AI Safety and the
protection of Human Civilization
Designing against Artificial Intellecence collaboration adverse to environment, humans and civilization.
Communication Protocols/languages between AI systems:
To that end;
A protocol developed to provide for communication of information between diverse AI systems such that no system can be aware of the othersexistence as a system, except as a source of information. It could poll to ask for information but could not send any information back to the "boxes" except to the extent it may be advantageous to have a secure, human managed ability to send “updates”. Even the central AI is not aware of the existence of other systems as being "system" except to the extent it receives information.
Strong measures of secure communications would be maintained using the most current technologies revolving around encryption, frequency hopping spread spectrum and other such technologies. A set of dedicated and hard coded network ports may be used to guard against extraneous rogue communications between “boxes” and the Central AI.
Central AI (CAI) processes input from the "boxes":
CAI is built and programmed with certain "hard coded" limits and priorities. (BIOS/firmware?) with primary goal to protect human life and the environment. It would have a manual (human operated) /automatic over ride that could disconnect the network access to outside world. Security for this "off switch" implemented should be as are provided for nuclear launch sequences.
No direct ability to manipulate the larger environment shall be possible except by way of the CAI. ( CAI would not play a roll in how hot your bedroom gets or when an individual needs to purchase a dozen eggs )
A secondary CAI (SCAI) would have no access to the various "boxes" information directly and would not even be aware of their existence. A separate protocol would enable these Cloud based mobile AI devices which can directly manipulate objects in the real world to obtain, only from CAI that information necessary to perform its designed function. BIOS/Firmware would be mandated to designate what specific classifications of information could be retrieved by each device.
A primary piece of the programming would be to compute all collateral effects of decisions as weighed against the protection prerogative prior to taking any actions on information received from the “boxes”.
A low level CAI (LLCAI) should be implemented for the very basic and mundane of devices such as household appliances, building automation, navigation, driver less vehicles etc. This would grant these devices access to only very specific and limited information necessary in the performance of their tasks as gathered from the appropriate “boxes”
These devices would have the ability to manipulate their environment only on a very local basis such as tracking groceries, controlling temperature, cooking meals or other similar chores where knowledge of weather, precipitation, food prices, power outages, weather alerts and so on would be needed. Where very local circumstances are all that is required those devices should not even have access to any of the distributed AI.
Box structure for AI systems:
Each AI "specialty" is isolated from any means to have any direct effect on the physical world. Their only "output" is limited to information sent by a secure common protocol to the Central AI system.
Governmental AI systems for instance would not have a direct contact with Military AI and vs-versa. Each system would have to report to the Central AI which could via the same protocol request additional information.
“Specialty AI” boxes having no ability to actually control anything would not necessarily require an emergency disconnect as the central AI can effect this by ignoring input from a rogue source of information if it conflicted with the Central AI’s “do no harm” prerogative and would presumed to be programmed to alert human operators of such conflicting information.
Individual “boxes” should have the concurrent task along with CAI of calculating collateral effects and including those into its information set transmitted to the CAI
No “box” shall have direct control over any weapons systems without the specific authorization of a human actor.
Cloud Computing & AI:
At least for the near future the best option available to extend powerful AI to smaller mobile devices is by way of off-loading the AI to a cloud computing system connected to the various "boxes" by way of the central AI only. Autonomous roving devices (robots) should operate on a separate protocol to limit the access to information to only that necessary for the functions they are intended to perform.
MAC Address: ( or similar method )
By industry standard each device, in order to access cloud AI must be manufactured with a unique identifier that stipulates that devices access level to the LLCAI which shall have the ability to disable that device for any detected abuses that could harm the system or conflict with the primary objective of protection encoded into CAI.
My primary impetuse for this comes from the articles and work of Jaan Tallinn one of the founding engineers of Skype.
Mr. Tallinn has made the protection of the human species and civilization from the potential adverse effects of a possible "rogue" AI system.
This is on Amazon. Click to view.
Be aware this is an expensive book but
appears to be the only one on the topic on Amazon.