AI and the potential of Open-Ended models

I’m really excited about an article from Google DeepMind this week, so we’re going to talk about it. Let me start by saying I am not a trained AI professional. I am self-taught, and learning as we build*, so I try to spend an hour or two each week catching up on the latest developments in the very fast-moving world of AI (or at least glance over the chatter). The article that caught my eye is Open-Endedness is Essential for Artificial Superhuman Intelligence by a consortium of researchers from DeepMind and University College London. What struck me in wading into this concept of open-ended models was how plainly human the logic flow behind these models is and how closely it is tied to evolutionary theory. But first let me try for a quick summary. The authors are providing a formalized definition of what constitutes “open-endedness” in terms of an autonomous system. Where foundation models are trained on static data, open-ended models are continually adding new information to existing training to produce outputs that are i.) novel and ii.) learnable to an observer. The ‘observer’ can be a human (where ‘novel’ is subjective, but requires a gain in unrepeated information) or the system itself. The outputs of the system must become increasingly unpredictable (novel) while also, when added to past information, improve the predictability of future outputs.

It is easy to fall into the trap of seeing AI as ‘other’, but, more than anything, humans are modeling intelligence off of ourselves and what we know. Open-ended models are a reflection of the human experience of learning - taking in new information, identifying it as novel, combining it with existing knowledge, and thinking creatively about what this combination means (i.e. having an idea), and then, depending on how interesting this idea is to you, potentially going to test this idea through experimentation or by gathering more novel information. They also add that the system should share the findings in a way that humans can interpret and understand it, which will obviously depend on the outputs and the humans doing the interpreting. This sounds a lot like academic research, no? Or, in a simpler capacity, like talking to a small child that has suddenly discovered dinosaurs and is absorbing, synthesizing, and distributing their findings in rapid iteration because there is so much that is new to them. 

A lot of AI research goes into trying to replicate what humans have evolved to do with our nifty little primate brains, so it makes sense that areas like Reinforcement Learning (a system learning what action in an environment results in positive or negative feedback - ouch, the stove hurts if I touch it when its “condition” is hot) have so many pathways where they’re being applied. Now, the question becomes if we are limiting ourselves by prioritizing human intelligence in AI development (the authors did specify “superhuman” intelligence, rather than just “super” intelligence, which I appreciated). I somewhat disagree with the authors when they state that they “posit that superhuman intelligence will be interesting to humans only as far as humans can learn to understand it” (pg. 17-18) as I think that’s a rather pessimistic take. One of the best attributes of humanity (in my opinion) is our curiosity. We’ve been trying to understand octopus intelligence for decades and while we still don’t know how they process their world, we haven’t found them uninteresting. We like puzzles, and we’re just creating more for ourselves.

But back to why this is relevant to what we’re building. I’m excited about open-ended models because humans, by definition, are open-ended. Not just our processing of the world around us through sensory input, but everything we physically experience on a day-to-day basis. You may argue that our lives aren’t inherently novel and we aren’t absorbing new information constantly. But! Your physical system is constantly changing in response to your environment and, more fundamentally, it’s changing as you age. Every minute is new because your body has changed, maybe only slightly, but it has changed. Past models have struggled to capture human movement with granularity because it is so variable and input-responsive. There are bounds, of course, but also many combinations and variations to movement within those bounds, just within an individual. I see open-ended models combined with data from wearables as the most logical approach to capturing this constant system change and the way to tackle building AI for humans.  🧠

*This is generally my MO. I did something similar when I jumped into an Anthropology PhD having never taken an Anthropology / Human Evolution course in my life. Shoutout to Dr. Kramer for facilitating + putting up with my learning on the fly!