IdeaWell

Project ideas from Hacker News discussions.

I built non-autoregressive decision models with RL a year ago

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

Five prevalent themes in the discussion

  1. Prior art vs. independent discovery
    Many commenters debated whether Laya (the OP’s open‑source work) actually predated Jev or represented a separate invention.
  2. “This project was built on the exact research on jev architecture research one year ago” – nandakishor_ml
  3. “Jev was built using the same architecture Laya's author proposed[1] in March 2025. Laya is an open‑source system based on that research from a year ago.” – klibertp

  4. “OP has published everything in the open, so I can take him at face value…. this is most probably two independent discoveries of the same idea.” – cgio

  5. Marketing/hype versus technical substance
    A recurring criticism was that Jev’s visibility stems from VC backing and polished branding rather than novel science.

  6. “Jev is only on people's mouths because they made friends with venture capitalists and used the publicity blowhorns that come with that.” – cmrdporcupine
  7. “It's an embarrassing showing for our community, seems like nobody has read anything. None of the claims of the blog post add up.” – prodigycorp
  8. “Jev is exceptionally‑well branded.... OPs ‘marketing’ is a single post on Reddit …” – johnfn

  9. “Ideas alone are near worthless and execution matters.” – malux85

  10. Technical assessment (architecture, hallucination claims, cost)
    Commenters examined whether Jev is merely a BERT‑style model, evaluated its hallucination‑free claims, and compared its inference cost to LLMs.

  11. “I can see why people would want ready made one shot classifier … but I wouldn’t call it breakthrough. And I believe many labs will replicate it in no time…” – Oras
  12. “Gemini 2.5 Flash Lite is $500/Gt, Jev is $42/Gt. AKA an order of magnitude cheaper.” – fastball
  13. “Anyone who has worked in ML for 10+ years would already know that the usage of LLMs for everything is lazy, wasteful and a high degree of marketing on it.” – tchalla

  14. “They’re definining hallucination as a property of iterative generation …” – seizethecheese (discussing the hallucination claim)

  15. Practical utility and real‑world benefits
    Several users highlighted concrete speed, cost, and workflow improvements when using Jev (or Laya) for classification, routing, and token‑heavy tasks.

  16. “I've been deeply impressed with Jev as it made a bunch of workloads we had on Luna or Gemini 10x cheaper and 2x faster…” – zurfer
  17. “Jev did it in minutes and for less than 1 cent.” – sharms
  18. “This approach has helped us save more than $1 Million annually against the straight LLM classifiers at scale.” – slrainka

  19. “I made it clear that it is useful and I can see many people using it including myself.” – Oras

  20. Community perception, credit, and meta‑discussion (HN voting, open source, reproducibility)
    The thread also reflected on how ideas gain traction on HN, the role of open source versus closed source, and concerns about attribution.

  21. “1052 pts on HN. I'd say they are pretty good at marketing actually.” – mcapodici
  22. “HN readers are awesome at saying something is great and upvoting it, but unless HN readers are your market it means absolutely nothing.” – onion2k
  23. “OP received a lot of echo chamber support in the subreddit, and recommended to post to HN, so here we are.” – verdverm
  24. “Did you see the carnage that typesafe's landing page was? every other post here is llm generated…” – prometheus1992
  25. “It’s a tale as old as time — people don’t understand that marketing and branding are just as important, if not more so, than the product.” – johnfn

🚀 Project Ideas

JevBench: Open Benchmark for Zero-Shot Structured Classification

Summary

  • Provides a standardized benchmark suite to evaluate models like Jev, Laya, and BERT variants on zero-shot classification tasks, measuring accuracy, calibration, latency, and cost.
  • Core value proposition: objective, reproducible measurements that cut through hype and help teams pick the right model for production.

Details

| Target Audience | ML researchers, model developers, product teams evaluating small classifiers | | Core Feature | Curated datasets (sentiment, ticket triage, medical, finance) with automated evaluation scripts, leaderboard, and calibration metrics | | Tech Stack | Python, Hugging Face datasets & evaluate, FastAPI for leaderboard, optional Streamlit UI | | Difficulty | Medium | | Monetization | Hobby |

Notes

  • Commenters complained about lack of benchmarks and marketing-driven claims (e.g., "we haven’t seen any of these copy cats play doom or street fighter"; a benchmark would give concrete grounds for discussion).
  • Enables practical utility by letting anyone quickly see how a new model stacks up against baselines without reinventing evaluation.

StructuredLLM: Schema-Guided Output Library

Summary

  • Library that wraps any LLM (local or API) with a JSON schema to enforce valid structured outputs, retrying on invalid results and reducing hallucination.
  • Core value proposition: type-safe, reliable outputs for AI agents and classification pipelines without prompt engineering guesswork.

Details

| Target Audience | Developers building AI applications needing reliable structured outputs (e.g., tool calls, classification) | | Core Feature | Schema validation via Pydantic, constrained decoding logic (similar to outlines/lm-format-enforcer), automatic retries, confidence scoring | | Tech Stack | Python, Pydantic, optional Rust core for performance, compatible with Hugging Face Transformers and vLLM | | Difficulty | Medium | | Monetization | Hobby |

Notes

  • HN users highlighted the need for "type-safe outputs" and hallucination mitigation (e.g., "cjalmeida: You can use structured outputs and validate them against a schema today").
  • Addresses frustration with LLMs hallucinating tool calls or invalid JSON, providing a practical tool that would earn upvotes for its utility.

ClassifyFast: Cheap Zero-Shot Classification API

Summary

  • Pay-per-use API exposing open weight models (like Laya/Jev equivalents) for zero-shot classification tasks, with optional private fine‑tuning per customer.
  • Core value proposition: sub‑100 ms latency, CPU‑friendly inference, and predictable pricing for teams that want Jev‑class performance without managing GPUs.

Details

| Target Audience | SaaS builders, indie hackers, enterprises wanting cheap classification without GPU ops | | Core Feature | REST endpoint accepting text + label list, returns calibrated probabilities; supports ONNX Runtime CPU inference; optional LoRA fine‑tuning via private endpoint | | Tech Stack | FastAPI, ONNX Runtime or TensorRT, Hugging Face model zoo, Docker/Kubernetes (or AWS Lambda container images) | | Difficulty | High | | Monetization | Revenue-ready: $0.001 per 1K classifications (or tiered monthly plans) |

Notes

  • Many commenters praised Jev’s low cost ("fastball: Jev is $42/Gt vs Gemini 2.5 Flash Lite $500/Gt") and wished for an API ("jwpapi: Where can I subscribe to a hosted version of this?").
  • Provides a ready‑made solution that sidesteps GPU maintenance and offers the zero‑shot convenience users requested.

AI Research Launchpad: From Paper to Product

Summary

  • Template repository and guide that helps researchers turn a paper into an open‑source project with model card, API demo, benchmark, and licensing.
  • Core value proposition: reduces the friction of sharing work, giving researchers the visibility and credit they deserve.

Details

| Target Audience | Academic researchers, ML practitioners who want to share work and get recognition | | Core Feature | Cookiecutter template generating README, model card, inference script, benchmark script, Dockerfile, CI workflow | | Tech Stack | Cookiecutter, Markdown, Python, GitHub Actions | | Difficulty | Low | | Monetization | Hobby |

Notes

  • OP expressed frustration about lack of recognition despite prior work ("nandakishor_ml: I built a pypi for it called hallunox"; "wild_egg: Last time I did anything with a BERT, you had to train or fine‑tune").
  • A launchpad would let researchers like the OP easily publish reusable code, attracting citations and community feedback—exactly the kind of tool HN’s research‑oriented audience would upvote.

SheetSense: AI‑Powered Classification Add‑In for Excel/LibreOffice

Summary

  • Add‑in that lets users define classification labels and run a local small model (like Laya) to tag cells based on text in other columns, requiring no internet or LLM fees.
  • Core value proposition: brings zero‑shot classification to everyday spreadsheet workflows, enabling analysts to categorize data instantly and privately.

Details

| Target Audience | Business analysts, ops teams needing to categorize data (e.g., support tickets, survey responses) | | Core Feature | Sidebar UI to select model, input columns, label list; runs inference locally via ONNX Runtime/WebAssembly; writes results to new column | | Tech Stack | JavaScript/TypeScript Office Add‑in, ONNX Runtime Web (or Pyodid), optional Python backend for model loading | | Difficulty | Medium | | Monetization | Hobby |

Notes

  • Commenters explicitly asked for an Excel/LibreOffice add‑in ("badatnames: This is crying out to become an Excel or LibreOffice Calc add‑in").
  • Solves the pain point of sending data to costly LLMs for simple tagging, offering a private, CPU‑friendly alternative that fits naturally into existing workflows.

Read Later