“I used to believe consciousness is evolution's solution to dealing with big data. I've now come to wonder whether life evolved to embody consciousness.”

KV: Jared Tarbell suggests that what distinguishes life from computation is resilience. Alexander Mordvintsev argues that learning systems begin to overcome the fragility of Conway's original Game of Life. From your perspective, what quality still separates today's AI from something we would genuinely recognise as alive?

MA: What's interesting to me is not only the emergence of higher level complexity and behaviour, but also the emergence of higher-level goal-oriented behaviour -- higher order formations developing capacity to solve problems to achieve goals. And ultimately, even the emergence of goals themselves. 

Many popular emergent systems like Conway's Life, physarum simulations, boids etc. show so-called emergent behaviour. From relatively simple rules, highly complex -- and often unpredictable -- behaviour emerges. And they look really interesting -- even pretty. And one might argue that the higher order formations do appear to have some kind of goals. E.g. Conway's 'gliders' seem to 'want to' glide across the screen, 'glider guns' seem to 'want to' create gliders etc. But arguably, these aren't very sophisticated goals or complex problems. And the system definitely isn't very good at pursuing them. You put one block in the way of a glider and it 'dies'. Mordvintsev et al's Neural Cellular Automata is indeed an absolutely brilliant approach to overcoming this brittleness and I love NCA :). And yes, we can say this relates to resilience - finding different pathways to achieving a certain goal. But what is that goal? Where does it come from? Where is it stored?

This is for example the motivation behind our recent work on ant simulations https://superradiance.net/ants.

This is something life does remarkably well: creating goals, subgoals, subsubgoals, hypergoals, hyperhypergoals. And lots of them, interlinked, crosslinked, and hyperlinked across so many spatial and temporal scales. 

And this brings me to my second point, which I also hinted at in my first reply. 

The machines that we typically build can be extraordinarily intricate, complex, even nonlinear, and adaptive. But they're generally designed to be modular and decomposable, with components performing specific functions that have relatively limited and legible causal pathways. This is generally how human design and engineering works, because it's tractable. 

Living systems are not like this. We can break living systems down, and look at a flagellum and see how it resembles a propellor, or look at the Electron Transport Chain and visualize it like a beautiful piece of complex machinery. And we look at all of life this way -- a beautiful clockwork machinery (thanks Descartes). I also used to subscribe very much to this 'life as machinery' view. I'm now very much against it. Not because I think life has some magical non-physical aspect to it. But because life is nothing like the kinds of machines we are accustomed to when we use the word machine. The machine metaphor becomes dangerously misleading when we mistake a small partial decomposition for an adequate model of living organization as a whole. And every time we say life is a machine, I think we are subconsciously dangerously radically over-simplifying -- and killing -- it. 

Living systems are not easily decomposable. They are full of recursive coupling across so many scales, dense networks of nested and cross-scale feedback loops operating simultaneously across different spatial and temporal scales. Molecular networks interact with each other, and affect cells; cells constrain and affect molecular networks. Cells interact with each other and affect tissues; tissues constrain and affect cells. Tissues affect organs and organisms. Organs and organisms affect tissue. Organisms interact with each other. With the world. The whole thing is a deeply entangled beautiful mess. And causality doesn't flow from the smallest parts upward. It flows upwards, downwards, sideways, in all directions. There are feedback loops on every level. And on radically different timescales. From milliseconds to millennia. All participating in the process of maintaining, repairing, and reproducing the organization that makes those processes possible in the first place.

There is a serious danger in reducing a living system to a legible decomposable model, then mistaking the radically oversimplified model for the real thing. This is demonstrated tragically beautifully in the opening chapter of James C. Scott's wonderfully depressing book Seeing Like a State using the catastrophic failure of 18th-century Prussian Scientific Forestry as an example. I won't go into it now, curious readers can get a glimpse in the introduction of Tega Brain's Essay The Environment is not a System.