Woman Bakes DeepSeek R1 From Scratch

PURE RL DEEPSEEK R1 NVIDIA G... GPT-4o

A woman humorously prepares an AI model like homemade bread in a kitchen. She adds labeled ingredients such as pure reinforcement learning, DeepSeek R1, NVIDIA GPUs, and GPT-4o, then kneads and bakes the result. The joke satirizes AI training jargon and DeepSeek’s claims of achieving strong reasoning with fewer resources and little human feedback.

My husband was complaining about how expensive and processed the big box AI models we're getting so I decided to make him one from scratch. Today we're doing something completely different. We're going all natural using pure reinforcement learning. While the others drown their models in process fine-tuning and human feedback at every step, we discovered a simpler free-range approach. Now for the chips. I had some old GPUs lying around but somehow got my hands on a few high-end GPUs. This recipe usually calls for a lot more but I was able to make it just as good with a lot less. Now between you and me, we borrowed a little starter from another model, but the secret to this recipe is in how we mix it. Instead of relying on humans to manually label every answer, we're letting our model figure it out on its own. Now see these little bubbles? That's our model developing its own chain of thought. It's checking its own work, catching mistakes, even having its own little aha moments. You might call it emergent reasoning. Everyone thought you needed mountains of GPUs and armies of human labellers to make good AI. But turns out, you just need to let it learn through pure trial and error. no processed human feedback, no artificial guidance, just pure learning from experience. And what's even better is that this whole recipe only costs 6 million to make, compared to the hundreds of millions it costs to create those big box models. Unlike the others, I'm not gatekeeping. I'm making it all open source. So comment recipe and I'll send you the full method, weights and all.

• @jemalloc / 1843 2025-02-04