OpenLabs.live

Backpropagation, built by hand · The AI builds 8 cards from one sentence; the autograd and numeric-check branches merge at the comparison card. Play writes that card out like a board, ▶ runs its code in a cloud environment (loss curve and both gradients open beside it), and a pen circle and a typed note mark the result on the card.
KV Cache 到底省在哪 · AI 从一句话建出 8 张卡，显存与速度两条支线在「量一次」汇合。播放这张实验卡，它像板书一样写出来；点 ▶ 在云端环境里真跑代码，每步耗时曲线出在卡片旁边；最后用笔圈出结果、写一句批注。

$ srun --gres=gpu:h100 srun: job 48213 queued and waiting for resources

H100 80 GBRTX 4090 24 GB
$ pip install flash-attn ERROR: Failed building wheel for flash-attn

flash_attn 2.8.3 + FA3mamba_ssm 2.3.1gsplat · PyTorch3D sm_89Qwen3 · GPT-2 · SAEs
$ ls ~/notebooks Untitled.ipynb Untitled(1).ipynb final_v2_REAL.ipynb

1,517

01


  
02
Two Python code blocks on the board, each showing run-by-agent timing, the python-cpu badge and a returned chart

  
03
$ python ~/sim-demos/diamond_dream.py --game Breakout --horizon 40
world-models-diamond · RTX 4090 · machine ready in 6.6 s
[ 4.3s] real game: 300 steps in episode 0
[ 5.4s] open loop: 40 frames, 26.4 ms/frame, 3 denoising steps
[ 8.2s] dream play: 80 imagined steps
DREAM_OK game=Breakout total 8.6s
16.3 s in the sandbox · 2 figures returned · billed $0.04
Five pairs of Breakout frames, real beside dreamed, at steps 1, 10, 20, 30 and 40

 






 

 

 

 
$20

 
    
  
  •  
    • 
  
  •  
    • 
  
  •  
    • 
  
  •  
    • 
 

 






 

  

  
# proposed by the agent, confirmed by you (measured: installs in 13 s)
name      MPI basics (Open MPI + mpi4py, CPU)
base      python-cpu
apt       build-essential libopenmpi-dev openmpi-bin
pip       mpi4py
status    ready · bound to tb:…/E4

 






 

  

  
    
   
  •  
    • 
   
  •  
    • 
   
  •  
    • 
  

 






 

 

 

 
 
Johns Hopkins logo: Wikimedia Commons via HandWiki, CC BY-SA 3.0, unmodified. Source · License · All asset sources