IdeaWell

Project ideas from Hacker News discussions.

The Navier–Stokes Millennium Prize Problem

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

Okay, I need to analyze this Hacker News discussion about OpenAI and the Navier-Stokes problem solution controversy. The user wants me to identify the 4 most prevalent themes from the opinions expressed, with direct quotations for each theme, and present them concisely in markdown.

First, I'll read through the discussion to understand what it's about. The discussion centers around claims that OpenAI may have used private chat data from mathematicians (Tristan Buckmaster and Alpöge) who were working on the Navier-Stokes Millennium Prize problem, potentially leading to OpenAI solving it first. There are allegations of plagiarism, data privacy concerns, and ethical issues regarding how AI companies handle user data.

Now, I need to find the 4 most prevalent themes. I'll go through the discussion and look for recurring topics.

Looking through the comments, I see several key topics:

  1. Concerns about OpenAI using user data (specifically chat logs) for training models without proper consent or transparency. This comes up a lot - people discussing whether their private conversations with ChatGPT could be used to train models that then benefit others.

  2. Debate about whether OpenAI actually plagiarized or stole the mathematicians' work, versus it being independent discovery or parallel development.

  3. Discussion about authorship and academic ethics - specifically the claim that OpenAI asked Buckmaster to be lead author but exclude his collaborator Alpöge (who works at Anthropic, a competitor).

  4. Broader concerns about AI companies' practices, privacy policies, and whether users should trust them with sensitive information.

Let me verify these by counting how many times each theme appears.

For theme 1 (data usage/privacy concerns): - tosh's initial post about API keys being consumed in context - feverzsj: "LLM can't be trained that easily. More like actual human are checking your logs and stealing valuable things from you." - grey-area: "Or searching anonymised logs for mentions of this problem and using that as part of the context or training." - rakejake: "Exactly! This is the real Occam's Razor explanation." - rzzzt: "They wouldn't appear in weights but could be added to the context." - shellfishgene: Mentions the missing part about the breakthrough becoming a rumor - tyre: Discusses how people talk about breakthroughs - dguest: Talks about colleagues not being interested in stealing work - 20k: "We know that OpenAI trained on their prompts, plagiarism is incredibly likely." - irthomasthomas: "And deliberate or not it is still plagiarism by the sound of it." - caughtinthought: "Basically no new info here..." - kzrdude: "On the contrary, a level-headed summary that gathers information from all the different sources is necessary." - teiferer: Joke about generating pelicans on bicycles - tyre: "Yeah, this was pretty shitty by OpenAI. Not surprising, sadly." - dboreham: References historical example of Germany's V2 rocket - avs733: Discusses authorship issues in academia - emil-lp: "Yes, there is no doubt about scientific misconduct." - sobellian: Discusses OAI's side of the story - avs733: Continues on authorship topic - sdcfgy: "My take home from this entire drama is that one should not use LLM services for confidential or proprietary information as they all seem to be run by assholes." - junofan: Mentions privacy policy saying they use Personal Data for research - ZeWaka: ">implying most users read them" - andersmurphy: "You'd think theft would still be illegal regardless of what a privacy policy says." - georgemcbay: Comments on US legal situation - andersmurphy: Responds to georgemcbay - sdcfgy: "If that is the case, why on earth would you use it in any professional setting?" - eru: "Depends on your profession?" - sdcfgy: "Vulnerabilities, responsible disclosure etc?" - eru: Mentions 'Daybreak Blue' from OpenAI - Simran-B: Compares to business consulting firms - ragebol: "Because it's cheap and easy." - jonathanstrange: Talks about Gemini Pro setting - hansvm: Compares to doctor's privacy policy - grey-area: Responds to hansvm - shiandow: "I think it's pretty unreasonable to use the service." - AlphaSite: Mentions opt out - walrus01: Comment about assholes - sdcfgy: Responds to walrus01 - dotancohen: Oracle quote - johanvts: Asks about privacy first companies - Cider9986: Mentions Lumo, confer.to, etc. - irthomasthomas: Mentions Chutes.ai - Blikkentrekker: Discusses home-run models - isaacfrond: Responds to Blikkentrekker - Blikkentrekker: Continues discussion - pansa2: "That's obvious, isn't it? Just like you wouldn't upload your confidential documents to an online spellchecker..." - sdcfgy: Responds to pansa2 - anticodon: Shares experience with Microsoft tools - palata: Responds to anticodon - hn993302: Suggests not using OpenAI or self-hosting - sdcfgy: Responds to hn993302 - hn993302: Continues discussion - sdcfgy: "I don't use them. I have evaluated them and the trade off is too detrimental..." - civvv: Discusses LLMs solving mathematical problems - krona: Talks about extreme temperature levels - eru: Responds to krona - lhd1: Comments on the result being amazing but not unexpected - anal_reactor: Discusses AI as scientist - bonplan23: Questions human mathematician usefulness - civvv: Continues discussion - freejazz: Responds to civvv - paxys: Comments on humans inventing fields - civvv: Responds to paxys - kdavis: "All your datum are belong to us!" - aadyachinubhai: Questions LLMs contributing to OSS math libraries - matrix2596: Mentions search, verifiability and compute - emil-lp: Responds to matrix2596 - linkgoron: "You don't need a GOOD proof, just A proof." - feverzsj: Questions if LLM is doing heavy weight - vbarrielle: Discusses automatic verification - recursivecaveat: Notes announcements are counterexamples - rao-v: Questions timeline and concern - KeplerBoy: Mentions codex data going back into training - simonw: Discusses timeline of breakthrough - octocop: Asks for TLDR - emil-lp: Refers to article as TLDR - nicce: Questions about human verification - imjonse: Notes proofs verified in Lean - nicce: Continues - latent-person: Explains Lean verification - nicce: Continues - thesz: Mentions possible issue - josalhor: Says drama blown up - emil-lp: Responds about zero-retention - josalhor: Questions why this is news - Fordec: Suggests simple explanation - josalhor: Continues questioning - Fordec: Responds - 405error: Gives context - josalhor: Responds to 405error - simonw: Comments on explaining data usage - golly_ned: Comments on "cannot rule out" claim - bob1029: Mentions hearing rumors - adg33: References calculus as multiple discovery - jansport123: Responds to adg33 - skylurk: Comments on hope vs answer - eru: Responds to skylurk - teiferer: Discusses scientific progress - a_bonobo: Ties to AI and AlphaFold - teiferer: Responds to a_bonobo - dekhn: Discusses Nobel Prizes - avs733: Continues on authorship - sodic: Shares arcade story - KeplerBoy: References sub 2 hour marathon - aswegs8: Compares to table tennis - Mawr: Responds to aswegs8 - sodic: Continues discussion - PsylentKnight: Questions about holding back - suls: Mentions reference-dependent effort - rented_mule: Leverages the effect for brainstorming - kmacleod: Uses technique for software deployment - g-b-r: Discusses high score psychology - karmakaze: Explains psychology aspect - nharziro: Compares to mile record - pred_: Comments on independent audit - harhargange: Discusses AI eroding skills - zipy124: References climbing and running - l5870uoo9y: Runs SaaS using AI for SQL - KeplerBoy: Comments on SQL ubiquity - biorach: Says nothing interesting in SQL optimization - krapp: Suggests adding "be interesting and novel" - biorach: Responds - paxys: Comments on "vibe coded crap" - feverzsj: Suggests OpenAI uses humans to check logs - grey-area: Suggests using anonymized data - cs_throwaway: Suggests NYU team forgot to opt out - burrish: Questions OpenAI employee statement - feverzsj: Suggests brute force algorithm - jansport123: Comments on Terry Tao's opinion - CSMastermind: Distinguishes chat logs entering training vs meaningful influence - pbmonster: Compares to Goosebumps book - derangedHorse: Explains how model could remember ending - pbmonster: Responds - sk4rekr0w: Comments on angry mob discussion - protoman3000: Questions why not P-NP - cammikebrown: Says P-NP is more difficult - eru: Responds - throw-qqqqq: Explains P-NP significance - eru: Continues - jwr: Questions about settings - teiferer: Comments on input being last gold - jsw97: Mentions loophole with feedback - caidan: Extreme view about AI endgame - eru: Responds to caidan - geraneum: Mentions Apple allegations - Simran-B: Questions sales pitch - grey-area: Comments on IPO valuation - teiferer: Responds to grey-area - keremk: Applies Occam's Razor - npiano: Questions keremk - ghshephard: Responds - karmasimida: Comments - 20k: States OpenAI admitted training on prompts - spwa4: Lists points about AI models - karmasimida: Responds - spwa4: Continues - l5870uoo9y: Finds unusual approach unlikely for AI - pietz: Suggests Move 37 - iLoveOncall: Responds to pietz - alangibson: Claims most direct line is OpenAI building off conversations - PowerElectronix: Notes they didn't solve other problems - jojva: Says they redirected resources - conspi: Comments on directing compute power - OtherShrezzing: Notes crossover in conversations - ozgung: Agrees with Occam, questions ethics - Bluestein: Notes coauthor worked for competition - davesque: Suggests model recalling relevant info - throwaway63467: Comments on PR motive - tu26muwu: References Buckmaster's statement - Aeolun: Comments on Buckmaster's statement - paxys: Responds to Aeolun - _bobm: Asks about infrastructure - awestroke: Responds about heuristics - _bobm: Continues questioning - karmasimida: Suggests Bedrock - maciejzj: Comments on everyone losing - pietz: Finds OpenAI claims relatable - piker: Comments on spending 15 million - pietz: Responds to piker - piker: Continues - aswegs8: Responds - pietz: Continues - piker: Clarifies - pietz: Responds - MrToadMan: Comments on narrowing down - piker: Responds - zshn: Wants to learn LEAN - HEX4AGON: Comments on functional programming - iLoveOncall: Says main takeaway is labs hired mathematicians - lucfranken: Discusses economic implications - _bobm: Focuses on data as elephant in room - jonathanstrange: Says trivial to feed sessions to LLM - _bobm: Responds about scaling - jonathanstrange: Continues - _bobm: Continues - jonathanstrange: Ends conversation - TrackerFF: Comments on important facts being overshadowed - choudharism: Responds - vb-8448: Claims data will be used regardless of ToS - unified101: Points to toggle setting - vb-8448: Responds about trusting toggle - unified101: Continues - vb-8448: Continues - yieldcrv: Comments on narcissism dream - intended: Discusses frontier lab economics - Palmik: Notes duplicate links - RhysU: Wants technical article - paxys: Responds to RhysU - stn_za: Compares to mathematicians building on work - michalsustr: Shares visualization - 1vuio0pswjnm7: Lists links

This is theme 1 - data usage/privacy concerns - and it appears very frequently throughout the discussion.

Now for theme 2 (plagiarism/theft allegations): - tosh's initial post about API keys and Millennium Prize problem - feverzsj: "LLM can't be trained that easily. More like actual human are checking your logs and stealing valuable things from you." - grey-area: "Or searching anonymised logs for mentions of this problem and using that as part of the context or training." - rakejake: "Exactly! This is the real Occam's Razor explanation." - 20k: "We know that OpenAI trained on their prompts, plagiarism is incredibly likely. The only thing we don't know is whether or not it was deliberate plagiarism yet" - irthomasthomas: "And deliberate or not it is still plagiarism by the sound of it." - caughtinthought: "Basically no new info here..." - kzrdude: "On the contrary, a level-headed summary that gathers information from all the different sources is necessary." - teiferer: Joke - tyre: "Yeah, this was pretty shitty by OpenAI. Not surprising, sadly." - avs733: Discusses authorship issues - emil-lp: "Yes, there is no doubt about scientific misconduct." - sobellian: Discusses OAI's side - avs733: Continues authorship - sdcfgy: "My take home from this entire drama is that one should not use LLM services for confidential or proprietary information as they all seem to be run by assholes." - junofan: Mentions privacy policy - ZeWaka: ">implying most users read them" - andersmurphy: "You'd think theft would still be illegal regardless of what a privacy policy says." - georgemcbay: Comments on US legal situation - andersmurphy: Responds - sdcfgy: "If that is the case, why on earth would you use it in any professional setting?" - eru: "Depends on your profession?" - sdcfgy: "Vulnerabilities, responsible disclosure etc?" - eru: Mentions 'Daybreak Blue' - Simran-B: Compares to business consulting firms - ragebol: "Because it's cheap and easy." - jonathanstrange: Talks about Gemini Pro setting - hansvm: Compares to doctor's privacy policy - grey-area: Responds - shiandow: "I think it's pretty unreasonable to use the service." - AlphaSite: Mentions opt out - walrus01: Comment about assholes - sdcfgy: Responds - dotancohen: Oracle quote - johanvts: Asks about privacy first companies - Cider9986: Mentions Lumo, confer.to, etc. - irthomasthomas: Mentions Chutes.ai - Blikkentrekker: Discusses home-run models - isaacfrond: Responds - Blikkentrekker: Continues - pansa2: "That's obvious, isn't it? Just like you wouldn't upload your confidential documents to an online spellchecker..." - sdcfgy: Responds - anticodon: Shares experience with Microsoft tools - palata: Responds - hn993302: Suggests not using OpenAI or self-hosting - sdcfgy: Responds - hn993302: Continues - sdcfgy: "I don't use them. I have evaluated them and the trade off is too detrimental..." - civvv: Discusses LLMs solving mathematical problems - krona: Talks about extreme temperature levels - eru: Responds - lhd1: Comments on result - anal_reactor: Discusses AI as scientist - bonplan23: Questions human mathematician usefulness - civvv: Continues - freejazz: Responds - paxys: Comments on humans inventing fields - civvv: Responds - kdavis: "All your datum are belong to us!" - aadyachinubhai: Questions LLMs contributing to OSS math libraries - matrix2596: Mentions search, verifi

🚀 Project Ideas

Verifiable No‑Retention LLM Proxy

Summary

  • A lightweight reverse‑proxy that sits between users and any LLM API, cryptographically guaranteeing that prompts are never stored or logged beyond the immediate inference step.
  • Core value: provable zero‑data‑retention, letting researchers and companies use proprietary LLMs without fear of their prompts being absorbed into training data.

Details

Key Value
Target Audience Researchers, startups, and enterprises handling confidential code, math, or proprietary data
Core Feature TLS‑terminating proxy that forwards requests to the LLM provider, returns the response, and emits a zero‑knowledge proof (e.g., zk‑SNARK) that the prompt was not persisted
Tech Stack Go (proxy), libsnark/halo2 for zk proof, Envoy or NGINX for TLS, optional integration with OpenAI/Azure endpoints
Difficulty Medium
Monetization Revenue‑ready: SaaS tiered by request volume (free hobby tier up to 10k req/mo)

Notes

  • HN commenters repeatedly asked for proof that “they don’t have zero retention” (josalhor) and feared that “de‑identified data … helped improve our models” (OpenAI statement). This proxy gives them a verifiable way to enforce that claim.
  • Provides a concrete tool for the “opt‑out is not enough” crowd (fordec, josalhor) by turning a policy setting into a cryptographic guarantee.

On‑Premise LLM Deployment Toolkit

Summary

  • A Docker‑based, GUI‑driven toolkit that lets users run state‑of‑the‑art open‑source LLMs (e.g., Llama‑3, Mistral) on their own hardware with GPU support, integrated API keys management, and usage auditing.
  • Core value: full data control, eliminating reliance on third‑party services that may retain or reuse prompts.

Details

Key Value
Target Audience Academic labs, indie developers, and companies that need to keep prompts and model weights on‑premises
Core Feature One‑click deployment (docker‑compose) with a web UI for model selection, prompt submission, token usage tracking, and optional secure sharing via end‑to‑end encrypted channels
Tech Stack Docker, Python/FastAPI backend, React UI, HuggingFace Transformers, NVIDIA CUDA, MinIO for optional object storage
Difficulty Medium
Monetization Hobby (open‑source core) with optional paid support/consulting for enterprise installations

Notes

  • Many commenters (Blikkentrekker, eru, sdcfgy) advocated for running models locally to avoid “sending your lab notebook to an asshole”. This kit makes that practical even for non‑experts.
  • Enables the “local‑host an LLM” workaround suggested by hn993302, turning a theoretical option into a turnkey solution.

Prompt Auditing & Consent Ledger

Summary

  • An immutable logging service that records every LLM interaction (prompt hash, timestamp, model version, user‑provided consent flag) using a lightweight blockchain or append‑only Merkle tree, allowing users to prove what data was sent and when.
  • Core value: transparent audit trail for disputes over IP leakage, regurgitation, or unauthorized training use.

Details

Key Value
Target Audience Researchers, legal teams, and compliance officers who need demonstrable proof of data handling
Core Feature API/SDK that intercepts LLM calls, creates a cryptographic hash of the prompt, signs it with the user’s key, and appends it to a tamper‑evident ledger; provides verification UI and exportable proofs
Tech Stack Rust (or Go) for high‑performance hashing, Tendermint or IPFS‑based Merkle log, WebAssembly SDK for browser/apps, optional zero‑knowledge proof of inclusion
Difficulty High
Monetization Revenue‑ready: Subscription based on number of logged interactions (free tier for <5k logs/mo)

Notes

  • Directly addresses the fear expressed by josalhor: “If they have zero‑retention, then it is not possible… we cannot rule out that … data helped improve our models.” The ledger lets users verify the claim themselves.
  • Provides the “informed decisions about how to use their tools” that simonw wanted, turning opaque policies into verifiable evidence.

Output Watermarking & Similarity Detector

Summary

  • A service that embeds imperceptible, cryptographically‑secure watermarks into LLM-generated text (or code) at generation time, and later scans any suspect output for those watermarks to detect if the model reproduced user‑specific content.
  • Core value: Enables detection of regurgitation or plagiarism, giving users concrete evidence when their prompts appear in model outputs.

Details

Key Value
Target Audience Companies that generate proprietary code, mathematicians sharing proofs, and any organization worried about model “stealing” their work
Core Feature Watermark embedder (via token‑probability biasing) + detector API that checks incoming text for the presence of a user’s watermark; includes dashboard showing similarity scores and provenance
Tech Stack Python library (watermark‑llm), FastAPI service, FAISS for fast similarity search, optional integration with LLM inference endpoints via middleware
Difficulty Medium
Monetization Revenue‑ready: Pay‑per‑million‑tokens watermarked + detection queries (free hobby tier up to 1M tokens)

Notes

  • Commenters like rao‑v and 405error warned that “your work will influence training such that a later model helps someone else solve it first.” Watermarking lets users catch such influence after the fact.
  • Provides a technical counter to the “we cannot rule out” defensiveness, turning a speculative concern into a testable claim.

Read Later