IdeaWell

Project ideas from Hacker News discussions.

On the Navier–Stokes Millennium Prize Problem

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

1. Massive compute and cost estimates

“Across all attempted problems, the agents sent 4.9 million messages and used about 300 billion output tokens.” – minimaxir

2. Allegations that OpenAI may have trained on user chats

“While unlikely, we cannot rule out that de‑identified data derived from their usage of our products helped improve our models.” – OpenAI (blog post)

3. Demand for proof / transparency about data use

“OpenAI needs to definitively prove that their agents did not look at the existing work that was about to be published. Otherwise OpenAI simply stole the glory…” – enraged_camel

4. Corporate scooping motive (beating Anthropic)

“Clearly the goal was to scoop Anthropic … OpenAI heard the rumor that Anthropic solved an open problem. So they went nuts pulling all plugs to scoop them.” – tzone

5. Ethical concerns – sociopathic behavior & threats to remove a co‑author

“I said that if OpenAI released its result in the way proposed I would go public with what happened. The reply was, ‘Why would you ruin your career?’ … The reply was, ‘If you don’t want me to be nice, then I don’t have to be nice.’” – Tristan Buckmaster (via peri‑cl)

6. Role of human input vs. AI agents

“I was shown a prompt and told the internal research model had simply been given the problem statement… Over the course of the call … it emerged that an entire team had been working on the problem.” – contemporary343

7. Novelty and significance of the proof

“Our proofs differ significantly and even the precise results proved are different in the Euler case (forced vs unforced).” – heaney‑555

8. “Dark forest” secrecy incentive created by AI’s ability to scoop

“Massively incentivize keeping results secret, to avoid being scooped by someone willing to throw enormous compute at your partial solution. … less about hiding civilizations, and more about hiding information.” – recitedropper

9. Zero‑data‑retention / privacy‑opt‑out concerns

“If they opted out of training, then we definitely did not train on them.” – derangedHorse

10. Singularity / AGI‑era hype

“Are we at the dawn of the singularity now?” – echelon
“The uncanny period during the Singularity when superintelligence is already accomplishing feats that seem like magic, yet everyday life still looks mostly the same.” – baq (normal‑cy overhang)

🚀 Project Ideas

Training Data Provenance Auditor

Summary

  • A CLI/web tool that lets users upload a hash of their chat logs or code snippets and checks whether similar token sequences appear in a provider’s training data via membership inference or n‑gram indexing.
  • Core value: gives users concrete evidence about whether their private interactions contributed to a model’s behavior, addressing the “we cannot rule out” uncertainty.

Details

Key Value
Target Audience Researchers, developers, privacy‑concerned users of LLM APIs
Core Feature Upload hashed prompts → similarity search against provider‑published n‑gram index → probability score of inclusion
Tech Stack Python, FAISS for vector search, MinHash/LSH, FastAPI, React
Difficulty Medium
Monetization Revenue-ready: Tiered SaaS (free limited queries, $15/mo for bulk)

Notes

  • HN commenters repeatedly asked “Can we just check if their data was used?” (e.g., lossolo, lambda). This tool would satisfy that demand.
  • Enables independent verification and could become a standard audit step before publishing sensitive work.

Zero‑Data‑Retention (ZDR) Verification Service

Summary

  • Automated audit platform that tests whether an AI provider truly respects a user’s “do not train on my data” setting by sending canned prompts and checking for leakage in later model outputs.
  • Core value: provides verifiable ZDR compliance certificates, turning a contractual claim into provable evidence.

Details

Key Value
Target Audience Enterprises, academic labs, legal teams requiring ZDR guarantees
Core Feature Periodic probe‑and‑detect runs; generates attestation reports and alerts on violations
Tech Stack Go, Docker, Prometheus for monitoring, GPU‑enabled inference workers, PostgreSQL
Difficulty High
Monetization Revenue-ready: $200/mo per audited model endpoint

Notes

  • The discussion highlighted frustration over unverifiable ZDR claims (e.g., pertekk, dkolli). A third‑party audit would be welcomed.
  • Could be integrated into procurement workflows, giving HN’s security‑savvy audience a practical tool.

Influence‑Attribution Library for LLMs

Summary

  • Open‑source library that approximates the influence of individual training examples on a model’s prediction using influence functions or gradient‑based methods.
  • Core value: lets researchers quantify how much a specific chat log (e.g., a proof draft) affected a model’s output, moving beyond “we cannot rule out” to measurable impact.

Details

Key Value
Target Audience ML researchers, AI safety auditors, model developers
Core Feature Compute influence scores for given input‑output pairs; supports transformer models via HookedTransformer
Tech Stack PyTorch, Triton, Numpy, Jupyter notebooks, MIT‑licensed
Difficulty High
Monetization Hobby

Notes

  • Commenters like lambda and tedsanders debated the feasibility of proving influence; this library gives a concrete research tool.
  • Could spark new papers and discussions on HN about responsible AI training.

Transparent AI Research Log (TARL)

Summary

  • Encrypted, append‑only log where researchers record every prompt, model version, and timestamp; optionally publishes a Merkle root to prove log integrity without revealing content.
  • Core value: provides provable independence when claiming a result was not derived from AI‑seen private work.

Details

Key Value
Target Audience Academics, independent researchers, AI‑assisted proof writers
Core Feature Client‑side encryption, hash‑chaining, optional public verifier API
Tech Stack Rust (for cryptography), Svelte frontend, IPFS for backup, WASM
Difficulty Medium
Monetization Hobby

Notes

  • Tristan Buckmaster’s statement shows a desire to prove his sessions were not used (he asked for confirmation). TARL lets him cryptographically demonstrate non‑usage.
  • HN’s privacy‑focused crowd would appreciate the zero‑knowledge proof angle.

Lean Proof Collaboration Platform

Summary

  • Web‑based environment for co‑authoring Lean proofs with fine‑grained attribution, change‑tracking, and release‑gating to prevent premature scooping.
  • Core value: replaces informal email/chat exchanges with a transparent, auditable workflow that clarifies contribution and avoids author‑removal disputes.

Details

Key Value
Target Audience Mathematicians, formal methods engineers, Lean community
Core Feature Real‑time Lean editing, per‑line Git‑style blame, configurable review workflows
Tech Stack TypeScript, React, Lean 4 server via WASM, Postgres, WebSocket
Difficulty Medium
Monetization Revenue-ready: Free tier, $10/user/mo for private projects

Notes

  • The OpenAI/Anthropic dispute centered on authorship and credit (e.g., “it would be simpler if Levent was not an Anthropic employee”). A formal collaboration tool would make such demands unnecessary.
  • HN’s mathematically inclined users often discuss Lean; this would be a natural fit.

LLM Experiment Cost Estimator

Summary

  • Calculator that translates agent‑message counts, token usage, and model pricing into realistic dollar estimates, including amortized training‑inference overhead.
  • Core value: lets teams anticipate the multi‑million‑dollar budgets seen in the Navier‑Stokes episode before launching large‑scale agent swarms.

Details

Key Value
Target Audience AI startup founders, research leads, budget officers
Core Feature Input: #agents, avg messages/agent, token length, model → Output: estimated cost (training + inference)
Tech Stack Python (Flask), vanilla JS, Chart.js
Difficulty Low
Monetization Hobby

Notes

  • Multiple commenters did napkin math on the $15 M figure (e.g., gcr, hmate9). A shared estimator would prevent repetitive back‑of‑the‑envelope calculations.
  • Useful for grant planning and investor pitches on HN.

AI‑Generated Proof Similarity Checker

Summary

  • Service that compares a newly submitted proof (text or Lean code) against a corpus of known training data (e.g., public GitHub, arXiv) using embeddings to flag suspiciously high similarity.
  • Core value: helps detect potential plagiarism or inadvertent copying when a model may have seen private drafts.

Details

Key Value
Target Audience Journal editors, conference chairs, math departments
Core Feature Upload proof → cosine similarity search over embedding index → report with highlighted matches
Tech Stack Sentence‑Transformers, FAISS, FastAPI, Svelte
Difficulty Medium
Monetization Revenue-ready: $0.01 per check, enterprise plans

Notes

  • The debate over whether OpenAI’s proof “copied” Buckmaster’s approach (e.g., davidguy, keeda) would be clarified by an objective similarity metric.
  • HN users love tools that bring rigor to contentious claims.

Opt‑Out Compliance Dashboard

Summary

  • Browser extension / portal that reads a user’s ChatGPT/Codex privacy setting and periodically queries the provider’s API to verify that the setting is honored (e.g., by attempting to trigger training‑eligible probes and checking for leakage).
  • Core value: turns an opaque opt‑out toggle into a visible, continuously monitored compliance signal.

Details

Key Value
Target Audience Privacy‑conscious consumers, GDPR compliance officers
Core Feature Dashboard shows “Opt‑out respected?” status with timestamps and audit log
Tech Stack Typescript, WebExtension API, Node.js backend, encrypted storage
Difficulty Medium
Monetization Hobby

Notes

  • Commenters like lossolo and tbhf demanded proof that opting out actually works (“Can't you guys just check their account settings?”). This gives them that proof.
  • Could become a de‑facto standard for evaluating AI providers’ privacy claims.

Decentralized Training Data Provenance Registry

Summary

  • Blockchain‑based registry where model providers publish Merkle roots of hashed training‑data chunks; anyone can prove a specific document was (or was not) included by providing a Merkle proof.
  • Core value: offers a trustless, auditable way to settle the “we cannot rule out” debate without revealing raw data.

Details

Key Value
Target Audience LLM vendors, AI auditors, open‑source communities
Core Feature Submit training data → compute chunk hashes → store root on-chain; verification API for third parties
Tech Stack Rust (Substrate), IPFS for data blobs, Polkadot, WASM verifier
Difficulty High
Monetization Revenue-ready: $0.001 per hash submission, verification free

Notes

  • The discussion repeatedly questioned whether OpenAI could rule out usage of specific chats (e.g., recursivecaveat, nulld3v). A public registry would give a definitive answer.
  • HN’s cryptoculture and decentralization enthusiasts would likely champion and debate this approach.

Ethical AI Scoop Alert Service

Summary

  • Monitoring service that watches public preprint servers, code repos, and social media for signals that a competitor is scooping a user’s unpublished work via AI‑assisted research, and alerts the originator.
  • Core value: gives early warning of potential AI‑mediated scooping, letting researchers decide whether to publish jointly or seek attribution.

Details

Key Value
Target Audience Academics, independent researchers, corporate R&D teams
Core Feature Scrape arXiv, GitHub, Twitter; match against user‑provided private‑work fingerprints (hashes, topics); send Slack/email alerts
Tech Stack Python (Scrapy, Transformers), Elasticsearch, React, Webhooks
Difficulty Medium
Monetization Revenue-ready: $25/mo per monitored topic

Notes

  • The thread alleged OpenAI used rumors to “scoop” Anthropic’s work (e.g., tzone, jrflo). An alert system would let users detect such behavior early.
  • HN’s audience frequently discusses scooping and ethics; this tool would feed directly into those conversations.

Read Later