kache

@yacineMTB • 330,319 subscribers

reinforcement learning, robots. prev eng @ x, stripe. 6'3 (height) tensorpunk subscribe to read my blog!

Shorts

92,444 views

152,633 views

36,148 views

68,441 views

45,350 views

31,586 views

35,284 views

29,728 views

27,376 views

26,064 views

33,993 views

35,400 views

19,248 views

136,084 views

138,298 views

16,064 views

60,094 views

85,568 views

11,734 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

741,183 views • 14 days ago

131,596 views • 10 days ago

832,946 views • 1 month ago

494,261 views • 1 month ago

128,536 views • 1 month ago

1,208,266 views • 9 months ago

91,129 views • 1 month ago

147,188 views • 1 month ago

1,652,675 views • 1 year ago

75,355 views • 29 days ago

87,281 views • 1 month ago

376,789 views • 7 months ago

60,492 views • 1 month ago

55,933 views • 1 month ago

47,586 views • 1 month ago

38,927 views • 1 month ago

208,582 views • 9 months ago

157,742 views • 8 months ago

180,686 views • 1 year ago

41,239 views • 2 months ago

Live Cam

kache

Shorts

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 :)

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

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

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

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

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

I'm building the Canadian DJI with a debit card

my first fluid sim

Fixed a numerics bug on the training loop, and got the training down from 52 minutes to 45 minutes Only 15 minutes away from beating the deepmind mujoco playground baseline!

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

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

live on my website again. i'm designing the internet condom. Trying to understand the difference of search and recommendations, figuring out what an MVP can look like, trying to design something that can actually beat all social media networks, something that would destroy them

right now my reinforcement learning model is basically doing this and i came up with a solution for it all by myself (by stealing it from pufferlib) that not only solves it, but also helps keep training stable and fast

just added skins. do you guys see this?

well it's getting one of them up..

Videos

Watch Anya Live

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

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

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.."

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!

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

everyone is building humanoid robots when they should be building this

scales great! each rectangle here is a set of discs. running on cuda. now to painfully go through all of the weird collision bugs and figure out why they're happening and why it isn't settling

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

programming in 2024

I think I'm going to become a game dev

2023, 3 pendulums 2025, 4 pendulums 2026:

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

I AM SO SO CLOSE TO SOLVING 6 PENDULUM CARTPOLE just added erection reward and it's almost got it but this task is really really hard. I've been pouring compute at it I'm not leaving my computer seat until I have this solved. it's time to bust out curriculum learning

i love this stuff so much

i wonder when this task becomes impossible to solve

I'm launching dingbotics: speeding up RL for robotics go to dingbotics dot com to check out the latest model I've trained. this will work on your phone's web browser built with mujoco, nvidia warp, pufferlib, raylib &amp; Rich the Kid

your product is 3D printed???????? you didn't even bother to get CARBON FIBRE END CAPS ordered from SEND CUT SEND (3 day delivery)?? for my hardware startup, i've been using SEND CUT SEND, this month's sponsor of my profile. kudos to them for helping support our development!

dingcad is now open source needed cad software for my startup, so i made it i'll keep it pushed with library and updates. i push straight to master. there is a file called scene.js that will live reload the screen every time you change it. it's openscad without the bloat

pretty insane that it's been nearly 2 years and no one has made anything nearly as good as this

my second fluid sim

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

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

I'm launching dingbotics: speeding up RL for robotics go to dingbotics dot com to check out the latest model I've trained. this will work on your phone's web browser built with mujoco, nvidia warp, pufferlib, raylib & Rich the Kid