Yes I agree that there are many challenges ahead on the path to AGI. Right now, I think we should focus on acquiring a better understanding of how the brain works from an algorithmic perspective and try to derive a hypothesis of general intelligence from it. After all, the brain is the only implementation of a general intelligence “platform” currently known to us.

I am sure that our brains represent nothing but one design out of a multitude of possible general intelligence implementations. However, I believe that the search-space for viable AGI architectures is just too large to be traversed by anything other than a super-civilization. Think about the staggering amount of computation mindlessly performed by evolution over millions of years to come up with the design of what we carry in between our ears.

I think it must be clear to you by now that I sit on the bio-inspired AGI camp and I definitely share your newfound fascination with the brain. Just recently, I started to tell people I am a hobbyist neuroscientist.

Reactions are interesting, sometimes hilarious.

I see whole-brain emulation as the worst-case scenario or “plan B”. If everything else fails, we will achieve AGI once we become able to emulate a whole brain down to an arbitrary level of precision yet to be determined.

That begs the question – what would be the best-case scenario?

Ultimately, I believe there is an algorithm for general intelligence yet to be discovered. Something I hope it will be as simple, beautiful and powerful as the genetic algorithm: a small set of rules that give rise to ever growing complexity and intelligence after many generative iterations.

It is unquestionable that this elusive algorithm is engraved not only on the neuronal topology of the brain but also in the rules that govern how topology changes over time. That is why any simulation of the brain must take into consideration plasticity and generative topology to be useful.

I also believe that only a very small subset of the human brain is actually responsible for general intelligence. In the best-case scenario, we will be able to identify the bare minimum amount of brain tissue necessary for general intelligence and derive powerful algorithmic insights from it. I am not talking about generating conectomes or maps but about understanding how to replicate what the brain does, not the minutia of how it does it.

Because truth be told: I don’t want an artificial brain, I want to automate work. I want to copy-and-paste scientists.

Unfortunately, the human brain is the farthest thing from a von-neumann computer (a.k.a. a stored-program computer) we could possibly imagine. Natural processes lead to the emergence of neuronal topology that then give rise to complex human behavior. Your mind is not your brain’s software – because in reality there is no software at all – information flows through the brain and computation happens naturally due to the physical properties of the neuronal pathways.

The key concept I want you to embrace is that your mind is fully described by the physical configuration of your brain. To “edit” your mind – for example, to implant a memory or instantly learn a skill – it would be necessary to either physically rewire your neurons or have your brain significantly augmented to support on-demand topology modification.

Input/Output interfaces are the most feasible in the short term

Right now we are only able to communicate with the brain by stimulating neurons (input) and measuring specific properties of neurons (output). There a lot of incredible things we can do using this approach, the key concept is to think in terms of what could be done using real-time input and output streams:

• Give people senses they don’t have (vision to the blind, GPS to the willing);
• Give people actuators they don’t have (arms to amputees, drive a car with your mind);
• Read active thoughts and intentions, including memories a person is actively conjuring;
• Give people artificial experiences using multi-sensorial stimulation;
• External knowledge databases (Google in your head);
• Ultimately, we could have an isolated brain with full-digital I/O, enabling for example, full-prosthetic bodies and disembodied living;

Science-fiction examples of I/O interfaces:

• The Matrix: the Matrix simulated world;
• Ghost in the Shell: full-prosthetic bodies, “the net”, external memories;
• Avatar and Surrogates: remote control of a prosthetic body;

Read/Write interfaces are possible but they will probably require advanced brain augmentation

There are things however, we might never be able to do using I/O interfaces because they require being able to read and modify the brain’s neuronal topology directly (read/write):

• Read a memory, without the subject actively conjuring it;
• Write a memory without generating an experience (“imprinting”);
• Significantly faster-than-real-time learning or instant knowledge transfer;
• “Editing” personality traits;

We currently lack significant understanding of how to address the challenge of building such R/W interface to the brain. First we would need significant advancements in neuroscience in order to learn how to design useful neuronal pathways. Secondly, we will need a few fundamental breakthroughs in nanofabrication and nanorobotics to gain the ability to manipulate matter with the degree of accuracy needed to make useful (and desirable) changes to a living human brain.

Science-fiction examples of R/W interfaces:

• The Matrix: instant learning through downloads;
• Ghost in the Shell: hacked memories, “puppet” agents;
• Dollhouse: personality imprints, “tabula rasa” programming;

Talking to the brain and altering the brain are two fundamentally different tasks

Although limited, I/O interfaces are the easiest to build. Even though every bit of information that enters the brain indirectly leads to neuronal topology change, the minutia and scope of these changes are not under our direct control. This means that there are fundamental limits of what we can do with I/O interfaces alone.

However, I/O brain-computer interfaces will significantly expand our mental landscape in the near term by adding new information streams to our conscious experience of the world. Yet, the dream of instant learning and mental imprints might never be achieved before we move on to considerably enhanced or artificial brains that provide easy R/W access to neuronal topology.

In other words, for the foreseeable future, you will not be downloading a kung-fu app into your brain. And when you are finally able to do so, you might not have what you currently call a brain anymore.

I see AGI as the ultimate force multiplier and as the final solution for the workforce problem. As such, I expect that at some point in the future I would be in control an AGI system that could act as my online proxy-agent, taking care of my interests, investments, relationships, etc.

The objective of such AGI proxy agent (AGI-PA) would be to intelligently automate my life as much as possible and to eventually convince me that I am better off letting it handle most of my obligations for me.

Given enough time and feedback the AGI-PA should learn to think like me (to a degree) and start making decisions on my behalf. Its decisions would initially need to be audited but just as I have learned to trust my spam filter I should eventually learn to trust my AGI-PA’s judgement.

The process of training a new AGI-PA should be similar to the process of training an off-shore virtual assistant (VA) hired from any of the currently popular outsourcing services (oDesk, GetFriday, etc).

If the AGI-PA is able to (by any means) reduce my workload, I would consider it successful by a factor that reflects how much less work I had to do in average compared to my workload before commissioning the system. Naturally, hours spent teaching and managing the AGI-PA would count as work hours.

The Tim Ferriss Test for Artificial Intelligence

I have named this test after Tim Ferriss is the author of the best seller “The 4-Hour Workweek” and a vocal advocate of the concept of outsourcing your life to off-shore workers. The test consists in having a human judge distribute several (lawful) tasks to two remote assistants over e-mail, one being an experienced human VA and the other being a machine. If the judge isn’t able to tell which assistant is the machine solely by observing the resulting work, the machine is deemed to have passed the test.

I just can’t wait to have copy & paste employees…

Brains in Silicon lab @ Stanford

An information integration theory of consciousness (PDF paper download)