Mind Children

Summary

In Mind Children, Hans Moravec describes how humans will, eventually, create machines that are able to think, reason and feel, just like we do. In the book he provides an interesting and detailed description for how he thinks intelligence works and an account for why he thinks consciousness is a property of information processing systems and therefore independent of the substrate it is instantiated in.

He proposes the idea that eventually we will transcend our biological bodies, creating "Mind Children" that can evolve culture, ideas and scientific insights at a pace that we can't even comprehend. Max Tegmark calls this same idea Life 3.0. What is astonishing to me is how long ago Hans Moravec wrote this book. It was originally published in 1988, long before the age of the internet and it is only now that we begin to see how right he might have been. The book is very similar in spirit to The Singularity is Near or The Metamorphosis of Prime Intellect. Moravec even predicted the ideas of scaling laws, and tried describing potential ways of building general robots that can traverse and act in the world, much like we can. We are now finally building those robots.

Moravec proposed that intelligence evolved, and that machine intelligence is the next step of this evolution. Robots and Machine Intelligence will be the intellectual descendants of humanity.

He believes that Mind Uploads are possible, that we could in theory build a machine, that without the patient losing consciousness, transfers their whole cognitive machinery into a very powerful computer. The way he describes this procedure is fascinating. He essentially envisionins a robotic neurosurgeon that while taking away neurons from the brain, maps out their exact connections and firing patterns. And then replaces them with "wires" that produce the same electrical currents and actions potentials, but entirely driven by a computer simulation. This way, a whole brain could be taken apart and replaced neuron by neuron, layer by layer, while the patient stays awake and conscious throughout the entire procedure, even though more and more of their consciousness is transferred into the simulation running on the computer instead of being instantiated in their brain.

He also introduces the philosophical conundrums that this brings with it, such as questions of identity and the "problem" of copies. In Moravecs mind, these questions have very clear answers, even if we do not like those answers. Moravec thinks that the only important thing about a brain (or a person for that matter) is the exact pattern of neuronal (and cellular) connections. If you perfectly capture and recreate that pattern, than you have made a copy that is equivalent in all aspects to the original. Just like words on a page have the same meaning, even though they exist in different places and are written with different ink, so would human minds in different substrates have the same meaning.

Furthermore given that it is the patterns that are important for conscious experience, we have to accept that machines can also have conscious experiences. It follows that some of those experiences might be incomprehensible to humans. It might feel like something to be an advanced AI recommender system or a ChatGPT. Just as we can not judge the conscious experiences of other animals (what does it feel like to be a bat?) we can not judge (or experience) the inner worlds of our potential Mind Children.

All of this leads to a vision of a future where machines dominate the cosmos. As robots become more intelligent, they will not only take over tasks on Earth but also start to explore and colonize space. What he describes is essentially the Beginning of Infinity. In this vision, machines will expand human influence far beyond Earth, possibly even beyond the physical universe. He is slightly worried about the prospects of creating superintelligent machines that are misaligned with the interests of humanity but overall still optimistic in the future.

Detailed Notes

Book Recommendation: Seven Clues to the Origin of Life — Cairns Smith

Today we rely more on information stored outside our genes than inside.

Our culture will then be able to evolve independently of human biology and its limitations, passing instead directly from generation to generation of ever more capable intelligent machinery.

We are uncomfortable halfbreeds, part biology, part culture, with many of our biological traits out of step with the inventions of our minds.

Organisms with only minds–where software and hardware can evolve together, undying– are what Hans Moravec calls Mind Children.

Chapter 1 - Mind in Motion

Our minds might be amplified by computers just as our muscles had been amplified by the steam engines of the industrial revolution.

Things can be easy for humans but incredibly hard for machines to do and vice versa. There is a clue to how to build better machines in this phenomenon.

Control theory is important for robotics.

Encoded in the large highly evolved sensory and motor portions of the human brain is a billion years of experience about the nature of the world and how to survive in it.

The bulk if intelligence is not properly defined because it happens in the unconscious and we don't exactly know what happens there (yet).

Once robots become more readily available their costs will fall and they will get better, which will lead to even more sales, in a virtuous cycle.

The graph of declining unit cost versus number of units produced is called the manufacturers learning curve.

As utility and numbers of robots increase, there is a tradeoff between people doing their jobs and people programming robots how to do their job. Successful program creation could yield royalties for years to come and is eventually more fruitful than practicing the craft itself. This has never happened so far.

Universal robots were expected by Hans Moravec within ten years. I.e. before the turn of the Millennium. This is nuts. All the components were there, where did their integration go wrong?

Robots back then had world models that might have given rise to consciousness.

Consciousness is a co-evolutionary phenomenon. It could arise once more in robots. Robots that have world models and decide based on those models are just like us and other animals with consciousness. A "routine programm" to avoid obstacles would be analog to what we know as fear.

Book Recommendation: Animal Thinking — Griffin

Book Recommendation: The Growth of Biological Thought — Ernst Mayr

Robots conscious precepts would come from the good/bad concepts given to them by their reward functions. They would behave just like animals with desires and things they'd rather avoid... they would have free will. Why wouldn't they feel the same way?

Repeated states of bring giving off a pain signal is the same as boredom.

Having a world simulator would be very useful in robots. Altering the conditions of the simulation would allow a robot to daydream, or think of creative solutions.

Tinkering with better world simulators is something that will become important for robotics in the early 21. century... It didn't...

Chapter 2 - Powering Up

10 teraops of computation should be sufficient to reproduce a human brain.

There is a theoretical maximum to how much information a single computational operation can create. This measure can be used to compare computers.

Computation got a LOT cheaper, and it got so very very fast. Trillionfold reduction of cost in ~80 years, through many changes of the technology used for computing.

He was already dreaming of faster than light speed computation back then, as well as nanotechnology sized computing machines, similar to Eric Drexler... What happened that we didn't get anywhere close to what these people dreamed up back then?

He even dreamt of turning the ultra densely packed matter of neutron stars into incredibly powerful computational machines. 10^30 times more powerful as a human mind. Crazy.

Chapter 3 - Symbiosis

Von Neumann architecture. Have general computational substrate that fan look up numbers and execute specific functions based on those numbers. Then store programs - sequences of instructions in a storage of numbers for the computer to fetch and execute from.

Handling increasing complexity of programs comes with a compute cost attached to it. Simulating human brains needs many computations, that are "unnecessary" and just translating, for things like assemblers and operating systems etc. hence the raw amount of compute needed to implement brains when accounting for this is much much higher.

Time sharing terminals are literally built on servers that split their computation time between each program that users are interacting with. Computing a bit of one, then a bit of the next and so on, without users noticing that most of the time their programs are not running. This interactivity is really awesome for programming. It also gave birth to hackers, modifying the time sharing machines to do all kinds of cool things. Like playing text based video games or chat apps to communicate with other users connected to the same mainframe.

To a hacker the interface with the computer was instinctive, fast, and immensely powerful—almost anything was possible to one who could construct the right spells.

Interactivity for computation and layers of abstraction to reduce the chaos and complexity necessary for invoking it are hallmarks of progress in computers. It's making magic more easily accessible and we have come a long way from wiring up cables in just the right way, to AI assisted smartphones automatically determining what music is playing and other such magical things.

Choosing your level of abstraction that you want to work on is very difficult. Maybe the people who built things on the lower end of the spectrum did not find the perfect solutions and there could be huge performance and efficiency gains when transcending to lower levels. However assembling new spells at the highest levels, like inventing new ML algorithms, can make magic more powerful and readily available too. And doing both is kind of impossible because the knowledge required is vast, and there are many steps in between that could be improved and for each step, to know that you are indeed improving it, you need to know a lot of things about that step before embarking on the journey. What improvements could be made on the design of the mouse, or the screen, or the keyboard, the assembler, the compiler, the operating system or the programming languages? It's difficult to say for sure, but what is also sure is that there is a LOT of improvements to be made.

The Dynabook thought up by Alan Kay one of the gurus at Xerox Parc is pretty much what a tablet in reality is these days.

Magical glasses, a small part in the book, essentially describes the invention of VR/AR much like the Apple Vision Pro, but in 1988 that radically visionary?! With magical gloves controlling temperature, resistance and touch, when grabbing objects, this merge between very advanced computation and our perception can be made real. But instead of crazy advancements there, we got smartphones, social media, google and less utopian versions of using computation. It's weird how at once he is so precinct while at the same time so wrong.

The differences between transportation and communication will become less distinct as we become more able to project our full awareness and skills to remote locations.

Chapter 4 - Grandfather Clause