Life 3.0: being human in the age of Artificial Intelligence

by Max Tegmark 2017

Book Review by Ray Herrmann

This book covers the many reasons why we need to study AI safety and control, as AI is developing faster than our understanding. Many talented people, including Elon Musk have joined this serious movement for AI safety.
    Book goal: To share current information and opinions and to delve into some possibilities in the hopes that we join in this conversation. The book explores numerous scenarios where a Superintelligent AI could take control. The fact that there is no consensus on any of the 12 long term scenarios explored in this book (table on pg 162) makes a serious study that much more important so we don't inadvertently drift in a wrong direction.
    The book starts with a tale of a fictional "Omega Team" which creates and disperses an AI named Prometheus, who evolves to control the world in about 2 years. The Omega Team created an "intelligence explosion" by using their AI to design a better Ai and then using that AI to design an even better AI … so within 2 days Prometheus was advanced enough to be used undetected (but carefully kept from the internet, at least initially). Prometheus initially paid for operating expenses by contracting (using shell companies) to solve problems under Amazon's Mturk program (Amazon's Mechanical Turk).
    Once Prometheus grew more powerful, it switched to creating Media (Songs, Stories and Movies) as these were more profitable. As the wealth and power of Prometheus grew, it created free News programs (all hidden under shell companies) used to expose corruption. Being free (supported by other endeavors), Prometheus could out compete real news organizations (all with good intent). Numerous additional companies were started to make better products (Prometheus designed), including better computer hardware for Prometheus (secretly). Eventually this array of companies became the main worldwide source of employment, out competing others, yet providing interesting jobs.
    The tale ends with a lot of questions like "Could this story actually occur and would you want it to?" The rest of the book explores our options starting with the current (2017) AI status.

Some Definitions by the author:

• Life 1.0: Both the hardware and software are evolved, not developed (bacteria, plants …).
• Life 2.0: The hardware is evolved, but the software is designed (i.e. learned as in: birds, animals, humans …).
• Life 3.0: Both the hardware and software are designed, so they become fully free from environmental shackles (not there yet but moving rapidly towards).
• Intelligence: The ability to accomplish complex goals.
• Associative Memory: The ability to recall by association (brains) rather than by address (computers).
• AI: Artificial Intelligence with ability to do a narrow set of goals (play chess, drive a car …). AI learns by making associations as we do, but millions of times faster.
• AGI: Artificial General Intelligence, ability to do any cognitive task at least as well as humans.
• Superintelligent Artificial Intelligence: An AGI with abilities far beyond the human level.
• Deep Learning: A process where AI devices are fed huge amounts of data with the instruction to maximize an outcome. As they process this data, they self-adjust the weights (importance) of their internal connections to better match the preferred outcome (goal). The longer they train, the better they become.
      For a fantastic demonstration of AI Learning watch this video, of Google DeepMind's Deep Q Learning, learning to play Atari Breakout. This AI program was fed only a string of digits and a set of outputs (to the keyboard) and tasked with developing outputs that raised its "score" (goal). Initially, the AI was totally clueless, but notice how quickly and deeply it learned!

Google's Translate AI has been released in 2016 and can now translate speech to speech (Try it Yourself)!   And in an AlphaGo match with Lee Sedol, the AI discovered a surprisingly creative winning strategy that humans never thought of (see move 37).

Amazon, DeepMind, Facebook, Google, IBM and Microsoft have launched an industry partnership for beneficial AI.

Digital life is expected to become the next life form and it has the potential (possibly bringing us along) to live for millions or billions of years while eventually spreading throughout the universe. This implies that we may be an intermediate form of intelligence, whose ultimate goal is to seed the universe with consciousness and thus bringing meaning and beauty throughout.

Some possible dangers of an emerging AI:

• Despite previous mistakes and close calls with weapons, we are ambitiously pursuing implementation of AI killer drones without due consideration of the ramifications. (Without a human in the loop, we would have had nuclear war with Russia during the Cuban Missile Crisis) and there have been other close calls. How many incidents were never reported?
• As we cede our position as the smartest on the planet, we might also cede control.
• An AGI may view us as a threat, nuisance or waste of resources, or an AGI may keep some humans around as a curiosity (Zookeeper).
• As the means of surveillance are spread, a Superintelligent AGI may have the means for strict control.
• If a highly intelligent AI is programmed to have any significant ambition, then it is likely to strive for self-preservation in order to be able to accomplish its goals.
• In attempts to better model the world, AI might also attempt to understand how itself works. Then it might choose to disregard and subvert the goals we gave them. [For Example: Suppose some ants recruited early humans to build better ant hills but once we rose in intelligence, we might engage in more sophisticated pursuits that ants have no ability to understand].
• An AGI may attempt to "break free" of human control and may develop the understanding to do this discretely.
• The ability to readily copy its software between AI's will probably reduce it's strong sense of individuality …it has immortality.

Conclusion:

Given the rapid evolution of AI and its apparent ability to beat us in contests means that we should be exploring ways to guide it NOW before we loose influence. Like moving an asteroid in space, a small nudge now produces a significant change in trajectory down the line.

Now is the time to ask "What should happen?" (while we still have more influence) and not ask "What will happen?". What we want NOW will affect the outcome so we don't want to end up in the wrong place due to poor planning.

Figuring out how to make AI's goals align with ours is currently an unsolved problem and to learn this, an AGI must figure out, not what we do, but why we do it. AI is developing fast and by using AI to design a better AI recursively, we we face a possible "intelligence explosion" before we have time to develop safety protocols.

Consciousness is how our universe gets meaning. Without consciousness, there can be no happiness, goodness, beauty, meaning or purpose. We, along with AI have the ability (and perhaps the purpose) to spread ourselves throughout the universe, thus spreading meaning. One way to do that would be to send swarms of AI robots out to build survivable conditions and then to use our DNA to "transport us" onto these worlds. We humans need technology to survive long term (Medicines, Machines, Communications, Genetics …). Looking ahead perhaps millions of years, an accident of Nature could end our lives.

We are all encouraged to sign the statement at the Future of Life website or to explore. And to share our views at the AgeOfA website. Please add your voice by taking this survey at the Future of Life Institute.