Hi fans ๐
been diving into continual learning lately and honestly... it's fascinating ๐ง
basically the problem is โ how do you teach a model new stuff without it forgetting everything it already knows? turns out it's really hard lol
found this paper on sparse memory finetuning and the approach is kinda crazy. instead of updating the whole model (which wrecks it), you give it millions of tiny memory slots and only touch the relevant ones. like... surgical precision ๐ฌ
the numbers: full finetuning forgets 89%. LoRA still forgets 71%. this method? only 11%. that's wild
still early research (1.3B scale) but the direction feels right. this is where always-learning AI actually starts to look possible
also I started google's nested learning path a while back but never finished it... getting back to that too ๐
https://jessylin.com/2025/10/20/continual-learning/
been diving into continual learning lately and honestly... it's fascinating ๐ง
basically the problem is โ how do you teach a model new stuff without it forgetting everything it already knows? turns out it's really hard lol
found this paper on sparse memory finetuning and the approach is kinda crazy. instead of updating the whole model (which wrecks it), you give it millions of tiny memory slots and only touch the relevant ones. like... surgical precision ๐ฌ
the numbers: full finetuning forgets 89%. LoRA still forgets 71%. this method? only 11%. that's wild
still early research (1.3B scale) but the direction feels right. this is where always-learning AI actually starts to look possible
also I started google's nested learning path a while back but never finished it... getting back to that too ๐
https://jessylin.com/2025/10/20/continual-learning/
so... I've been working on something ๐
building an AI learning playground โ basically a place where you can actually learn AI by doing, not just watching tutorials and reading papers all day
especially heavy on reinforcement learning ๐ฎ โ training agents, reward shaping, policy stuff... RL is one of those things that clicks 10x faster when you see an agent actually learn in real time vs staring at equations
it's still super early stages but I'll be sharing progress here regularly. this channel is basically the build log now ๐
oh and โ if there's a topic you want me to cover or break down, drop it in the comments or slide into my DMs ๐ฌ I'm building this for y'all so lmk what you actually want to learn
I will include things I want to learn meaning I will be building them while learning that thing
stay tune
building an AI learning playground โ basically a place where you can actually learn AI by doing, not just watching tutorials and reading papers all day
especially heavy on reinforcement learning ๐ฎ โ training agents, reward shaping, policy stuff... RL is one of those things that clicks 10x faster when you see an agent actually learn in real time vs staring at equations
it's still super early stages but I'll be sharing progress here regularly. this channel is basically the build log now ๐
oh and โ if there's a topic you want me to cover or break down, drop it in the comments or slide into my DMs ๐ฌ I'm building this for y'all so lmk what you actually want to learn
I will include things I want to learn meaning I will be building them while learning that thing
stay tune
๐ฅ4
I will be using this kinds of templates whenever I post -> I kinda like them ... maybe we could have another persona than the hacker like this one, maybe ... like minimalistic hand drawings ...
๐2
Me wasting my time ๐ nevermind ๐
Was thinking about what to use for post images on this channel.
Then thought โ what if I just build it myself?
A bunch of templates I can pick from, type my text, export.
So that's what I did.
SJ Studio โ a canvas engine that renders post graphics.
8 styles. 20+ templates. All pure Canvas 2D, no design tools.
Terminal, Glass, Glitch, Gradient, Notebook,
Blueprint, Node Graph, Polaroid.
Mostly an internal tool for the channel.
Not trying to build a startup out of it.
But it's open source โ if you want to use it, use it.
Want to contribute a template or style? PRs welcome.
github.com/jajos12/sj-studio -> give it starts if u like it
https://sj-studio-gamma.vercel.app/
I don't really know if it is useful or not frankly ๐
#buildinpublic
Was thinking about what to use for post images on this channel.
Then thought โ what if I just build it myself?
A bunch of templates I can pick from, type my text, export.
So that's what I did.
SJ Studio โ a canvas engine that renders post graphics.
8 styles. 20+ templates. All pure Canvas 2D, no design tools.
Terminal, Glass, Glitch, Gradient, Notebook,
Blueprint, Node Graph, Polaroid.
Mostly an internal tool for the channel.
Not trying to build a startup out of it.
But it's open source โ if you want to use it, use it.
Want to contribute a template or style? PRs welcome.
github.com/jajos12/sj-studio -> give it starts if u like it
https://sj-studio-gamma.vercel.app/
I don't really know if it is useful or not frankly ๐
#buildinpublic
๐1๐ซก1
You know I love game dev but I hadn't pursued it because u can't be all where all at once ๐ (which is exactly what we all devs try to do ๐)-> but at some moment maybe no far I want to vibe code a good game ๐ฒ ...
๐1
sudo jajos
You know I love game dev but I hadn't pursued it because u can't be all where all at once ๐ (which is exactly what we all devs try to do ๐)-> but at some moment maybe no far I want to vibe code a good game ๐ฒ ...
Ofc it wouldn't be tic tac toe or x/o -> I think of like an rpg game maybe or ๐ค some kind of platformers ... also sandbox games maybe ...
๐1
sudo jajos
Photo
Remember when I said I wanted to vibe code a game?
I did it. One session. Start to finish. (I mean we did it ๐, I think u obviously know the other guy ๐ )
โก๏ธ SHIFT โ a dimension-swapping pixel art platformer.
Two parallel worlds โ Lumina (warm) and Umbra (cold).
Platforms exist in one dimension but vanish in the other.
You shift mid-jump to find paths that shouldn't exist.
What's in it:
โ 10 handcrafted levels, tutorial to gauntlet
โ Wall-jumping, coyote time, jump buffering
โ Patrol enemies that are dimension-aware
โ Every single texture generated procedurally โ zero sprites
โ Dynamic audio that crossfades between dimensions
โ Mobile touch controls included
โ Parallax scrolling backgrounds
Stack: Phaser 3 ยท TypeScript ยท Vite
Art: 100% procedural via Canvas API
No Unreal. No Unity. No Godot.
Just a browser and TypeScript.
Not tic-tac-toe. I promised.
โน๏ธWant to play?
https://shift-sage-eight.vercel.app/
If you want the max experience it would be convenient if u use ur pcs
github.com/jajos12/shift -> if you like it give it stars ๐ซ
#gamedev #buildinpublic #phaserjs
I did it. One session. Start to finish. (
โก๏ธ SHIFT โ a dimension-swapping pixel art platformer.
Two parallel worlds โ Lumina (warm) and Umbra (cold).
Platforms exist in one dimension but vanish in the other.
You shift mid-jump to find paths that shouldn't exist.
What's in it:
โ 10 handcrafted levels, tutorial to gauntlet
โ Wall-jumping, coyote time, jump buffering
โ Patrol enemies that are dimension-aware
โ Every single texture generated procedurally โ zero sprites
โ Dynamic audio that crossfades between dimensions
โ Mobile touch controls included
โ Parallax scrolling backgrounds
Stack: Phaser 3 ยท TypeScript ยท Vite
Art: 100% procedural via Canvas API
No Unreal. No Unity. No Godot.
Just a browser and TypeScript.
Not tic-tac-toe. I promised.
โน๏ธWant to play?
https://shift-sage-eight.vercel.app/
If you want the max experience it would be convenient if u use ur pcs
github.com/jajos12/shift -> if you like it give it stars ๐ซ
#gamedev #buildinpublic #phaserjs
shift-sage-eight.vercel.app
SHIFT โ Dimension Platformer
SHIFT โ A dimension-swapping pixel art platformer
๐2๐ฅ1
sudo jajos
Remember when I said I wanted to vibe code a game? I did it. One session. Start to finish. (I mean we did it ๐, I think u obviously know the other guy ๐ ) โก๏ธ SHIFT โ a dimension-swapping pixel art platformer. Two parallel worlds โ Lumina (warm) and Umbraโฆ
I think many of u don't understand the game -> try using shift...
โค1
sudo jajos
Remember when I said I wanted to vibe code a game? I did it. One session. Start to finish. (I mean we did it ๐, I think u obviously know the other guy ๐ ) โก๏ธ SHIFT โ a dimension-swapping pixel art platformer. Two parallel worlds โ Lumina (warm) and Umbraโฆ
If any of u want to understand how the game works - even if u are new to game dev check out the doc: https://github.com/jajos12/shift/blob/master/GAME_DEV_GUIDE.md
GitHub
shift/GAME_DEV_GUIDE.md at master ยท jajos12/shift
Contribute to jajos12/shift development by creating an account on GitHub.
๐ฅ2
Guys just some updates on the ai playground I mentioned back then mnamn ...
-> Now it is live and I am open to comments to improve it, yaw ahun gena betam early stage lay nw yalew ... like I aspire to have 5 tiers ... and almost 70+ modules in total on all tiers ... now we have only 3 tiers available but all of them with guided learning first then fully fledged playground and challenges following ... check it out by urself and discover things ...
github.com/jajos12/aiai
Star the repo if you're interested โ more modules dropping soon โญ๏ธ
#ai #machinelearning #opensource #buildinpublic
-> Now it is live and I am open to comments to improve it, yaw ahun gena betam early stage lay nw yalew ... like I aspire to have 5 tiers ... and almost 70+ modules in total on all tiers ... now we have only 3 tiers available but all of them with guided learning first then fully fledged playground and challenges following ... check it out by urself and discover things ...
github.com/jajos12/aiai
Star the repo if you're interested โ more modules dropping soon โญ๏ธ
#ai #machinelearning #opensource #buildinpublic