially infinite) and not beyond. The spaces of functions that could lead to some self-improving AI are very far away from current research. This means that the progress in AI is manual, it has to be crafted, progress may be empowered by a growing population and PhD numbers, but challenges become harder, costs grow, and a winter may be near.
There may be many things to fear or to hope for in the next decades, but the technological singularity is not one of them. This is one thing I would be happy to be wrong about.