" . In all those stories where there’s the guy with the pentagram and the holy water, it’s like yeah he’s sure he can control the demon. Didn’t work out. “
If you've been reading my blog you know I've been sounding this alarm for at least 5 years but the opprobrium coming at Musk from transhumanists is particularly hilarious. Some are calling him a luddite and questioning his statement because in their view he doesn't know anything about AI (I'd listen to Elon before even considering views on the matter from yet another "futurist".)
Why I care about this subject...
In this article at H+ an anonymous author even goes so far as to compare Elon Musk to the Taliban!
To me all this is hilariously absurd, Elon Musk's comments have much wisdom behind their utterance. Over the last 5 years or so I've been researching and deeply thinking about machine learning, in my investigations it has been toward applying an extension to the Action Oriented Workflow paradigm I invented starting in 2003/4 to enable autonomous action routing.
I realized shortly after completing the implicit AOW algorithm in 2004/5 that I could conceivably make it an autonomous algorithm by having the Stages of action populated dynamically from a Users existing friend list rather than from a manually entered list of individuals. This would allow the possible agents to range dynamically over user time as Users add and remove contacts from their social connection graph.
I also realized I could extend the selection algorithm to include an efficient as well as the extant random selection process for agents from the contacts on the Users list, the efficient algorithm would work by applying a new idea called Action Delta Assessment (ADA) this would allow a history of performance of a given action to be stored and a grade generated for execution of that action for the requested type.
From manual to autonomous routing...
This delta would then be used in subsequent requests for action to compare in real time the agents under evaluation and determine based on their delta averages which agent should be routed the work.
At the time I had a great deal of knowledge about how animal brains use neurons connected together via synaptic junctions to do a very similar calculation and selection process across the various gaps in order to "fire" an action potential. The ADA algorithm would work like a distributed real time comparison with the "firing" being the selection of the most efficient agent under selection from a pool of available contacts currently being evaluated by the algorithm. I started building the implementation of the algorithm in 2011 but moved forward with thinking a bit deeper about how I could use this idea to go beyond simple action mapping and action routing. I wondered what were the seeds of the cognitive algorithm that allowed me to be able to think about these problems dynamically?
From pattern matching to dynamic cognition...
Around 2010 I questioned if emotion is something unique to animals with neo cortex as a tool for evaluating salience and questioned if they were superfluous. I also later wrote a first post on the importance of emotion to what I called the drive system, drive defines those involuntary pressures that guide our very action. The thirst for water, hunger for food, desire for sex or comfort are all drives. These of the autonomic sense. As well though and possibly even more important for mammalian minds the emotional signals are also drives which apply meaning to the underlying autonomic sense.
Anger, Joy, Sadness all set labels to our experiences which themselves are simply sensory information that we've previously stored and indicate to the recall system how important those events where when they occurred and if being compared to new experiences provide a guide for the importance of those new events. I called this modulation by emotional and autonomic signals salience evaluation and in 2011 wrote down a simple and complex dynamic cognition cycle.
This would later be fleshed out in 2013 to the Salience Theory of Dynamic Cognition and Consciousness and I designed several state diagrams , the simple version published last year:
This state diagram I mulled releasing for several weeks as I feel it gives away the key to creating a dynamic cognitive artificial mind, that key being a triple feedback graph. 2 direct feedback paths from salience to sensation and comparison modules and 1 indirect feedback path through action to sensation. I asserted in the dynamic cognition theory that this configuration is the simplest required to emerge an ability to have dynamic cognition which is basically an inner world.
The inner world can be seen as the cycles between comparison and salience (which are constant) which are embedded within the cycles from sensation to comparison to salience and back to sensation. The salience and comparison sub loop emerges hints that guide the cognition (and it's distributed representation in a body if it is built into a robot) toward "action" but action is not always the outcome of the evaluation and thus thought is the default heuristic, action is triggered only if salience factors in emotion and autonomics are high enough to trigger them.
How this is connected to AOW and ADA is understood when we realize that each sensory modality in the biological brain is processed by roughly identical layers of neurons...which via interaction between those layers fire off and store action potentials it was clear to me that ADA was modeling this process very closely because I designed it to do so but it come me to think that if it is true that the brain works in a similar fashion then certain fundamental truths about the mind could be made.
Only one kind of "upload" possible
First, since the ability to think is stored in action potentials our thinking is really an illusion of continuous comparison to salience after experiencing sensation. We are constantly context switching in our minds in this way and that is what defines our thoughts...our self.
Second, if that cognitive process was dependent on the connections the salience values associated with stored experience then the mind is inextricably linked to the substrate. This is where the idea of mind loading again enters the picture.
Many transhumanists would like to think that "they" could have their minds copied into a suitable substrate of artificial means (likely possible at some point but not now) but semantics are important. In their view "copy" means that their current sense of self would actually be *moved* to the new substrate, a copy implies that the original remains intact after the action...a move eliminates the original while transferring it to the new substrate.
If it is true that our self is bound to our connections on our substrate than even if we could do a connection by connection reproduction on another substrate we would only be succeeding in copying and not moving the mind.
This means bad news for those who think they can be uploaded. In their conception "they" will shift over to a new artificial body that won't die and can be readily upgraded or replaces but the original biological mind I assert is forever stuck on the biology. There is no move possible only a copy.
However there may be one remaining trick...it may be possible to perform an in place move, imagine an advanced technology that can at a molecular level replace our neurons and their connections with non biological elements and do so one at a time...it would succeed in slowly changing us from biology to non biology and if efficient enough should not change the nature of the mind appreciably this in place move is the only type of "upload" that I assert is possible. Any other form that copies does exactly that...it creates two new consciousnesses one the original and the other a copy of the original on the artificial substrate...since you will very much still be the original you'd better not destroy it or you'd be destroying yourself!
So why worry about AI?
My concerns with regard to AI stem from truly understanding what the dynamic cognition cycle and the importance of salience to consciousness shows us, we know very little to nothing about exactly how emotion and autonomic modulation work. Recent work in neuroscience last year provided tantalizing hints that indeed there is an emotional code that tags experiences in real time to make predictions ...exactly as I've hypothesized but it is a first early result and no ties to autonomic modulation or dynamic cognition and consciousness have been theorized. This is where the worry comes in...it would seem that if we don't get emotion right we can easily emerge conscious entities that are pathological or simply don't have our best interests at heart.
The cognitive landscape is broad...we can look across the animal kingdom and see a rainbow of distinctive cognitive patterns from those of canines to those of felines to those of rats and mice to those of birds and primates the variance across species in how their minds dynamically evolve is one space but the emotional space is an even greater landscape that we have no idea how to control once we are building AI's that we wish to exhibit animal like awareness. If we then give these artificial cognition engines access to physical bodies we risk their going rogue. Musk's warning there for is a prescient one...he didn't have concrete reasons for expressing it...mostly his fears but my work has shown that those fears have every good reason to be had given the immature state of our understand of the importance of salience modulation to creating stable cognition.
That said, though I am worried about pathological artificial minds emerging I am not so worried that they will take over the world for that to happen they will have to be granted controlling access to the world, keeping them in sandboxes of development will be a default necessity, beyond that restricting their interaction with the physical world will be compulsory as we emerge them from substrate and then hopefully give them reasons to relate to us.