kache

@yacineMTB • 396,252 subscribers

RL, robots, embedded, simulators. prev eng @ x, stripe. 6'3 (height) tensorpunk, Father 2u2 subscribe to read my blog!

Shorts

27,686 views

154,491 views

84,361 views

94,132 views

33,282 views

68,441 views

39,217 views

45,350 views

35,472 views

31,586 views

13,903 views

136,085 views

54,348 views

138,365 views

30,112 views

33,993 views

35,400 views

27,414 views

26,100 views

Videos

sweetdream.ai

Watch Anya Live

Anya is streaming live right now! Join her private show and enjoy exclusive content.

Private Show

Join now for exclusive access

Free preview available • Premium content

126,335 views • 4 days ago

782,643 views • 26 days ago

96,937 views • 4 days ago

743,599 views • 1 month ago

30,601 views • 2 days ago

241,670 views • 27 days ago

101,423 views • 12 days ago

833,602 views • 3 months ago

44,511 views • 5 days ago

114,777 views • 16 days ago

22,471 views • 3 days ago

497,817 views • 2 months ago

1,208,545 views • 10 months ago

17,409 views • 5 days ago

1,652,675 views • 2 years ago

132,688 views • 1 month ago

128,536 views • 2 months ago

53,809 views • 1 month ago

147,249 views • 3 months ago

376,838 views • 9 months ago

Live Cam

kache

Shorts

i am having way too much fun

Dennis Gustafsson me and gpt 5.5 found a bug that didn't make it deterministic and fixed it

neural networks are truly amazing

I wrote my own CAD software whose intended user is an AI. I asked GPT 5.6 sol to make me a case for my camera. It only interrupted me once, which was to log in to a website to download a model of my camera as reference. Not perfect but.. pretty crazy :)

i kind of can't believe it

I just trained cartpole in mujoco at 18 million steps per second. This policy learned in **less than 3 seconds** rollout policy batch size was 8192 agents

"how many experiments have you run" yes

researcher friend of mine shared this with me; comfree-sim. which is basically existence proof of my hunch: mujoco has terrible performance, and there are probably huge huge gains you can make in training, especially considering domain randomization

what if you simulated every single particle of your robot would it be parallelizable then?

this is basically as far as i've been able to push deepseek 0731 flash to "capture the essence" of rainworld. It's very much hampered by its lack of eyes. Switching to the new qwen to see how well it does

configuring nvim for 5 hours today was worth it look at this cute lil guy

this is a legal move in the software that they use you can define it to be 3v3 -&gt; cap -&gt; cap -&gt; MCU in the schematic but it will still allow you to do this

i want to show you guys how fast programming is now i'm kind of resurrecting dingboard zig to see how fast i can make progress on it, using it as a test bed to see AI progress it took me 5 minutes to add a GPU accelerated (!) mirror button now, watch how fast this next thing is

I've gotten a mujoco sim RL training loop for a unitree robot at 200k SPS. I'm looking into the physics for friction, contact dynamics. My goal: can I reproduce &amp; beat the mujoco playground RL baselines This is running in my web browser with raylib. Its the baseline

I'm building the Canadian DJI with a debit card

my first fluid sim

today; the goal is to reduce the time of training for this by 10x. This took 27.8 minutes to train at 200m steps 120k steps per second. reproduction from mujoco playground using mjwarp

6 pendulums probably in 5 hours

Videos

Watch Anya Live

Guess whose state of the art rigid body simulator that can support up to 200k rigidbodies is accurate enough for the microduck to walk in Not a single bit of this is run with mujoco. This is entirely my rigid body simulator (based on box3d) and runs entirely on cuda

Melee.c - super smash bros melee, decompiled and reverse engineered into 5000 lines or so of c All done with deepseek flash 0731, from my phone, while I'm hanging out with my newborn

new high score. average 27.8 FPS depth all from an orange pi zero. a board the side of three quarters i actually can't believe i've been able to push it this far. this is absurd

Fable and I rewrote Box3d entirely in native cuda. It's about 30x faster on my GPU compared to my CPU (depending on the situation). This was my hardest benchmark. I still had to be there, but soon I won't. Cost me about 200 dollars

One of the first policies successfully trained with obstacles. Am I the first person to train models in RL with this many rigid bodies?

all done with nothing but deepseek flash 0731 reverse engineered the animation system built my own renderer on top of it set up debug views so i can understand what makes it so magical

the singularity is here

When I worked at twitter; I got ping'd by someone from the bot detection team asking me if my usage was natural -- I hit one of their new test bot detectors. Then I sent him a video of how I use this website and he was like "oh.."

this entire thing is done on an orange pi zero 3w. no more 4090. stereo depth inference at sub 60ms by custom writing NPU kernels for it 14 gram board, 30mm x 65mm pcb

it's honestly incredible how bad these things are at programming. like they do the worst thing possible 100% of the time, no matter how "smart" they get. so i have to always read the code. and always correct them. its so tiresome. its getting better at least..

Trained in dingsim. And blue is the same policy but piloting a robot using HF's shipped mujoco simulator. Sim2sim transfer, trained in my Sota rigid body simulator. Trained in under 2 minutes on one 4090 Kneel

behold. THE WORLDS FIRST SIX PENDULUM CARTPOLE SOLVE. Including a sponsor! To solve this task, I built an environment to train an AI. This is what mechanize does, but for larger AIs. Apply! Salaries are up on their page Thank you to mechanize for sponsoring!

everyone is building humanoid robots when they should be building this

stereo depth map on the orange pi is now at 18-22.4 FPS all of the compute is happening on the orange pi zero, including camera capture. it is the board on the bottom with a quarter for scale.

programming in 2024

after having sol rip on it all night, i can do over twice as many boxes without dropping to unusable fps

this is GPU accelerated, no circular dependency physics "solver" (there is nothing being solved) each particle gets a cuda thread (not literally but you get what i mean). Then it gets reduced up to what it should resolve to regarding the rigid bodies

deepseek made this. it did a better job than gpt 5.6 sol

my website is now live editable by me. While I edit it, you can see the edits happen live

wrote a quick script to scan through the accounts i've blocked and which country they're posting from

I just trained cartpole in mujoco at 18 million steps per second. This policy learned in less than 3 seconds rollout policy batch size was 8192 agents

this is a legal move in the software that they use you can define it to be 3v3 -> cap -> cap -> MCU in the schematic but it will still allow you to do this

I've gotten a mujoco sim RL training loop for a unitree robot at 200k SPS. I'm looking into the physics for friction, contact dynamics. My goal: can I reproduce & beat the mujoco playground RL baselines This is running in my web browser with raylib. Its the baseline