Okay, lets consider the concept of the design space. The design space for motorized vehicles consists of just about anything that has enough wheels to be stable on the ground, has a structural frame, has an energy source, and has an engine of some kind to apply that energy source, in the form of torque, to one or more of the wheels. We have seen many examples of motorized vehicles for more than two centuries. ( https://en.wikipedia.org/wiki/History_of_steam_road_vehicles )
These days just about any bozo can produce a basic vehicle using off the shelf parts. The key there is that a massive amount of effort was required to produce those off the shelf parts. Today, all of the parts that are available were designed for purposes other than AGI, meaning that it will require a massive amount of effort to adapt them to purpose, compounded by the fact that we don't quite know what we are building. Nevertheless, we must suspect that there is a set of parameters that define the set of workable AGI designs and that this space is fairly large. We also assume that we have freedom within a basic design to swap one thing for another and still end up with a working design just as you can swap factory rims with pimp rims or race wheels or opt for some more rubber to give you a more LX ride. Similarly you can swap a steam engine for a gas engine for an electric motor without violating the basic concept of the motor vehicle. All of them will move down the road. The first thing that is required for an AGI is a system complex enough to allow it to manifest it's general intelligence. That is either a robot of a reasonable complexity, or a simulation of some kind with parameters of sufficient complexity that it demands an agent be capable of generating concepts on the fly. Here is a game that I've wasted far far too much of my life on, it's similar to Minecraft but it lets you build starships, here's one I remodeled: https://steamcommunity.com/sharedfiles/filedetails/?id=929823165 The critical point here is that the game only provides a few dozen block types. It is your job to assign them a meaning or purpose and to create concepts such as wings, compartments, hulls, etc... I'm not 100% certain this is sufficient, but it is certainly a necessary level of complexity. This is not an easy problem. Google is using Atari games these days. Building a more satisfactory simulation is probably a hundred million dollar development project. =~( ( I am flat broke w/ no income). The AI's avatar should be reasonably similar to human modalities, at least not completely alien. Secondly, AI is not a blank slate. It definitely isn't just a grey block of empty computronium.... Now I'm not exactly sure what level of complexity is necessary, or whether it is possible to start with a minimal kernel of some kind and then bootstrap complexity from that, but there almost certainly does need to be a framework sufficient enough for the AGI to start exhibiting basic behaviors almost immediately. In many cases, the brain seems to be cobbling together a solution for a high level algorithm using neurons (the only available hammer...), in those cases, you can get a 1,000x+ speedup immediately from just implementing the high level code. Normally, we would say that we just want to get it done, for the first generation AGI. In this case, however, a 1,000x speed up means getting to the singularity ten years sooner, so carefully chosen early optimizations are a big part of getting it done. Okay, my original idea was to base everything on abstractions of a form similar to A -> { W, X, Y, Z } and then construct a mind out of several billion of those. The advantage to that concept is that it can use perfectly conventional memory management techniques. The other approach is to allocate a matrix that's Big Enough (tm) for your deep neural network and basically accept that it has a learning-limit when it hits its maximum entropy. At that point you have two options, really do make it big enough to achieve enough intelligence to design version 2.0, or design an introspection algorithm that can optimize and re-allocate neural matricies, and basically patch up some of the pathological conditions neural systems can get into. Now the big problem that people are fretting about is how to "transfer knowledge between tasks". There was even a [poorly constructed] contest on this problem last year. Want to know the secret? DON'T! =P The brain does not transfer knowledge between tasks, it simply re-uses entire neural sub-units in different combinations to produce different behaviors, or to solve different problems. While a linear neural model seems to be sufficient to re-produce the quirks seen in human psychology, it is clear that a switching network also exists that dynamically reconfigures the brain for each task it performs. Replicating this functionality is probably the most important problem right now internal to the AGI itself. Anyway, I'm getting off into the woods here, but the basic questions about what functionality is actually needed remains. -- Please report bounces from this address to a...@numentics.com Powers are not rights. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T58c366e7715267bb-M9e8de2dc7b43fa646a8500f5 Delivery options: https://agi.topicbox.com/groups
