4 min read
Artificial general intelligence or AGI is an area of AI in which efforts are made to make machines have intelligence closer to the complex nature of human intelligence. Such a system could possibly, in theory, perform tasks that a human can with the ability to learn as it progresses through tasks, collects data/sensory input. Human intelligence also involves learning a skill and applying it to other areas.
For example, if a human learns Dota 2, they can apply the same learned experience to other similar strategy games, only the UI and characters in the game that can be adopted will be different. A machine cannot do this, AI systems are trained for a specific area and the skills cannot really be transferred to another task with complete efficiency and the fear of causing technical debt. That is, a machine cannot generalize skills as a human can.
Come 2018, we saw Deepmind’s AlphaZero, something that is at least beginning to show what an idea of AGI could look like. But even this is not really AGI, an AlphaZero like system may excel at playing a variety of games or even understand the rules of novel games but cannot deal with the real world and its challenges.
Some groundwork and basic ideas for AGI were set in a paper by the US Air Force. Dr. Paul Yaworsky, in the paper, says that artificial general intelligence is an effort to cover the gap between lower and higher level work in AI. So to speak, try and make sense of the abstract nature of intelligence. The paper also shows an organized hierarchical model for intelligence considering the external world.
One of Packt’s authors, Sudharsan Ravichandiran thinks that: “Great things are happening around RL research each and every day. Deep Meta reinforcement learning will be the future of AI where we will be so close to achieving artificial general intelligence (AGI). Instead of creating different models to perform different tasks, with AGI, a single model can master a wide variety of tasks and mimics the human intelligence.”
Honda came up with a program called Curious Minded Machine in association with MIT, University of Pennsylvania, and the University of Washington. The idea sounds simple at first – it is to build a model on how children ‘learn to learn’. But something like this which children do instinctively is a very complex task for a machine/computer with artificial intelligence. The teams will showcase their work in various fields they are working on at the end of three years since the inception of the program.
There was another effort by SingularityNET and Mindfire to explore AI and “cracking the brain code”. The effort is to better understand the functioning of the human brain. Together these two companies will focus on three key areas—talent, AI services, and AI education. Mindfire Mission 2 will take place in early 2019, Switzerland.
These were the areas of work we saw on AGI in 2018. There were only small steps taken towards the research direction and nothing noteworthy that gained mainstream traction. On an average, experts think AGI would take at least a 100 more years to be a reality, as per Martin Ford’s interviews with machine learning experts for his best selling book, ‘Architects of Intelligence’.
OpenAI released a new language model called GPT-2 in February 2019. With just one line of words, the model can generate whole articles. The results are good enough to pass as something written by a human. This does not mean that the machine actually understands human language, it’s merely generating sentences by associating words. This development has triggered passionate discussions within the community on not just the technical merits of the findings, but also the dangers and implications of applications of such research on the larger society.
Get ready to see more tangible research in AGI in the next few decades.