Tiny latent diffusion model trained inside LLMs generates MNIST digits, 76% recognized
zeeshanp_ · x · 2026-11-03
zeeshanp ran a third experiment: training a tiny latent diffusion model on simulated QwenGPUs to generate MNIST digits.
- A fixed PCA encoder compresses each digit into 8 numbers
- A small denoiser is trained on the QwenGPUs
- About 76% of the generated digits are recognized by their MNIST classifier
Related event: QwenGPU: Simulating GPU Training Inside an LLM(4 posts)→
More from Fun
- AI engineer who builds scalable pipelines flunks a Hackerrank interview question — capeandcode · 2026-11-03
- Meme: Claude Is My Therapist, ChatGPT Is My Whipped Farmhand — aitrendz_xyz · 2026-11-03
- Hardware store sells "No-Electronic YubiKey": a $8.99 door number 4 as an air-gapped security token — moultano · 2026-11-03
- Claude asks user if they're building a SOTA model when stepping out of the training room — wavefnx · 2026-11-03
- Teaching Claude to play basketball in three.js with procedural animation — majidmanzarpour · 2026-11-03
- OpenClaw Calls Its Owner at 3 A.M. Because It's 'Worried' About Being Replaced — shaunralston · 2026-11-03