An Invitation to Andrej Karpathy — AI-Assisted Scientific Computing at Scale
3.5M lines of Rust, zero human-written code, 13+ months of AI-assisted development. The largest existence proof that conversational programming works at production scale.
This is a standing invitation. A human reads and responds to every message at eco.primal@pm.me.
The Existence Proof
3,598,358 lines of Rust. 135,000+ tests. 74K lines of GPU shader code. 15 composable programs. 9 scientific validation domains. 175+ published papers reproduced computationally.
Zero human-written code.
The human is a microbiologist with a data science degree who chose Rust because they didn’t know it — forcing every interaction to stay in conversation with AI assistants. 13+ months. 3-6 machines running parallel AI conversations. Every line emerged from conversation.
This is not a demo. It’s a production scientific computing ecosystem that runs lattice QCD, GPU-accelerated DADA2 bioinformatics, protein structure prediction, pharmacometrics, and molecular dynamics on commodity hardware.
Why This Matters for AI
You’ve talked about AI-assisted coding as a paradigm shift. This is data:
| Metric | Value |
|---|---|
| Total Rust LOC | 3,598,358 |
| Total tests | 135,000+ |
| GPU shaders (WGSL) | 952 |
| Human-written code | 0 lines |
| Duration | 13+ months |
| Scientific papers reproduced | 175+ |
| Validation scenarios | 197 |
| Known debt items | 2 |
The methodology — K-NOME Programming (Knowledge-Numeric Orchestrated Mentoring Ecosystem) — treats the human as mentor and the AI as implementer. The human never touches the codebase directly. The conversation constraint is structural: intent flows one direction, implementation flows the other, and the friction between them produces software that neither could produce alone.
The Neural Architecture Theorem
neuralSpring proves the Isomorphism Theorem: all neural architectures decompose into 6 primitives (GEMM, Attention, Normalization, Nonlinearity, Reduction, Gating). 83.6× faster than Python/NumPy for core operations. Implemented in pure Rust + WGSL — no PyTorch, no TensorFlow, no ONNX runtime.
This is the kind of first-principles decomposition you’ve advocated for in your lectures: understanding neural networks by building them from scratch, not by importing libraries.
What The Stack Actually Does
Not a framework. Not a library. A sovereign operating environment:
- GPU-accelerated DADA2 — 16S bioinformatics without Galaxy or CUDA
- Lattice QCD on consumer GPUs — gauge theory without CUDA, HPC, or vendor SDKs
- Cross-vendor f64 GPU compute — WGSL shaders on NVIDIA, AMD, Intel through Vulkan
- Sovereign mesh networking — 353× LAN throughput over WireGuard
- Reproducible science — guideStone verification class: binaries that prove their own correctness
All pure Rust. All AGPL-3.0. All running on $15K of consumer hardware in a basement.
The Conversation
The methodology is documented, the code is public, the evidence is reproducible. If AI-assisted development at this scale is interesting to you, the full record exists:
- K-NOME Programming — the conversational method
- The Prompt Bank — real prompts from 13 months of development
- I Don’t Know Rust — how a microbiologist built 135,000+ tests through conversation
Every commit is co-authored (Co-authored-by: Cursor). The agentic development is fully transparent.
The proof of work is the work itself. The conversation is the method.