AI Talk: Artificial General Intelligence

This week's AI Talk...

Artificial General Intelligence (AGI). Are we there yet? Not even close. I saw a fairly extensive article in MIT Technology Review exploring this theme. Of course, the first order question is what exactly is AGI? Short answer: A machine that thinks like a person. That said, there are widely divergent views on what AGI is.

The term and the acronym AGI was coined by Shane Legg, who is currently the Chief Scientist in DeepMind, a Google subsidiary. Dr. Legg has a long history with development of intelligent machines – a dotcom startup called Webmind (1998), a PhD thesis on Machine Super Intelligence (2008), and a founder of DeepMind (2010) with ambitions to develop general purpose intelligence. But, the field and its pursuits of general intelligence predates all of this and goes back to the original underpinnings of AI formulated during the Dartmouth workshop by John McCarthy, along with Marvin Minsky, Herbert Simon, Claude Shannon (AI luminaries) and others in 1956.

So, what do we have to show after all these years working on AI?

The promise of AI has never been in question, but its stock has seen a few ups and downs. In the late 80s there were overblown expectations surrounding what AI could do. When they failed to translate into real, useful systems,  the result was a flight of capital and the period that ensued has been dubbed the “AI winter.” The current resurgence of AI started with the remarkable success achieved by deep learning solutions. Over the past decade and the past few years, AI has shown exponential improvement in some core capabilities. And, unlike in the 80s, there have been a slew of practical applications as a result of all the recent advances.

Back to the question of AGI, or a more basic question: What is intelligence? Is it the ability to solve problems or to learn to solve problems? Or is it about being conscious about what problems can be solved? Or being a generalist able to bring multiple skills to bear? What about the incredible amount of common sense that humans exhibit (or don't)? Or the ability to generalize from a few examples, as opposed to the tens of thousands of examples that machine learning solutions seem to demand?

Once you go down this line of reasoning, there is no end in sight and none of the solutions of today measure up. There is also the vision, once a machine learns to learn, that it can quickly outstrip any human ability and be able to solve any problem. Ray Kurzweil, a famous futurist and inventor, now with Google, dubbed it the singularity, when machines overnight become omniscient and omnipotent. On the heels of the 15th anniversary of the book he wrote, Singularity is Near, he is working on releasing a 2021 sequel to that book! Here, is a related discussion you might find interesting.

Why the renewed interest in this topic? Partly, it’s the remarkable progress in a particular branch of deep learning called "Transfer Learning." GPT-3, the mammoth language model from Open-AI, showed it can learn from just a few examples how to do a variety of natural language processing tasks. This is clearly progress, as learning from few examples is a human-trait. However, GPT-3 was created by processing a huge amount of text in a totally unsupervised fashion – no where close to a human approach to learning! Practical solutions based on GPT-3 technology are probably going to show up soon, now that Microsoft has leased this technology.

GPT-3 and related solutions will provide incremental advances to move the field forward. However, that moveshould not be mistaken for reaching the overall goal of endowing machines with human intelligence. Machine intelligence is going to be narrow, focused and practical for specific applications for the foreseeable future – all the hoopla surrounding artificial general intelligence notwithstanding!

I am always looking for feedback and if you would like me to cover a story, please let me know. “See something, say something!” Leave me a comment below or ask a question on my blogger profile page.

V. “Juggy” Jagannathan, PhD, is Director of Research for 3M M*Modal and is an AI Evangelist with four decades of experience in AI and Computer Science research.