Simulating Life on my Computer like a Scientist

Chapters:
- Coming Soon
References:
- SmoothLife Paper: arxiv.org/pdf/1111.1567.pdf
- Lenia Paper: arxiv.org/pdf/1812.05433.pdf
- RealLife: arxiv.org/pdf/math/0503504.pdf
- Coq Assistant: coq.inria.fr/
- Prefetching in Functional Languages: dl.acm.org/doi/pdf/10.1145/33...
- sourceforge.net/projects/smoo...
- Source Code: github.com/tsoding/SmoothLife
Support:
- BTC: bc1qj820dmeazpeq5pjn89mlh9lhws7ghs9v34x9v9

[@noctavel]

RealLife paper is actually pretty consistent with real life. nobody understands shit

[@miran248]

So you're saying those who understand it, now understand life?

[@npas76]

@@miran248he did not say that at any point in that comment

[@lloydbush]

Give me the area of a circle: - Mathmatician: π*r^2 - Engineer: 3*r^2 - Tsoding: just sum all the points of an arbitrary interval

[@maksasj]

1:07:55 In 2 years I will have opportunity to write proper scientific paper for my bachelor degree, and I swear I will use one of yours projects as a reference.

[@korigamik]

Challenge accepted

[@garethma7734]

52:30 I mean, the definition of M here is actually the correct one (instead of the usual "pi R^2" or whatever) since you want to clamp the result to [0, 1]. Random mathematician knowledge dump, computing the exact value of M given just R is actually quite hard, especially mathematically. It's called the Gauss circle problem and is one of those hard number theory problems

[@futurisold]

yes, please do more paper series!

[@valovanonym]

Paper reviews would be a cool series 👀

[@BirdbrainEngineer]

When I was working through the SmoothLife paper, I found it a little bit wonky myself as well, especially as I had actually worked through the Lenia paper first which in my opinion explained things much better. Ontop of that, the mathematical way of explaining these systems has always been really difficult for me to reason about, and instead, for me it's easier to think of them in the programmer way, which is to say operation by operation. Your simulations are running correctly, but the parameters need fine-tuning; both in SmoothLife and in Lenia the "interesting" parameter sets are fairly chaotic and require fine tuning of the parameters to at least 2 decimal places, usually more like to 3 decimal places.

[@MrYefh]

I love you, please keep doing streams about simulating life-like systems

[@FacePalmProduxtnsFPP]

The tip on staying in boundaries of arrays using Euclidean mod was super helpful. Thank you!

[@michalbotor]

when you calculate sigma1(x=n, a) and sigma1(x=n, b) i would use alphan, but when you calculate sigma1(x=m, 0.5) i would use alpham.

[@JonnyDeRico]

who doesn't move in the direction of their thickest part?

[@lovely-shrubbery8578]

You calling me fat 😭

[@nomadvagabond1263]

Thats why i always find myself sitting😏

[@davidaloysparrow219]

Love these streams!

[@drdca8263]

Something I’ve wondered for a while about systems like SmoothLife and Lenia, is that, these systems are translation invariant, they have a continuous translation symmetry, but, it seems like, because they aren’t described by Hamiltonian mechanics, that doesn’t result in them having conservation of momentum (or of angular momentum). It would be interesting, I think, if they could be modified in a way to make them fit into the framework of Hamiltonian or Lagrangian mechanics, in order to get a version that does have conservation of momentum (and if angular momentum). That would be cool, I think! But, I’m not sure how that would be best done... I guess for one thing, would need to have a continuous time version of the system in order to be able to put it into those frameworks. I think there are some theories of how to take a system that doesn’t quite obey Hamiltonian mechanics due to losing energy (e.g. because of friction with an external environment not considered part of the system), and come up with like, a Hamiltonian system that has a pair of interacting systems, one of which is the one one wants to model? Or something like that? Idk. Hm, ok, but, suppose we first are just looking at the state space for smoothlife, and we like, want to describe what sort of thing momentum should be, and well, it should be the generator of translation symmetry.. I guess we have a field which describes the “aliveness” at each point, and you could talk about, for each point, an operator that takes a state and returns the aliveness at that point in that state. and, the translation symmetry would be like, an family of automorphisms of the algebra of those kinds of operators, and I guess, the derivation of the function at a point in a given direction, would be the momentum at that point in that direction? ... maybe the problem is just that the total momentum is always zero, and is therefore trivially conserved? Well, assuming that the aliveness far away is zero in all directions. If that’s why, then maybe having complex values like in quantum mechanics would help? Or... Maybe it is due to like, stuff not being linear enough? Idk

[@bartpelle3460]

I'm pretty sure I understood nothing

[@koktszfung]

Is it similar to Hamiltonian in fluid mechanics?

[@drdca8263]

@@koktszfung Is what? Hamiltonians in general? Yes. Hamiltonians are Hamiltonians. They do the same kind of thing. SmoothLife is, I think, not described by a Hamiltonian, but I was hoping that maybe a small modification to it could be made to make it so it could be described by a Hamiltonian (or a Lagrangian) . But I’m not really sure how one would go about making such a change. If one could make such a change, the result would, I think, have conservation of momentum, conservation of angular momentum, and conservation of energy, which would be cool I think.

[@koktszfung]

@@drdca8263 sorry I realised what I said makes no sense, I mean maybe you could treat the aliveness as a scalar field and write the Hamiltonian of the field

[@drdca8263]

@@koktszfung Oh, ok! Makes sense then. That sounds like it could work, but I’m unsure about how specifically that would be done. Thanks for the pointer though. I had been mostly thinking about Lagrangians (even though I might have said Hamiltonian), and maybe thinking about it that way could be clearer? I will look some stuff up

[@GTS00000]

paper implementation, yes, more please

[@Jack-sy6di]

You're not dumb, the SmoothLife paper is just not very well written and clearly not written by a pure mathematician (in the author's defense though it looks like they probably just kinda threw it together, it's probably not the best representation of their writing abilities). It would actually be clearer if it were written MORE in the style of a pure math paper. I'm also not sure what they're talking about with alpha_n and alpha_m, but the whole set of formulas they use for the transition formula is basically just fuzzy boolean logic. sigma_1(x, a) is basically 1 when x > a and 0 when x < a, except it doesn't actually swap from 0 to 1 instantly, it fades continuously between them. It's like a fuzzy x > a. If you read the other formulas and pretend that sigma_1(x, a) is just boolean x > a, it makes sense. sigma_2(x, a, b) is 1 if a < x < b and 0 otherwise. sigma_m(x, y, m) selects either x or y depending on if m > 1/2. And finally in s(n, m) we're basically saying that if m > 1/2 let I = [b1, b2] (the birth interval) and if m < 1/2 let I = [d1, d2]; and then just return 1 if n is in I and 0 otherwise. So alpha controls the "fuzziness" of the fuzzy boolean logic operators, higher alpha meaning more fuzzy. In total though, the function s(n, m) involves four different calls to sigma_1 and you could in principle use a different fuzziness for all of them, so if we want we could even have alpha1, alpha2, alpha3 and alpha4. My best guess is that they want you to use alpha_n when you're making a call that "involves" n and alpha_m when it "involves" m, so in practice that would mean sigma_m uses alpha_m and sigma_2 uses alpha_n. (maybe you say all this in the stream, I'm kinda skipping around)

[@adrian66696]

1:04:08 alpha_m and alpha_n are just the respective alpha for when you compute eq. 3 for either m or n. By looking at eq.3, you just need to set alpha to alpha_m when x is m, and set it to alpha_n when x is n. s(n, m) calls sigma_2 with x = n, which then calls sigma with x=n. s(n, m) also calls sigma_m which calls sigma_1 with x=m. I think the idea is that you want to smooth the edges of the circles since you have aliasing, for the smaller circle you want more smoothing, and for the bigger one you don't need as much. EDIT: he points it out in the next live

[@ghostroll]

brilliant video as always. side note: for emod you don't have to do the first modulo, you can just do { return (a+b)%b}. works for both negative and positive cases.

[@SlimyDash]

this doesnt work for a < (-b) unfortunately

[@johnpavel1913]

The different alphas are used to smooth in the inner and outer rings

[@afmikasenpai]

54:37 damn this is so relatable, that was me a few months ago. I hate how 'easy' yet so damn confusing this paper looks xD

[@omaramo190683]

another fucking great tsoding session

[@davideferrara6274]

Awesome stream!

[@spacewad8745]

yoyoyoyoyo. fking brilliant as always

[@tdoc666___]

tson actualy showed us how complicated life is, and how much of engineering you actually need to create life, something weird actually bumped into my head, could actually life have been engineered by an intelligent engineer or all this complexity came by itself, bu if it came by itself, then how can it know all this formulas by itself, you know this kind of things is really interesting.. what do u think?

[@taileebao]

I think your coding tool is VIM and you use it very well through keymaps, so can you make a video about it? Thank you

[@OpenKeith]

Actually, they use Emacs. If you want to try it out, I'd personally recommend checking out the Doom Emacs configuration framework (it comes with a lot of useful modules and defaults, and even has a Vim-style keymap mode)

[@Golimko]

1:24:49 in the paper when updating the current cell state they multiply S[s(n, m)] by current value f(x, t), and you don't. maybe this is the problem

[@angelomarano8458]

thank you!

[@mertunsal1389]

Anybody understands what kind of terminal he's using?

[@mellowyellow7523]

drinking game, take a sip every time he says right

[@keremardicli4013]

Holy moly

[@xanhx]

so cool

[@andreypopov6166]

а я только кнопочки по экрану двигать могу...Спасибо за инспирейшен!

[@gabef9538]

Next, train a neural net for generating the next frame of a microscope.

[@DaRza17]

I would love to see that being done in Pygame.

[@ConnorMcCormick]

damn this guy is good at vim

[@BlueDippy]

Bro you are cracked

[@Paxsali]

I like stimulating life on your computer.

[@_orangutan]

I understand some but not all. Maybe if I dedicate a week to it, I'll understand it.

[@piotao]

The problems with the paper explain WHY it is posted on arxiv - it will never pass the rigid review required for good scientific papers. So many unexplained things left.

[@ammarahmad4872]

wow

[@doxnit4343]

24:00 Not impressed. Almost all editors have substitution with regex. BUT Vim has Vim Disel. Does Emacs have its own Luke Smith???? Checkmate

[@Neuer_Alias_erstellen]

wow a real Tsientist

[@bizarrapmusic]

30:04 lmaoooo i can't hahahha

[@douglasgabriel99]

I'm scared

[@user-un1fq7uz7z]

Очень классное видео) хороший английский прям, не понял ваще, что ты русский

[@TurtleKwitty]

vim can do it /better/ actually ;3

[@fxs2008]

Instead of ((x%100) + 100)%100 you could do simpler: (x + 100)%100, it does not break negative numbers.

[@sirynka]

Unless you'd go below -100 His impl would work for -250 your wouldn't

[@Rookie_AI]

hey, are you drunk when you're live,

[@dysb]

God, I wish you wrote Rust. Watching a skilled someone tackle a relatively complex problem is a great way to learn a language. I have zero desire to deepen my knowledge of C when there's more modern alternatives that, crucially, compile to WASM. Why not Rust?

[@albusdandelion719]

C is more fun

[@dandreani]

just translate that to rust... you will learn rust and c at the same time

[@niculaelaurentiu1201]

He has Rust streams that you can watch

[@andrewgr144]

Really? No one's going to comment on the 6.7GiB Porn Folder self-disclosure? Are you all actually paying attention to the content, rather than looking for opportunities to be snarky? Are you even trying?

[@evasiveutopian]

Why r_a? Answer: because the shape of the "2d donut" (r_a - r_i) is called an "annulus". See en.wikipedia.org/wiki/Annulus_(mathematics) (I am guessing)