Demis Hassabis's new AGI test from Google DeepMind
Published on
I just watched a video of Demis Hassabis, founder of Google DeepMind and a Nobel Prize winner, that really got me, talking about what could be a true test of whether we’ve reached an AGI (Artificial General Intelligence), a term that has been trivialized by marketing-happy CEOs.
In the video (which I leave below), Demis said his definition of AGI has never changed, and that it would be “a system capable of exhibiting absolutely all the cognitive capabilities that we humans have”.
In other words, it’s not about passing the bar exam or winning a gold medal in coding competitions. It’s all the capabilities, and more importantly, having them consistently (and not that thing of correctly planning a rocket’s orbit at 11:59 PM and getting the capital of Australia wrong at midnight). That’s something today’s models still don’t have, which has been called “jagged AI”.
A very interesting insight he brought up is the reason they study deep neuroscience over at DeepMind. It’s not just for inspiration, but for empirical data. The human brain is the only living proof we have, in the entire known universe, that general intelligence is physically possible.
But the most interesting part of the video is the final test he proposed. Basically, train an AI model with all of human knowledge, but with a knowledge cutoff in the year 1911. And then see if the model can, on its own, using only that data up to 1911, discover the Theory of General Relativity, exactly as Albert Einstein did in 1915.
If the model manages to do that, according to him, we will have reached AGI. If not, we’ll still just be building incredible pattern recognition systems.
And according to Demis, we’re “a few years” away from getting there. Not decades, but a few years. Imagine the impact of a system that can “be Einstein” simultaneously in every area of science and human knowledge?
As for whether this can happen within that timeframe, I believe neither I nor Hassabis is sure (me less than him), but I found the idea of the test extremely interesting.
I’m leaving the video below with that excerpt of his talk for anyone who wants to check it out in full.