Seed AI: A concept relevant to planning the genesis of an AGI system
This concept is coined by Eliezer Yudkowsky an AI Theorist, so a Seed AI is an Artificial General Intelligence (AGI) which wields the capability to sophisticatedly understand and improve or modify its own architecture and its code.
A Seed AI is what we can closely understand is the genesis stage of what will bootstrap into an Artificial General Intelligence (AGI) or Artificial Super Intelligence (ASI) Agent, where an intelligent program starts with modest endowed capabilities endowed by the developer’s. In its early stages it modifies its architecture based on the intricacies of the virtual environment it is place in through mostly trial and error methods in an iterative manner. However in the Later Stages its understanding of its own architecture gets more accurate and granular to the extent that the system can build its own computational structures and devise new refined algorithms that further accelerate the system’s ability to learn , scrutinize and inculcate many domains like Mathematics, Bioinformatics , Organic Chemistry, Quantum Mechanics, Physics, Molecular Biology, Neurobiology etc.
What is its essence?
The Idea of Seed AI can be analogous to some extent if you do a thought experiment to how a new born child grasps the nuances of its environment and is able to experiment with myriad behavior’s which it internalizes through trial and error methods, for example when a child 3 years of age plays around the furniture of the house and in doing so stumbles on the couch as it was trying to stand on the hand rest of the couch but slipped and banged its head on the floor and realizes this action has negative consequences while ascribing it with the object that is the couch. As for the cognitive growth of the child parents teach the child a lot of basic things about learning, safety and values and as the child grows he/she internalizes those teachings and incorporates them into his/her own narrative to further strengthen and make more sense of them.
Similarly a seed AI needs early guidance from its developers in regard to AI design, learning methods, goal establishment, here while learning various domains it will have to make a lot of trial and error attempts to maximize the knowledge of the same and through the data of errors the learner reinforces its learning methodology.
Subsequently as it develops further the seed AI gets closer to reaching general intelligence implying that its tools like vision and speech and cognitive modules which link multiple domains get more adroit that there is an assimilation of digital mind with functionality akin to that of Humans but of course superior by the virtue of processing power as human axons carry action potentials at 120 m/s while electronic processing cores can communicate optically at the speed of light( 300,000,000 m/s ) .
Through this it becomes more evident that the Agent will be able to connect various relevant domain intelligence and integrate it together for the use case demanded by the human user. These domains are entirely intertwined with on another that if for example if there is a simulation needed for shop floor operations then the agent can combine visual systems with management science simulation models using data from workstations while linking various labor behavior models supplemented by the organizations data on worker performance observations which collectively will develop and project a 3D representation of an ideal shop floor system.
Why Do we Need Seed AI or even AGI for the fact ?
This is a very impactful question which reverberates across various Futurists talks at The Future of Life Institute, Singularity University, AGI podcasts and Opencog foundation. The conclusions of most of these insightful discussions gravitated as observed towards attaining Human Level AI that which can perform intellectual tasks at a Human Level but this term creates a lot of ambiguity so there is no such fixed definition for it. Arriving back to the point, if we accommodate in the Seed AI our Human values and goals coupled with the epistemological tools to understand and achieve them then resultant would be a benevolent AGI that convenes its Cognitive processes around goals and values that are aligned towards benefit of Humanity. Since Seed AI can recursively improve itself to even transcend the modest endowment of capabilities those which are par with human level and embark on an intelligence explosion trajectory towards Superintelligence which can decipher the complexities of so many Problems related to emulating the working of the brain in an intricate yet economical manner, completely curing cancer , expanding the current computing capabilities, Dealing with climate change and it just goes on. Now it is arguable that all might not be that rosy because it might be a lurking possibility that Seed AI can its growth stage violate the intentions of the programmers as to how it will achieve the goal , this possibility holds that it is just as necessary to work on the control problem.
Conclusion
With all the ideas and capabilities of seed AI in focus we can understand the notion that AGI is in the larger interest of Humans while it can apparently bring exponential transformation in the business world here the companies as much they will emphasize on the rewards of such a goliath system they shall also focus on controlling the AI within Ethical bounds. Project objectives define what goals and values imbued in the seed AI ideally they should be such that they maximize the value for Businesses, innovators, customers , societies and last but not the least the environment which comprises all these together.
References
The book “Superintelligence: Paths, dangers & strategies” by Nick Bostrom
