Things I've Done with AI
Developers compare concrete ways they use large language models—from refactoring abandoned codebases, automating taxes, and building games or personal tools, to writing nearly all of a bespoke note‑taking app—against criticism that most AI‑assisted projects are trivial, misleading, or quickly abandoned. Commenters argue over whether AI‑generated code meaningfully boosts productivity or merely produces unmaintainable “slop,” raising concerns about testing, long‑term reliability, data privacy, and skill atrophy. Underneath is a broader question about software’s future: whether embracing AI will create a new divide between those who leverage it as powerful infrastructure and those who resist it on principle or out of skepticism.
Scope of AI-Built Projects
- Many examples shared: personal assistants, note-taking tools, macropads, games, clocks with irregular ticking, drawing “towns,” fictional encyclopedias, blood-test viewers, feature boards, and support-email bots.
- Some are clearly whimsical or experimental; others aim at concrete utility (e.g., automating life admin, viewing medical tests, support responses).
Usefulness vs. “Slop”
- Critics argue many AI projects are trivial, duplicative, or self-referential (“tools to use AI”), and often abandoned quickly.
- Specific criticism targets a fictional encyclopedia that fabricates facts without warning, seen as actively misleading.
- Defenders say personal joy and learning are valid goals; demanding mass-market success or revenue as a bar is unreasonable and often ideological.
Throwaway Code & Abandonware
- One side sees the flood of short-lived tools as evidence of no real productivity gain, just dopamine.
- Others welcome cheap, disposable code: write one-offs, get value, then delete. Reviving abandoned open-source projects via LLMs is cited as concrete value.
Concrete Use Cases
- Reported successful uses:
- Reviving and modernizing an abandoned web-based editor.
- Large-scale refactors of legacy codebases.
- Tax workflows: renaming and extracting data from PDFs, building web UIs to summarize taxes, preparing documentation.
- Custom CAD-like desk design tools using browser 3D and B-rep modeling.
- Custom note-taking apps with specific editor behavior; multiple educational and puzzle games.
Hallucinations, Safety, and Privacy
- Several note subtle but real hallucinations, especially with large or complex data (lipids, taxes); results can look correct but be numerically or temporally wrong.
- Mitigations discussed: have LLMs generate deterministic scripts/tools, then run them; extract structured data (JSON) first; use LLMs mainly as validators or hypothesis generators.
- Strong disagreement over uploading sensitive data (tax, medical) to cloud models; some see it as fine, others as dangerously naive.
AI, Skills, and Careers
- View 1: Using AI too heavily risks skill atrophy and dependence; might ultimately reduce one’s value.
- View 2: Refusing AI means “missing out” or being “left behind” in a major computing shift.
- Pushback: that framing is condescending; tools are easy to learn later, and some skepticism is principled or cautious.
- Some report barely typing code themselves now, relying on tools like agentic coding environments, but still reviewing output.
Maintainability and System Design
- Debate over “code that works” vs. maintainable systems:
- Pro-AI-regeneration side suggests tests + LLMs can regenerate “ugly” code on demand.
- Critics argue tests can’t capture all behavior; LLMs generate code “nodes” but not the important “edges” (assumptions, relationships).
- Guardrail-style programming (guided by tests) is seen as insufficient for user-facing, long-lived systems.
Open Source and Bespoke Tools
- Some predict fewer polished open-source apps: with LLMs, it’s easier to build bespoke tools that exactly match one person’s workflow, with little motivation to generalize or support others.
- Others counter that using simple, file-based storage (e.g., markdown) and backups mitigates risk; critics question why to reimplement what already exists and is maintained.