The kid wants to work at Anthropic. So he built a GPT-2-scale model in pure C and CUDA. From scratch. After they banned his previous project. That's the kind of spite that makes great engineering.
The Fable That Started It All
Anthropic, the AI safety darling, banned a project called Fable earlier this year. The details are murky — something about violating usage policies — but the creator, who goes by the handle nanoeuler, took it personally. Instead of whining on Twitter, he wrote a transformer from the ground up. No PyTorch. No TensorFlow. Just raw CUDA kernels and a C harness.
Why C? Because he wanted to understand every damn line. 'Interfacing with an LLM does not mean understanding how they are composed,' he said. That's the kind of line that makes old-school systems programmers nod slowly.
What's Under the Hood
NanoEuler isn't just another 'I trained a model on my laptop' flex. It's a full GPT-2-scale implementation — 1.5 billion parameters if you push it — with custom CUDA kernels for attention, feed-forward, and layer norm. He wrote the tokenizer, the batching logic, the whole pipeline. In C. With CUDA. On a single consumer GPU.
I've seen the code. It's clean, opinionated, and ruthlessly commented. There's no bloat. No over-engineering. Just the math, the memory, and the metal.
Why This Matters
We're drowning in abstractions. Every kid with a laptop can load a Hugging Face model and call themselves an AI engineer. But when the memory leaks, when the inference stalls, when the whole thing falls apart — they're lost. NanoEuler is a counterweight. A reminder that the stack isn't magic. It's just C, CUDA, and a lot of linear algebra.
Anthropic might have banned his toy. But they can't ban the knowledge he acquired building this. That's the real victory.
What's Next?
NanoEuler is open source. The repo is bare-bones — documentation is sparse, and you'll need to know CUDA to get it running. But that's the point. It's not for tourists. It's for people who want to get their hands dirty.
Will Anthropic hire him now? Who knows. But if I were Dario Amodei, I'd be watching. Because a coder who can build a GPT-2 from scratch in C is worth ten prompt engineers with PhDs.
Spite is a hell of a motivator. And sometimes, it produces beautiful code.
