Two-Person Lab Trains Video AI for $21K

🔥 What happened A two-person research lab trained Pyramid-JiT, a video model built entirely from scratch, for just $21K. The trick: instead of relying on a separate VAE for compression, the diffusion model learns its own latent space directly in pixel space, achieving 32×32 token reduction versus the usual 16×16. 💡 Why it matters The model converges with 11× fewer samples than its predecessor and halves the attention context window – a 512² image needs only 512 tokens instead of 1,024. Since attention scales quadratically, every token saved is real money. This is the cost bottleneck that makes today's video models impractical for iterative creative work. ⚡ Our take If two people with $21K can train a competitive model, the big labs have an efficiency problem – or we're at a real inflection point. Either way, the incumbents should be nervous.