IdeaWell

Project ideas from Hacker News discussions.

Jev meets the Trolley Problem

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

Three prevalent themes

  1. Decision models vs. LLMs – safety, constraints, and interpretability
  2. “well, it's a nice property of decision models that they can only choose from the options you give them. As I write in the article, I think that's a way in which they are safer than LLMs” – zlandes86
  3. “With a decision model, you can just constrain its choices to those you're comfortable with it making. An LLM requires another layer of interpretation of its outputs to get the same property.” – zlandes86

  4. “I would contend that you can enforce the same determinism with some plumbing and that the thinking process is more helpful.” – verdverm

  5. Human intuition and ethical implications of the trolley problem for AI

  6. “Would people really pull the level to save 5 people? It seems to me like the default would be to just do nothing, stand by and record what happens with your phone.” – deadbabe
  7. “Just standing by makes total sense. By participating you become part of the problem. If you pull the lever, your direct action kills a person. Legally saving 5 people does not offset the murder of a single person.” – throw9393u300

  8. “Surveys: Awad et all 2020 asked 70,000 people in 42 countries the 5 v 1 setup and 81% said they'd divert it.” – zlandes86 (citing empirical data)

  9. Robustness of the experimental design and calls for extension

  10. “I used resampling and multiple variations and orderings of each question so a random lucky phrasing wouldn't dominate the result.” – zlandes86
  11. “I did not go into this with any sense of what the results would be--I was kind of shocked. I reran it from scratch this week the results held up across the board.” – zlandes86
  12. “Would you mind adding Jeffy to your benchmark?” – nico (suggesting model expansion)
  13. “Any relation to https://github.com/firelex/jeff ?” – evanjrowley ( probing alternatives)

🚀 Project Ideas

Generating project ideas…

Failure Mode Exchange (FME)

Summary

  • A collaborative platform where researchers and developers can submit, tag, and search AI decision‑model failure cases (e.g., unexpected trolley‑problem responses, hallucinated outputs).
  • Core value: democratizes failure‑mode data so teams can quickly spot weaknesses and improve model safety without reinventing the wheel.

Details

Key Value
Target Audience AI safety researchers, ML engineers, product teams using decision models or LLMs
Core Feature Versioned repository of failure‑mode scenarios with metadata (model, prompt variant, outcome) and search/filter UI
Tech Stack React + TypeScript frontend, Node.js/Express backend, PostgreSQL database, Docker for deployment
Difficulty Medium
Monetization Revenue-ready: Tiered subscription (free public repo, paid private teams & advanced analytics)

Notes

  • HN commenters lamented that model labs hoard failure‑mode data; FME gives them a place to share openly (as verdverm suggested collecting more failure modes).
  • Enables reproducible studies like zlandes86’s trolley‑problem sweeps and can spark discussion on new edge cases.

Constrained Choice Wrapper (CCW)

Summary

  • A lightweight middleware that takes any LLM (or decision model) and forces its output to be a single choice from a predefined set, logging the internal “thinking” tokens for audit.
  • Core value: provides the safety guarantee of decision models (limited action space) while letting teams keep using powerful LLMs for tasks like content moderation.

Details

Key Value
Target Audience Developers building LLM‑based moderation, recommendation, or decision‑support systems
Core Feature Choice‑enforcement API (POST /choose) that validates model output against an allowed list and returns a structured decision object
Tech Stack Python (FastAPI) wrapper, optional integration with HuggingFace Transformers or OpenAPI SDK, Redis for caching choices
Difficulty Low
Monetization Hobby

Notes

  • Commenters noted that decision models are safer because they can only choose from given options (zlandes86) and that similar determinism can be added via “plumbing” (verdverm).
  • CCW offers exactly that plumbing, making it easy to adopt and discuss in safety‑focused threads.

Decision Model Benchmark Harness (DMBH)

Summary

  • An automated harness that runs decision models (Jev, Jeffy, etc.) across thousands of prompt variations, orderings, and resampling schemes, producing statistical reports and visualizations.
  • Core value: eliminates manual, error‑prone experimentation and lets researchers rigorously compare models on scenarios like the trolley problem or custom ethical dilemmas.

Details

Key Value
Target Audience AI researchers, model evaluators, educators teaching decision theory
Core Feature Scenario definition DSL (YAML), parallel execution engine, result aggregation with confidence intervals, web dashboard
Tech Stack Go for orchestration, PostgreSQL for results, React + Chart.js dashboard, Kubernetes (optional) for scaling
Difficulty High
Monetization Revenue-ready: Usage‑based pricing (per‑run compute credits) or enterprise license

Notes

  • zlandes86 emphasized resampling and multiple variations to avoid lucky phrasing; DMBH automates exactly that workflow.
  • Would be welcomed by HN users who want to replicate or extend the trolley‑problem analysis and explore new failure modes in a reproducible way.

Read Later