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:

MetricValue
Total Rust LOC3,598,358
Total tests135,000+
GPU shaders (WGSL)952
Human-written code0 lines
Duration13+ months
Scientific papers reproduced175+
Validation scenarios197
Known debt items2

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:

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:

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.