What would be the most revolutionary advance in AI?

Hi Olaf!

There has been a lot of progress and new applications for AI in the last decade, mostly thanks to Deep Learning. I think that recently the advances have been more incremental and less revolutionary. Do you think AI is pretty much mature and will be fine just continuing on that path, or is there any new result that would really blow your mind, and cause a complete shift in the AI landscape?

Absolutely agreed about the recent advances having been rather incremental, and overall dull. I sure don’t mind seeing such results, but I’d prefer seeing some that would surprise me. I guess there are many directions that would surprise me, here are a few off the top of my head:

• language understanding,

• common sense reasoning about physics of a system,

• learning from continuous information transfers in a group,

• theory of mind,

• design of new communication interfaces between agents,

• design of a open-ended system (emergent mechanisms leading to other ones),

• robust logic,

• temporality and multi-scale memory,

• keeping attention of another agent,

• ambiguity in Natural Language Processing,

• strong reasoning,

• playful agents.

It looks like these could be summarized into groups: environmental grounding (temporality, common sense physics reasoning, play, logic, language understanding), sociality (continuous information transfers in a group, theory of mind, new communication interfaces between agents, keeping attention), and evolution (open-ended system). What would you surprise you, that people work in those directions, or that they get results in solving these questions?

I think what would most surprise me would be a new type of AI model, possibly not even based on artificial neurons, that would be able to learn one “difficult” task in a simple (in hindsight!) way. Having such an advance would basically mean that we have found an efficient way to define our problems, so efficient that finding solutions is a piece of cake. I think for now we rely on our models to do the heavy lifting in terms of “solving difficulty”, because we don’t know how to define problems correctly. Maybe the biggest advance would just be Good Definitions! Among them the definition of all definitions: Open Ended Evolution…

Indeed, they could use some classification, which tells me this may use some longer brainstorming. The example you give about a new architecture is definitely another one. Also, not only varying the model, but also the rest of the setup: type of task, environment, substrate, etc. Otherwise, it feels like we are looking for aliens, without looking on different planets.

I guess each category on the wishlist also has its own roadmap, so that I’d be impressed by specific things.

Definitions do help.

I find an opposition recently between large models in which everything seems to emerge magically, implicitly (pattern recognition, logic, arithmetic, possibly AGI too at some point, if you keep scaling it up), and models that include a specific, explicit model, implementing some function.

Can you expand on the opposition between emergent functions and defined functions? I can’t remember the last time I saw a model where a function was emergent rather than defined. In ALife, some simulations have implicit goals (“here is an environment, now whoever has the most surviving offspring wins”), but in AI I feel like the goal is always well defined, never even partially implicit. For example even when genetic algorithms are used, the “winner” is not tied to the number of surviving offspring; on the contrary, the winner is decided using a well defined fitness function, and the number of offspring is decided from the fitness value.

Going back to the topic of “better definitions”, I should note that right now the opposite approach is the most popular and successful: let’s not really define the problem (“how do you recognize a cat from a dog?”), but have an extremely well defined goal (“this image should be classified as cat”). It’s an attractive alternative to the previous approaches, for example expert systems where every step of the task at hand and how it should be solved was defined. But maybe that increase in freedom was misplaced. We don’t really know/agree on what is the “goal” of evolution for example. Or even the “goal” of intelligence. Meanwhile in OEE simulations, we would like to do the opposite: not define a goal, as the goal must always be shifting; instead define mechanisms through which the diversity of goals might emerge.

Right. To clarify, I did not mean ‘pretty’ emergence as we are working with in ALife, but something similar to GPT-3, where you see that through exposure to a lot (sigh) of data, the agent becomes able to solve underlying problems. One example would be the emergence of arithmetic, such as the ability to multiply large numbers. I concur that this is as implicit as you get, and although not unimpressive, it’s not what one has in mind when talking about strong emergences.

The opposite trend is promising to me, and I was thinking of parts of representation learning that adds a piece of architecture dedicated to capturing some function, and demonstrates it on a much smaller/faster learning timescale in most cases. I guess one example would be something like David Ha’s world models, which although they use a specific, fixed architecture with a certain mechanism to represent some knowledge, produced some interesting results by nicely factoring out a relevant function from the system.

If I understood it well, I think I agree with your point on the misplaced increase in freedom. What I see in evolution is actually somewhat similar to gradient descent approaches: at the end of the day, there is a lack of transparency in the goals, which remain implicit post-learning (that may be fine) and are hard to get a read on (that in addition makes it more problematic). I’ve been thinking a fair bit on how to study emergent goals in agents, which is part of why I like to go with OEE simulations, explicitly aiming at increasing a second/higher degree of objective function (to me this makes it in many cases similar to metalearning, but that’d be a different discussion thread). My hope is that we get to construct a nice theory of how goals organize themselves, with respect to each other, in different environments and situations. This would make me happy.