A Step Toward Artificial General Intelligence

Will Douglas Heaven published an intriguing article in MIT Technology Review May 27, 2021 (AI is Learning How to Create Itself) beginning with a sketch of Rui Wang’s delightful “little stick figure with a wedge-shaped head shuffles across the screen. It moves in a half crouch, dragging one knee along the ground. It’s walking! Er, sort of.” Rui Wang leaves the little guy fend for itself in its software environment, “trying to navigate a crude cartoon landscape of fences and ravines without falling over.”  Its name is POET (Paired Open-Ended Trailblazer) – “a training dojo for virtual bots.”  But POET is learning for itself by trial and error as it generates its obstacle courses without human intervention. This experiment is an attempt at “a revolutionary new way to create supersmart machines by getting AI to make itself.” Machines that can outthink humans. (Wang is an AI researcher at Uber.)

Jeff Clune was a colleague of Rui Wang, both working at Uber AI Labs and attempting to build truly intelligent Artificial Intelligence. They think POET “might point to a shortcut. We need to take the shackles off and get out of our own way.” Clune wrote a lengthy, ambitious article while at the University of Wyoming, (home of Uber AI Labs), AI-GAs: AI-generating algorithms, an alternate paradigm for producing general artificial intelligence – https://arxiv.org/pdf/1905.10985.pdf.  

Clune claims that AI can discover “solutions that humans haven’t found,” as in the example of Alpha Go by a solution never thought of.  Clune moved to OpenAI, where “bots learned to pay hide and seek.” He hints that “AI might arrive at technical solutions human would not think of by themselves, inventing new and more efficient types of algorithms or neural networks.”  Clune sites the example of computer vision that eventually taught itself object recognition – (https://www.analyticsvidhya.com/blog/2018/09/mit-computer-vision-teaches-object-detection-45-minutes/). 

Clune believes that in Darwin’s theory of evolution we have a matrix of AI.

“The very simple algorithm of Darwinian evolution produced your brain, and your brain is the most intelligent learning algorithm in the universe that we know so far.” His point is that if intelligence as we know it resulted from the mindless mutation of genes over countless generations, why not seek to replicate the intelligence-producing process—which is arguably simpler—rather than intelligence itself? 

He further argues his point:

Intelligence was never an endpoint for evolution, something to aim for. Instead, it emerged in many different forms from countless tiny solutions to challenges that allowed living things to survive and take on future challenges. Intelligence is the current high point in an ongoing and open-ended process. In this sense, evolution is quite different from algorithms the way people typically think of them—as means to an end.

Returning to our little stick figure POET, we might grasp the depth of Clune’s reasoning in that the “real breakthrough may come from building algorithms that try to mimic the open-ended problem-solving of evolution – and sitting back to watch what emerges.”  We know that research has already engineered machines that are learning on their own, having been trained to find solutions like machines that can do more than single tasks to overcome catastrophic forgetting. Please check out the article posted in PNAS (National Academy of Sciences) – Overcoming catastrophic forgetting in neural networks – https://www.pnas.org/content/114/13/3521. 

Take-away: We might have taken a giant step with POET toward AGI. In the meantime, you will do well to read Will Douglas Heaven’s article AI is Learning How to Create Itself.