The simulation hypothesi.., p.9
The Simulation Hypothesis, page 9
In 2013, a team of researchers at MIT, while researching Alzheimer’s, found that they could implant false memories in the brains of mice. According to Susumu Tonegawa, professor of Biology and Neuroscience at MIT, these false memories end up having the same neural structure as real memories.9
If our memories of the past can be modified, does this also mean that the past can effectively be modified? Is there a meaningful distinction between these two?
Stephen Hawking raises the point that his research into black holes brought up a disturbing aspect of information loss of particles that go in to but do not come out of a black hole. “If determinism breaks down, we can’t be sure of our past history either,” Hawking said. “The history books and our memories could just be illusions. It is the past that tells us who we are. Without it, we lose our identity.”10
In common life, we are used to thinking of time as flowing from the past to the present to the future and that nothing in the future can affect what has happened in the past. But is it possible for the past to be modified in some way? We’ll explore some of the paradoxes in quantum physics that touch on this idea in the next part of this book. We’ll also look at the idea of artificial intelligence in detail in the next chapter.
While we may be years off from being able to implant false memories using chemical or electrical signals in the human brain, research on both suggestion and electrical signals show it is theoretically possible.
This raises interesting possibilities that once again start to sound like we may be in a Philip K. Dick novel. To paraphrase the title of an article written by columnist Laura Miller in the New York Times in 2002: It's Philip K. Dick's world, and we only live in it! 11
Chapter 3
Stages 9 to 10: Artificial Intelligence and Downloadable Consciousness
In this chapter, we will explore the final two stages of reaching the simulation point, both of which will require considerable progress in understanding not just video games and simulations but in understanding a term that, despite being used often, is not well understood: consciousness.
Other than video games, the other subfield of computer science most associated with the simulation hypothesis is artificial intelligence (AI). The two topics aren’t mutually exclusive, since the rise of artificial/simulated characters was integral to the evolution of video games—and as we look at the history of AI, it is intertwined with the history of gaming.
The first stage that this chapter deals with, Stage 9, is all about artificial consciousness within a simulation and gets into issues of neuroscience as well as computer science. Once again, we’ll see how something that science used to believe was biologically based may actually come down to being a set of information and computation. Once a civilization has a handle on artificially simulating consciousness, then the final stage, Stage 10, or downloadable consciousness, becomes technologically possible.
In this chapter, we’ll also look at the singularity, a term that originally meant the explosion of AI but has recently come to include downloading human consciousness from biological hardware to silicon or other digital systems, resulting in the new term digital immortality.
In the end, if consciousness, artificial or human, comes down to information and computation, then the distinctions between these two stages is somewhat artificial and starts to move into broader areas such as philosophy and religion, which we’ll explore in future chapters.
Stage 9: Artificial Intelligence and NPCs
We saw that video games have both PCs (“player characters”) and NPCs (“non-player characters”). The NPC is any character that is not a player—which means that almost all characters in early video games, which were mostly single-player games, were NPCs.
Under this definition, the aliens in Space Invaders, the ghosts in Pac-Man, while not particularly intelligent, were NPCs—since they were an automated part of the program.
Early text adventure games and subsequent graphical adventure games tried to use NPCs in a more sophisticated way. Over time, NPCs have come to refer to characters that you can interact with, and which are there to help or hinder you on your journey in some way.
From early text adventure games to today, the way that players interact with NPCs is to “talk” to them. This was originally accomplished by typing commands or questions in text and early graphics adventures. Early NPCs were very limited in what they could understand and in their responses. For example, there was a banker in Ultima that you could only say certain things to, such as “statement balance,” etc., but you really couldn’t have a full coherent conversation.
As games became more sophisticated, entire conversations were programmed into NPCs using the idea of “conversation trees” and “branching.” In these games, you had to ask the NPC the right question to get the answer; this answer was usually an important clue that you would need to complete a puzzle in the game. As MMORPGs emerged with more realistic 3D models and graphics, having more realistic NPCs became a requirement to give the player the illusion of immersion.
While the graphics generally became more sophisticated over time, the interface for interacting with characters didn’t necessarily improve very much. In fact, it was typically an overlay of what the NPC said along with several options that you could choose as a response. This offered some complexity but not much more than the old Choose Your Own Adventure books.
Eventually, some game companies, like Telltale Games, which specialized in adventure games, had the characters talk to you (via voices recorded by actors during the game development process), but you still had to choose a response (see Figure 12 for an example).
Figure 12: An example of dialog in The Walking Dead from Telltale Games
In some ways, this was a step back from typing conversation into text adventures, since the conversations could only branch down certain “paths” and were even more limited than some of the Infocom games. Some might object to calling these characters artificial intelligence, because they allowed only limited branching, rather than real conversation.
This brings us to a big question that hasn’t been fully answered: What is AI, exactly?
We all know that AI in the context of NPCs represents an artificially intelligent being, but what does that mean exactly? Common sense tells us that it’s a program that appears human—in some ways. In lieu of a formal definition, an informal definition is a computer program or artificial device that can pass the Turing Test.
The History and Rise of AI
The Turing Test
Figure 13: A visual depiction of the Turing Test 12
The Turing Test is more of a milestone than a definition, since most AI today cannot pass this test. Alan Turing, considered by many to be the father of modern computer science, conjectured a time when a machine would exhibit intelligent behaviors. In his 1950 paper titled “Computing Machinery and Intelligence,” Turing took on the question of whether a machine could “think.” Since it was very difficult to say what “thinking” would mean, Turing devised a party game to tell if a computer was “intelligent” enough in conversation that it could fool a human.
In this party game, an interrogator (in Figure 13, party C) was behind the curtain, and he was “interrogating” two parties: A and B. C is not told whether A or B is a computer, but one of them is a machine and one is a real person. Party C would start conversations (passing messages using something like a teletype machine—the best that Turing had in his time) and would have to tell the difference between A and B. If he was unable to distinguish which was the human and which was the machine, then the machine could be said to have passed the Turing Test. Of course, back then, he described it as a machine, but today we know it would be the AI program (which is software) that would pass the test, not so much the hardware.
This party game and the concept underlying it eventually became known as the Turing Test.
AI and Games: Claude Shannon and Chess
The Turing Test is not the only test of artificial intelligence. In a paper in 1950 (the same year that Turing proposed his test), MIT professor Claude Shannon posited that a computer would be capable of playing chess in a groundbreaking paper titled “Programming a Computer for Playing Chess,” and showed a computer he had built for such a purpose (see Figure 14).
This ingenious device had a chessboard on top of it so, when the computer indicated a move that should be made, the human operator could actually make the move. Once again, even though AI was thought of as a machine, over time researchers realized that the software was what was important to the AI, not so much the hardware. The software languages and abstractions didn’t exist at the time of Shannon or Turing to make this distinction.
Figure 14: The first practical AI, a chess-playing computer built by Claude Shannon at MIT in the 1950s.
Shannon’s paper laid the foundation for chess-playing programs in the future, one of which, Deep Blue, actually beat the reigning chess world champion, Gary Kasparov, in one game out of six in 1996. In 1997, the computer program actually beat Kasparov in the six-game rematch 3 ½ to 2 ½.
This was one of the major milestones that Shannon had described as aspirational for “future AI.” The milestones included, writing poetry, orchestrating music, translating from one language to another, and generally accomplishing other tasks that only humans would be capable of at the time.
Deep Mind, Alpha Go and Video Games
Not only is the history of AI and games intertwined, it continues to be in the near future. Google’s DeepMind group created AlphaGo, the first computer program to beat a professional Go player in 2015. It also beat the South Korean Go champion Lee Sedol in 2016.
An interesting twist on the “AI learns to play games” mechanic was when the DeepMind team trained the AI to play video games. This was done not through rules-based AI for a specific game, like the Tic Tac Toe algorithm I had written as a kid, but by watching the screen and controls. Arcade games like Space Invaders and Breakout were part of the initial research project’s seven Atari games.13
In this case, the AI watched the screen and decided which move was the right one. In these kinds of arcade games, this meant moving the joystick (or, rather, simulating the joystick movement by issuing an electronic command) to move in one direction or another. For a game like Space Invaders or Breakout, this meant only moving left and right, but all the old games were built on taking input in the form of 18 possible joystick movements. The team showed that it was possible for an AI to “learn” to play arcade-style games.
Given the response times available to AI algorithms, can we expect that AI will learn to play other video games, such as first- person shooters and fighting? Recently, Elon Musk funded OpenAI and announced that it had learned to play DOTA 2, an extremely popular fantasy-themed fighting game. Competitive video gaming, or eSports, is played by professionals and has become a popular spectator sport in the same way sports such as basketball, baseball and football developed in the last century. OpenAI announced that a team of five bots were competitive enough to qualify to play against professional teams!
This is an interesting twist, though not entirely unexpected. That AI could recognize what is happening in a 3D world (in this case, the 3D virtual world of the game) is part of what would need to happen to reach Stage 9 on the road to the simulation point.
A Digital Psychiatrist
Backing up a bit, one of the first simple AIs that attempted to fool the user into thinking the Turing Test had been passed was a computer program called Eliza. It was a sort-of game, but in actuality it was a digital psychiatrist created in the MIT Artificial Intelligence Laboratory in the 1960s by Joseph Weizenbaum.
Eliza used clever pattern matching and “substitution” to make you think that it had understood what you had said and repeated a question or statement about it back to you. Sometimes it would respond with “Tell me more about X” if you had mentioned X in your sentence. Sometimes it would ask questions like: “Why do you feel that way about X?” or “How come you don’t know X?”
While at first glance it seems like you are talking to an intelligent agent, if you spend more than a few minutes with Eliza you start to see some of the patterns repeat. Still, given that it was developed in the mid 1960s, Eliza was a considerable achievement in the development of AI.
In some ways, Eliza was the precedent for many of the NPCs in adventure games and a precedent to the chat-bots that we see in the early 21st century. Some of the chat-bots use very simplistic pattern matching, while others are starting to incorporate more complicated natural language processing.
Different kinds of AI techniques had to be developed in order for a computer to have a chance at passing the “Turing Test.” In the early 21st century, digital assistants like Siri, Alexa, and Google Assistant are much better at processing either text or voice than any of the video games that we have covered thus far. But just as video games drove early graphics technology, you can expect that simulated characters will drive more sophisticated AI in the future.
Figure 15: Eliza was an early digital psychiatrist that used simple matching.
NLP, AI, and the Quest to Pass the Turing Test
Of critical importance to passing the Turing Test is NLP, or Natural Language Processing. NLP is the ability of a computer to read (or listen to) and understand the meaning of natural language. How could we tell if a computer program had “understood” a sentence? This is another difficult question to answer, so it comes down to what kind of response the program gives us.
Early NLP systems were heuristic, meaning they were based on rules. The rules were initially hardcoded by the programmers of the system. This became very difficult to maintain because, in natural speech, there are many rules that overlap and change based on context, and different languages have different rules.
In the late 1980s and early 1990s, statistical NLP (SNLP)—where examples were fed into the machine-learning algorithm which then “learned” or “inferred” the rules from examples—became popular. While initially these rules, though generated by the machines, were the heart of NLP programs, the true breakthrough came with weighted probabilistic analysis.
When using weighted probabilistic analysis, the AI could choose from multiple responses based on past experience, each with different weights. The machine could “learn” the best responses over time. This approach has proved successful enough that assistants like Amazon Alexa or Google Home can recognize most commands spoken without needing to train the AI to your specific voice, a characteristic of early voice recognition systems.
This approach (SNLP) is used to get the right response, but it can be combined with other technologies such as voice output to make the output actually sound more natural. in 2018, Google Duplex, a research group at Google, demonstrated that an AI could make appointments on behalf of a human by making phone calls. This application, which was called Google Assistant, not only understood what a human wanted (in terms of making an appointment) but was capable of generating a voice that sounded, well, human, and making a phone call to, for example, a salon to make the appointment. The AI even put in natural conversation spacers like “Um.”
There was an initial euphoria that Google Duplex had passed the Turing Test. This turned out not to be true since the interactions here were limited only to making an appointment, while a true passing of the Turing Test would have to allow for longer, more open-ended conversations.
After the initial news about Google Duplex, though, there was widespread concern that robo-calls could now sound authentic and that this might lead to a whole new wave of spam phone calls! Google quickly backtracked and decided that it would always have autonomous agents making phone calls “self-identify” as an agent.
An AI that can pass the Turing Test and do other things that humans can do has been dubbed “Artificial Generalized Intelligence,” or AGI. Thus far, most AI applications have focused on specific tasks—reading handwriting, predicting certain patterns from numbers, helping a human with solving limited tasks, etc.
While developments in NLP technology have made incredible strides in the past few decades, many experts still believe that we are probably within a decade of being able to create artificially intelligent characters (or NPCs) that can pass the Turing Test, within games or in the real world. With the recent advancements in NLP, machine learning, and robotics, we may be within several decades of having robots and AIs that talk and move like humans.
Sophia, an autonomous robot created by Hanson Robotics, shown in Figure 16, has become the poster child for this coming “age of robots.”
Figure 16: Picture of Sophia, one of the first autonomous robots; (Source: ITU Pictures).14
While Sophia led many in the media to think we had passed the Turing Test, Hanson's chief scientist, Ben Goertzel, who created Sophia, disagrees. In an interview with The Verge, Goertzel expressed that Sophia was not AGI, and the only reason this was encouraging was to show that AGI was within our grasp. 15
In 2018 and 2019, Xinhua, China’s state news agency, released two virtual newsreaders. These newsreaders looked human, and were wearing professional clothes like any news anchor. The newsreaders could read news that was fed into the system and while the voice was somewhat artificial, the visuals were stunning. It was a far cry from Max Headroom, one of the first virtual characters to gain a following back in the 1980s, who could manage only a stuttering voice. These newsreaders sounded and looked real. Combinations of this virtual rendering with the output capabilities of a Google duplex, for example could solve half of the challenge of this stage.
