Hacker News Reader

Consider that the lower orders of integration of life are energetic molecular arrangements, then genetically controlled factories which make which make up the components of cells that in turn give rise to cellular replication, then cellular arrangements and multicellular organisms.

The creatures you are hoping to simulate which seem to be at the level of abstraction of a nematode, IOW a model of a very simple body with a brain of a few hundred neurons, operate in environments so diffuse and rarified that superficially they don't seem striking behaviorally, even as they do the stuff you expect simple animals to do. A casual onlooker will not be impressed by the richness of the world of the nematode. So one challenge is you could be on the right track but find appropriate results boring.

But it may be that the best you can do with a simulation is make arrangements seem interesting, because there may be an entropic blockade to stimulating life: life may be so efficient in its manifestation of degrees of abstractions, interactions, fabrication, layering, and interdependencies-- its complexity-- that organization of models using electronic computation devices can't get within many orders of magnitude of life's dynamical density and therefore simulation of life is inherently impractical. IOW you can get from life to computers, but entropy prevents getting from computers to life.

In life, is there any clear distinction between machine and code? Genomics suggests there might be, but so far fiddling with genomes is well known only for edits of taxonomically superficial traits. Has anyone shown edits that give rise to a new subspecies-- much less a new genus? Idk... but it wouldn't surprise me to find the answer is be no, and that epigenetic factors might inhibit any approach to editing new lifeforms into existence.

Is life is fundamentally emergent, without possibility of construction? I see a definitional hazard; that we may lack the language or cognitive capacity to deal with life's dynamics. These may be off limits. Or maybe not?

But it may be that the best you can do with a simulation is make arrangements seem interesting

It may be, but is there any evidence of that or reason to believe it? How would one even go about defining that problem or proving that? With computer graphics we can produce worlds that approximate the real world well enough to understand and reason about them in similar ways, even though they don't really work like the real thing at all (e.g. hollow 3D objects suffice for a lot of things). Why could we not produce code that makes a character in that world pass for "intelligent"? We don't need to copy the model of life. We just need to understand what parts/features are necessary to create a system that has enough potential to develop something we find sufficiently interesting.

> there may be an entropic blockade to stimulating life:

Interesting.-

PS. Taking this further, there might be a similar "blockade" to simulating intelligence. Or (definitely) consciousness ...

There is definitely some requirements to achieve that otherwise every simulation would be able to achieve it. The million dollar question is what are those requirements.

imo, your environment needs a certain level of complexity (what this level is no one knows), your fitness function (or whatever you use) needs a certain level of complexity and your simulation needs to run for a very long time (how long no one knows). If you think about it, our universe is incredibly rich in complexity and detail and even then, it took about 9 billion years for carbon based life to appear on Earth. So surely, in your much less complex environment, you will need the same or more "time"

The complexity and detail in our universe create an incredibly vast search space, which may hinder development rather than promote it. To speed it up we need to be able to narrow the search space by cutting some stuff out and simplifying the model. Do we need to model atoms? Molecules? Organelles? DNA? Some other made up building block?

I have been intrigued by artificial life because I think it takes a more holistic look at that question than AI. I feel like starting with the neuron is kind of a huge assumption, how have we ruled out or accounted for the importance of all the more basic building blocks and developmental milestones of life?

I also think the environment is critical. We formed and evolved as a product of our environment, we have to find food in it; our goals, behaviors and fitness level are largely dependent on/derived from it. An AI doesn't have an environment. It doesn't have any sense of living and being in a place. It can't care about anything. It's not going to see an apple fall and wonder why that happens. It's not alive or doing anything when it's not called to process an input into an output. It makes things up by stitching together words that it has no actual context or experienced meaning for. I find it hard to believe that we're going to achieve anything resembling the interesting characteristics of humans with these abstract Chinese Rooms.

I just had a thought that "intelligence" could be a misnomer or red herring in all this. A lot of modern AI seems to be built around "getting the right answer," performing measurably better on some task. But that's not really what makes humans interesting, and I don't think that's what most people dream of with AI. I would say that it's actually emotion that makes us interesting. I wonder how things would be different if it had been called Artificial Emotion. Emotion is more about experience than correctness. We need something that has some kind of "lived" experience that we can relate to. Feeding a billion words into a system of weights is just not it.

Seconded.-

(For what little it might be worth, I have recently posted/commented along these very lines ...)

Seconded.-

(For what little it might be worth, I have recently posted/commented along these very lines ...)

Edited: Found it!

- https://news.ycombinator.com/item?id=40898635

This is not an indication of any merit in the comment :)

... but as a way to showcase "prior art" in that we are thinking in similar directions (along with many people, I'm sure ...)

> have been intrigued by artificial life because I think it takes a more holistic look at that question than AI.

Totally in agreement.-

> I feel like starting with the neuron is kind of a huge assumption,

A warranted one, but a huge one, indeed, nonetheless ...

My first question would be, how exactly does the brain work? What inputs does it take, what outputs does it generate, what all is it capable of? My guess would be that the brain is too simple and kind of locked in or highly limited.

I think we need to start with modeling more basic things with as few built-in assumptions/structures as possible. It may take a long time before anything interesting develops in the simulation. Billions of years is impractical but I doubt we're gonna whittle that down to a span measured in days anytime soon. The ultimate challenge is how do you know that your system is capable of building/developing complexity, and how do you tell if it is making progress on that?

My goal was to limit how much I hard coded into the creatures function, so it could evolve as much as possible.

This was the design: The environment was made of a small number of materials:

1-Non-life - blocks etc., obstructions

2-Plant-life - is food but not a creature, not mobile etc.

3-Creatures - alive and mobile (can be eaten by other creatures)

Creatures:

1-Shaped like hexagons

2-Eyes: Located on one of the sides of the hexagon

3-Mouth: A medium length proboscis type thing on same side as eyes

Touching plant or food with mouth=eating (which transfers energy)

4-Movement/Motor:

Can move forward or backward (based on which way pointing)

Can rotate to face new direction

Special neurons connect to the motor+rotator

5-Senses - Eyes:

Each eye had 32 rays of "light" that spread out in an expanding pattern

Each ray provided instant information

Information returned per ray: distance the ray traveled before hitting something, and type of material it hit

Each ray within each eye connected to a group of input neurons so distance and material could be input

6-Senses - Sound:

Each creature emitted waves of sound in decreasing amplitude in a full circle, like ripples in a pond

This signal traveled faster than creatures top speed but much slower than instant

Sound sensor neurons covered the entire "skin" of the creature

7-Senses - Body or Mouth touching other materials:

The mouth and entire skin had direct touch sensor neurons which detected the type of material being touched

Just like the eyes, if non-life material was being interacted with then one neuron fired, if plant then a different neuron, etc.

8-Senses - Body - being eaten:

The entire skin had special sensors for when another creatures mouth is physically touching this creature (meaning it's being eaten)

9-Brain:

Randomly created, then evolved per generation Many random/semi-random/ajustable factors controlled the higher level structure of the NN, examples:

Number of layers

Number of sections per layer (section=a group neurons more tightly connected intra-section compared to inter-section)

Percent connectivity within section

Percent connectivity between sections

Percent conectivity between layers

Percent of recurrent connetions (for each of the above)

Type of neurons (in total or per section)

Types of synapes (experimented with various things here) etc.

10-Fitness function

Basically amount of energy at end.

Creature starts with X, eating adds to it, moving reduces it, getting eaten reduces it

I needed to tweak this a bit because the naive version rewarded limited movement.

The creatures needed to evolve the ability to use their eyes, their motors, to respond to being eaten by another creature, to eat another creature (just by touchig with mouth). During the early generations they would just be jiggling around randomly, but eventually some would settle in to more productive movements.

I noticed that for long running sessions (7 days), the best creatures frequently had a NN structure that was deep-ish, like 6 layers, and came together in middle layers with many signals whittled down to a much smaller number of neurons before exploding out again to connect to the control neurons.

I also built alife simulators that had emergent natural selection properties. I definitely observed there was some minimal but present tuning of the environment to get natural selection to emerge.

I also came to the conclusion you seem to have… that the environment and its complexity and attributes have as much to do with the creatures that emerge as the creatures’ inherent possibilities themselves.

In the decades since writing those (primitive Martin Gardener-inspired) simulations, I have observed from afar that there is a strain of academic evolutionary biology that has come to similar conclusions… that the environment and its attributes and complexity is a critical (coequal?) part of the evolutionary process.

> that the environment and its complexity and attributes have as much to do with the creatures that emerge as the creatures’ inherent possibilities themselves.

An interesting conclusion, and one that I am sure has implications for artificially intelligent systems.-

> Is there a level of complexity in the environment that is a base requirement?

I think you are unto something here. Either that or a necessary minimum of sensory/processing capacity to "take in"/parse/drawn information from the environment, which amounts to the same (you end up with a certain amount of environmental complexity that you can benefit from, as a "ceiling" ...

Really interesting topic. For about 5 years, I spent much of my hobby time writing a-life sims and messing with them. I wanted an "aquarium" that looked interesting that I could have running 24/7, constantly evolving. I guess I wrote maybe 10 sims from scratch in Unity, Java, Processing, and p5.js.

I have had (limited) success in getting emergent control and coordination behaviours. A couple examples:

I played with a "battleground" simulation, inspired by the "Gladiabots" game, where the entities were NN-based fighting robots organised in teams (4v4, 20v20 etc). Random sensor systems, random fire control systems, signal emitters, etc, hooked up to random neural nets where winning teams procreate and evolve/mutate. Over a LOT of generations (maybe like 20,000+), I'd see teams evolve what looked like coherent tactics, including teamwork, role specialisation, and time-based tactics/formations.

I had a "jellyfish" simulation, similar premise, with small aquatic animals of various "species" who could speciate, etc. The aquarium had qualities like light levels, food production, temperature zones, etc. You always have to be careful to not anthropomorphize too much, but I would clearly see battlegrounds develop and evolve, and successful species rise and fall with behaviours that (again over many thousands of generations) initially look like drunken sailors but evolve into what looks like basic behvaiours, at least at the level you may see in small multicellular organisms.

In my experience, getting emergent behaviours needs a difficult-to-optimize balance between system complexity and system stability. Every neuron, sensor, output, or attribute you add to the sim can cause a super-exponential growth in the "problem space" of your denizens. Behaviours that require any form of coordination take so much longer to evolve than uncoordinated behaviours because you tend to need to "get lucky" where both a signal and response evolve/mutate at the same time.

Another issue is that feed-forward networks are not great at temporal processing. Many basic behaviours require more temporal processing than I expected when I started building these things. I got these to evolve by having features like memory nodes, timer nodes, and recurrent nodes, but if I were to build another one, I'd experiment more with LSTM-like neurons and other recurrent/memory-based architectures.

In my experience (and with my horrendously under-optimised codebases), the amount of evolution you need to simulate is super-exponential as you grow the complexity of the sensors, outputs, and neural nets. In my playing around, doubling any single dimension almost always led to more than a 4x increase in the time you need to see new emergence. In my most complex sim, it got to the point where even tweaking a basic variable like temperature or day/night cycle would need me to run the sim for at least a day to properly observe the impact, and sometimes more like a week or two. I never really knew when the "evolution ceiling" was hit, which was part of the fun for me -- I love seeing a new behavior suddenly appear.

This ALIEN project is really interesting and it has a lot of cool ideas. I haven't played with it yet but I think it's going to be fun! The project I have been keeping an eye on is Neuraquarium, which has similar design goals to what I have built in the past.

> you tend to need to "get lucky" where both a signal and response evolve/mutate at the same time

If you count the total area of earth and the age of earth, how big/long should be your simulation to have the same breadth as earth ?

Well yeah, it's big numbers! These sims (at least the ones I did) can never be compared to earth, because they make assumptions that start from maybe billions of years of simulated time, and then further make tons of ridiculous assumptions about life based on our observations from this one planet, and then they massively compress our physical reality from atomic and sub-atomic interactions to insanely granular abstractions in these sims.

("How much energy would it take to 100% accurately simulate our universe from the big bang to now" is a pretty interesting thought experiment though!)

I like that interesting period on Earth from maybe 4 billion years ago to maybe 500 million years ago, where we went from amoebas learning to duel to arthropods learning advanced dueling techniques. So any given one of my sims assumes a minimum of 10 billion years of priors, and assumes that it can compress millions and millions of generations and population diversity into a much smaller number.

I've had single sims where basic building blocks go from drunken amoeba to the earliest of pretty intelligent arthropods - so that's 3.5b years of evolution. But all sims are a massive tradeoff between fidelity and duration (fast/good/cheap), and the sims I like the most are the ones that cover an era that creates emergence in even a single aspect of the simulation. I guess most of what I'm interested in simulating is vaguely on the order of 100m years of evolution, over vaguely 7 days of real-time simulation?

I wonder if we could develop any kind of rigor around knowing what certain systems are capable of, or at least some kind of framework for testing things out and being able to catalog and compare them.

I believe the sims need to tend towards less assumptions, more primordial soup, and more running time. We only have one reference point (billions of years of basic shit swirling around) and I feel like we need to start closer to that rather than skipping a bunch of stuff because it's too hard. Making things with so much baked in structure/assumptions doesn't seem to be providing the learning that we need. (I'm speaking in a very general sense, I don't mean this to come off as personal about you or your projects, they sound cool!)

PS. Not to mention the timescales involved ...