Kurzweil's dream is to one day resurrect his dead father:
Ray Kurzweil, a prominent inventor and "futurist" who has long predicted that mind and machine will one day merge, has been making arrangements to talk to his dead father through the help of a computer.
"I will be able to talk to this re-creation," he explained. "Ultimately, it will be so realistic it will be like talking to my father."
"You can certainly argue that, philosophically, that is not your father," Kurzweil said. "That is a replica, but I can actually make a strong case that it would be more like my father than my father would be, were he to live."
I remembered this while watching the new FX & HULU mini-series Devs, directed by Alex Garland. I won't spoil it for you, if you haven't seen it, except to say that it has a similar theme. Both Kurzweil's approach, and the approach in this mini-series, are probably not going to deliver the goods; however, as I will explain, there is a method that might work.
The problem with Kurzweil's approach (and somewhat similarly, the approach in Devs) is that just collecting memories and recordings of his father -- and maybe also DNA and other biological data -- is not going to be enough constraint to produce a good facsimile of the man he knew. One problem is that what makes us who we are is very heavily determined by all the small details of life, and you will never be able to collect enough of them to resurrect the dead. I could see it happening, perhaps, if we had good enough brain state recordings; but this will not help to bring back people that died in the past.
So let me propose a crazier idea, that actually has a better chance of succeeding: resurrect everybody!
Where to begin?... Let's say you don't have recordings of what transpired at a sporting event back many decades ago; but know that the details of that event were important in the life trajectory of someone who was there. Their recollections, alone, would not be enough to constrain your model of them; nor would writeups in the local newspapers. But if, say, thousands of people attended the event, and wrote about it, and it likewise influenced their lives in various ways, and if you have access to all of their writings and other data -- their diary entries; Facebook recollections years later; DNA and biological data; behavior patterns after the event -- then you can reconstruct more about that event. All the different constraints, from all the different people, fit together like puzzle-pieces, as they all have to be logically consistent.
So, the more people and more of the world you try to model, the more constraints you will have, and the more accurate your model will be about any one of them!
Here is an interesting example relevant to my point: how lenient court judges are on juveniles could be influenced by the performance of the "home team" in sporting events:
Kids who are sentenced by college-football-loving judges who are disappointed after unexpected team losses are finding themselves behind bars for longer than kids who are sentenced after wins or predicted losses.
I could believe there is some effect, even though it may be small. But if you had data for enough judges, that attended a given sporting event, you might be able to predict its outcome, just based on how lenient they were a short time afterwards. You, furthermore, might be able to use this to predict the emotional impact on all the other people who were there that day.
Now think bigger: imagine you compiled together data from every event imaginable, every person, every street, every photo, every video, everything you can get your hands on, and fed it into a giant model of the world over the decades and even back to 1900. The model won't have access to information about absolutely everything that happened. For example, if someone played computer games by themselves in the small hours of the night back in 1980, you won't have a record of what they did. Years later, there might be a trace influence on their brain patterns that might be barely detectable. Or, if they got a high score on some game, they might have written about it later -- it might have impacted their life significantly. But otherwise, at least you could attempt a plausible model of what they did that night, what games they likely played; plausible, in the sense that it would be consistent with all other predictions, of all other people in the model.
Obviously, this is going to require massive computing resources, though still not as much as one might imagine. First of all, very crude statistical models have been around for ages. Economists build these kinds of models to predict future trends in the economy. Slightly less crude, though still very crude, are models like you see in the latest versions of Sim City, which are "mechanistic" (rather than statistical) in nature.
On the other extreme would be to simulate the world all the way down to the atomic level, or even the level of quarks. We won't have the computing power for that for millennia.
In-between these extremes one could imagine a giant neural net model of the world and all the people, animals, and objects within it. The level of crudeness could be dialed up or down, depending on the amount of data and computing resources one had available. Several decades from now (probably more than 50 years; but hopefully less than 100), for instance, we might have enough computing power to build a model using the sum total of all information on the internet, including such things as:
* Maps of all parts of the world, through the decades.
* Old photographs and audio and video recordings of people, animals, plants, cities, landmarks, oceans, rivers, mountains, forests -- everything.
* All the writings of everybody, and information about when they produced it.
* Weather maps; architectural blueprints; health records; DNA information about everybody, including people who died and will be added to the database in the future; and so on.
* Brain recordings in the future, of people alive in the present.
If you had all that data, that would give you a lot of constraints about how the world looked at some time in the past. One could attempt to build a giant "maximum likelihood" model for the world over time, conditioned on all those constraints. It would be like a inconceivably vast generalization of "video interpolation", where a model attempts to predict what happens during missing frames in a video:
Vast it would be, but it wouldn't require anywhere near the computing power of the aforementioned "atomic simulation" model. This maximum likelihood model would cut corners -- a lot of corners -- to place the simulation within the computational budget, while still seeming pretty accurate. This is sort of like how deep learning can be used to mimic a physics engine for CGI video, while using orders of magnitude less compute than an actual physics engine:
Such a model could even extrapolate; and you could insert yourself into the works as a character, to interact with people long dead.
Kurzweil will probably not be alive to see all that I have described, and unfortunately will not get to see an accurate reproduction of his father. But several decades from now (more than 50, hopefully less than 100), some of us alive might get to see some version of both Kurzweils in a Matrix simulation of the world circa 1969, say. Maybe Marvin Minsky could be consulted, told he is living in a dream world, and asked whether it is like what he expected it to be...