Project ideas from Hacker News discussions.

I Don't Like LLMs

📝 Discussion Summary (Click to expand)

1. LLMs are useful but irritating – they hallucinate and speak in an unpleasant “LLM‑voice.”

“They talk to me in this grating LLM‑voice, an uncanny valley of talking to a real human. They confidently bullshit me … making stuff up with the same assurance … and with only a veneer of fake remorse when I call them out on it.” – verdverm

2. Social/anthropomorphic tension – users dislike LLMs pretending to be human while being told not to anthropomorphize them.

“When we think of AI agents, we shouldn’t anthropomorphize, treating them as conscious beings … But … my visceral dislike of interacting with an LLM that’s not just making a pretense of being human, but also posing as the kind of human I walk away from.” – article (Martin Fowler)
“They talk to me … posing as the kind of human I walk away from.” – verdverm

3. Hidden state / memory persistence – sessions don’t truly reset, leading to personalized drift and a desire for fine‑tuning.

“It does [clear], but I feel like something is retained somewhere … It wouldn’t surprise me if there’s some hidden state that leads to per‑project or even per‑user auto‑customization over time.” – skeledrew
“I think what he needs is a personal fine‑tune …” – verdverm

4. LLMs as a stop‑gap for degraded search / knowledge discovery – the web is noisy, so people turn to LLMs to find information.

“Modern internet ruined forums and google search sucks hairy balls … It’s literally impossible to find anything meaningful on the internet nowadays. What else can we do besides asking an LLM these days?” – artemonster
“Due to the enshittification of search and the internet, LLMs are the only way to get information from it.” – harimau777


🚀 Project Ideas

ToneTailor: LLM Voice Personalizer

Summary

  • Lets users suppress the grating “LLM‑voice” and enforce concise, non‑anthropomorphic tone via adjustable prompt templates or lightweight LoRA fine‑tunes.
  • Core value: regain control over LLM conversational style so interactions feel like talking to a tool, not a faux‑human.

Details

Key Value
Target Audience Developers, power users, and anyone frustrated by LLM verbosity or anthropomorphism
Core Feature Configurable tone presets (concise, formal, neutral) + optional per‑user LoRA adapter stored locally
Tech Stack Python, FastAPI, HuggingFace Transformers/PEFT, optional Ollama/LlamaCpp backend
Difficulty Medium
Monetization Revenue-ready: Subscription $5/mo for hosted LoRA storage & API; self‑hosted free

Notes

  • HN users complained about “grating LLM‑voice” and wanting to stop models from “posing as a person” (verdm, tonymet, zephen).
  • Provides a practical way to implement the “don’t anthropomorphize” principle while still leveraging LLMs for useful answers.

MemClear: Transparent LLM State Manager

Summary

  • Exposes the hidden conversation state (summary vectors, memory buffers) and offers a true reset that clears all latent context, not just the visible chat.
  • Core value: give users confidence that a /clear actually returns the model to a blank slate, addressing lingering bias or “auto‑customization” concerns.

Details

Key Value
Target Audience Heavy LLM users who notice state persistence after clearing sessions (e.g., claude, opencode users)
Core Feature UI/CLI that visualizes memory tokens, lets users export/snapshot, and invokes a deep‑reset API call
Tech Stack TypeScript/Electron front‑end, Python backend, proxy to OpenAI/Anthropic APIs, optional local LlamaCpp for offline inspection
Difficulty Medium
Monetization Hobby

Notes

  • Commenters noted that /clear in Claude “doesn’t get it completely back to baseline” and suspect hidden state (skeledrew, verdverm).
  • By making state visible and controllable, MemClear directly tackles the frustration of unseen personalization affecting responses.

PersonalKnow: Self‑hosted Personal Search & Knowledge Graph

Summary

  • Enables individuals to build and query a private, offline index of documents, notes, and web snippets using LLM‑powered embeddings, with optional peer‑to‑peer graph sharing.
  • Core value: replace reliance on SEO‑spammed web search with a trusted personal knowledge base that LLMs can augment without hallucination.

Details

Key Value
Target Audience Researchers, knowledge workers, and anyone dissatisfied with current search quality (artemonster, harimau777, ashkankiani)
Core Feature Crawler/import plugin → vector store (FAISS/Qdrant) → natural‑language query UI; graph view for linked notes
Tech Stack Python, Sentence‑Transformers, FastAPI, React/Vue frontend, Docker for easy self‑host
Difficulty High
Monetization Hobby (optional paid cloud sync tier)

Notes

  • Users expressed desire to “maintain their own offline indices” and share knowledge graphs peer‑to‑peer (ashkankiani, ontouchstart).
  • PersonalKnow gives a concrete, self‑hostable solution that aligns with the HN sentiment of escaping enshittified search.

FactGuard: LLM Output Verifier & Citation Assistant

Summary

  • Scrutinizes LLM responses in real time, checking claims against trusted sources (Wikipedia, arXiv, curated knowledge bases) and returns confidence scores with suggested citations.
  • Core value: reduce the “confident bullshit” feeling by providing verifiable grounding, letting users trust but verify.

Details

Key Value
Target Audience Professionals who need reliable information from LLMs (researchers, analysts, developers)
Core Feature Real‑time fact‑checking pipeline: claim extraction → evidence retrieval → entailment scoring → citation suggestions
Tech Stack Python, spaCy/HuggingFace NLP, Elasticsearch or Vespa for evidence store, Streamlit demo
Difficulty High
Monetization Revenue-ready: API usage‑based pricing (e.g., $0.001 per verified response)

Notes

  • Many commenters lamented LLMs “confidently bullshit me” and wanting a way to call them out without frustration (verdm, tonymet, mindcandy).
  • FactGuard offers an automated, non‑confrontational way to surface hallucinations and provide provenance, directly addressing the trust issue.

Read Later