IdeaWell

Project ideas from Hacker News discussions.

Ask HN: How to deal with shipping AI-written code you don't understand?

Hacker News Discussion

📝 Discussion Summary (Click to expand)

Theme 1: Trust through comprehensive testing and verification
Many argue rigorous testing enables confidence in AI-generated code, treating passing tests as sufficient validation.

"I have it write a bazillion tests, and I have real, non-trivial programs that it compiled. I have it run a ‘gate’ before releases that takes literally hours on 5 different architectures." - spacedcowboy
"Testing. Lots and lots of testing... Trust; but verify." - ChrisMarshallNY

Theme 2: Shift in engineer role toward oversight and specification
Commenters note engineers are evolving from hands-on coders to managers/specifiers who guide AI rather than implement details.

"Technically it's just a role change. You're now the project manager or something, not the developer." - skeledrew
"Project managers just take the engineers word for it." - jimbokun

Theme 3: Necessity of understanding shipped code for accountability
A strong contingent insists engineers must comprehend what they ship, warning against delegating responsibility to AI.

"knowingly shipping code you don't understand is grossly negligent" - netburst
"you need to understand the code you're merging. If you don't, ownership then lands on whatever AI agent you're using." - willio58

Theme 4: Critique of testing adequacy and AI-generated test quality
Some question whether testing alone suffices, highlighting flaws in AI-produced tests and verification gaps.

"I respectfully disagree. IME writing a good test suite is one of the areas where current LLMs perform worst. They certainly like to write plenty of tests but those tests aren't necessarily adding much value" - Silhouette
"Slop code doesn't stop being slop code when you add in slop tests to account for the sloppiness." - SimianSci

🚀 Project Ideas

Generating project ideas…

Adversarial LLM Code Verifier (ALCV)

Summary

  • Runs two or more LLMs to generate code from the same natural‑language spec, then compares outputs with differential testing, mutation testing, and property‑based checks.
  • Provides a confidence score and highlights divergent sections that need human review.
  • Core value: automated detection of hallucinations and logical errors in AI‑generated code before it reaches production.

Details

Key Value
Target Audience Development teams using LLM‑assisted coding (e.g., Copilot, Cursor, internal agents)
Core Feature Adversarial code generation + automated verification suite (diff, mutation, property‑based)
Tech Stack Python orchestration, Docker sandbox, LLMs via OpenAI/Anthropic APIs, Hypothesis for property testing, mutmut for mutation testing
Difficulty Medium
Monetization Revenue-ready: subscription per developer seat or per‑API‑call tier

Notes

  • Addresses the repeated HN concern that “you can’t trust code you didn’t write” by giving an objective, automated second opinion.
  • Would spark discussion on verification strategies and could be adopted as a CI gate similar to spacedcowboy’s multi‑architecture test gate.

Heterogeneous Auto‑Vectorizer Profiler & Scheduler (HAVPS)

Summary

  • Extends spacedcowboy’s autovectorizing compiler with a runtime profiler that measures latency/compute on each available device (CPU, GPU, APU) and schedules kernels opportunistically.
  • Delivers a drop‑in library/CLI that takes standard C/C++ and transparently accelerates data‑parallel regions without source changes.
  • Core value: automatic performance portability and near‑hand‑tuned speedups on heterogeneous hardware.

Details

Key Value
Target Audience HPC developers, performance engineers, game studios, scientific computing teams
Core Feature Data‑parallel extraction + runtime device characterization + dynamic kernel scheduling
Tech Stack LLVM MLIR front‑end, backends for PTX, SPIR‑V, MSL, WGSL; Rust/C++ core; CUDA/Vulkan/OpenCL abstractions; optional Python bindings
Difficulty High
Monetization Revenue-ready: enterprise license or usage‑based SaaS (pay per accelerated kernel‑hour)

Notes

  • Directly builds on the impressive speedups shown in the thread (e.g., 1312× on mandelbrot) and solves the pain point of “no programming overhead” while still needing trust in the generated kernels.
  • HN commenters interested in the compiler would likely discuss integration, open‑source licensing, and benchmarking workflows.

AI‑Generated Code Understanding Suite (AIGUS)

Summary

  • Takes AI‑written code (or a PR diff) and automatically produces interactive documentation: mermaid diagrams, markdown explanations, quizzes, and a sandboxed simulation harness.
  • Enables developers to “grokk” the code without line‑by‑line review, reducing the trust gap.
  • Core value: faster, higher‑confidence comprehension of AI contributions.

Details

Key Value
Target Audience Engineers reviewing AI‑generated pull requests, tech leads, audit teams
Core Feature Auto‑generated explainers, interactive quizzes, simulation playground
Tech Stack TypeScript/React frontend, Python/FastAPI backend, LLM (e.g., GPT‑4o) for natural‑language summaries, Mermaid.js, JupyterLite for sandbox
Difficulty Medium
Monetization Hobby (open‑source core) with optional paid team collaboration features (private repos, SSO)

Notes

  • Directly answers the HN lament “I don’t understand the code I’m shipping” by turning opaque AI output into teachable artifacts.
  • Would generate lively discussion on the balance between verification and understanding, echoing suggestions like “interview‑based design” and “design‑by‑contract” from the thread.

Formal Verification Guardrails for AI Code (FVG)

Summary

  • Integrates a contract‑first approach into CI: LLMs suggest pre/postconditions and invariants, which are then fed to a formal verifier (Lean 4/Dafny + Z3) to prove correctness or produce counter‑examples.
  • Fails the build if unverifiable contracts are found, forcing explicit clarification.
  • Core value: mathematically grounded assurance that AI‑generated code meets its specification.

Details

Key Value
Target Audience Safety‑critical domains (embedded, finance, aerospace) and teams requiring high assurance
Core Feature Automated contract generation + formal proof attempts
Tech Stack Lean 4 (or Dafny) for contracts, LLMs for contract suggestion, GitHub Actions CI plugin, Z3/SMT solvers
Difficulty High
Monetization Revenue-ready: per‑project or per‑seat licensing with tiered support levels

Notes

  • Responds to the thread’s anxiety about “shipping code you don’t understand” and the call for “formal verification” as a way to regain confidence.
  • Would likely be well‑received by HN readers who cited design‑by‑contract, Lean, and the need for stronger guarantees in AI‑assisted development.

Read Later