feat(misc): Neural net in 30 lines of K

[jamesrziggy]

Summary

A cookbook demonstrating Claude generating, explaining, and self-repairing neural network code in Kona (K3) — Arthur Whitney's array programming language.

• Claude writes a complete XOR neural net (forward pass + backprop + training) in ~30 lines of K
• If the generated code errors, a self-repair loop feeds the Kona error back to Claude for iterative fixing
• Includes a hand-tested reference implementation verified on Kona (converges to >99% accuracy on XOR)

Why this is interesting

K has virtually zero presence in LLM training data compared to Python. The notation is so dense that W1 _dot\: xi replaces what would be a nested for-loop in most languages. This is a genuine stress test of Claude's ability to reason about unfamiliar, terse array languages — not just parrot Python.

Verified on hardware

The reference K code was tested and confirmed working on Kona running on Android/Termux (ARM64). Output:

XOR predictions:
0 0.0 -> ,0.003721189
0 1.0 -> ,0.9952888
1 0.0 -> ,0.9952321
1 1.0 -> ,0.006605917

Fun fact

This PR was written entirely from an Android phone running Termux + Claude Code. The K code was debugged live in a Kona REPL on the same phone. No laptop involved.

🤙 Second PR from the phone. First one was anthropics/courses#142.

🤖 Generated with Claude Code

[nidhishgajjar]

Orb Code Review (powered by GLM 5.1 on Orb Cloud)

New cookbook: Neural Net Inference in 30 Lines of K. A creative exploration of Claude's ability to generate code in the K array programming language.

Observations

1. Creative and educational (Positive)
This is a unique cookbook that tests Claude's capabilities in a niche language with sparse training data. The approach — provide a reference card, generate code, explain it, validate — is well-structured.

2. The K reference card pattern (Positive)
Providing a domain-specific reference card in the prompt is a great technique for getting quality output from Claude in specialized domains. This pattern is transferable to other niche languages/frameworks.

3. Subprocess execution of generated code (Medium)

result = subprocess.run(["k", k_file], capture_output=True, text=True, timeout=30)

Running AI-generated code via subprocess has inherent risks. While K is a relatively safe language (no filesystem access by default), the notebook should note this consideration. The 30-second timeout is a good safety measure.

4. F-string syntax issue (Bug)

messages=[{{"role": "user", "content": generation_prompt}}]

Double braces {{ inside an f-string are correct for literal braces, so this is fine. Though the readability is slightly affected.

Summary

Fun and educational cookbook that demonstrates both Claude's capabilities and the reference-card prompting technique. The K language angle makes it memorable.

Assessment: approve