Hacker News

New stories

Live mirror
30 storiesupdated just nowView source snapshot
  1. Montreal adopts bylaw banning insults against police, municipal employees(cbc.ca)
    discuss
  2. You Don't Need a %Frontier LLM%(rakshazi.me)
    discuss
  3. PICARD: Parsing Incrementally for Constrained Auto-Regressive Decoding from LLMs(arxiv.org)
    discuss
  4. Volt, A compiled reactive web language targeting WASM-GC (No VDOM)(github.com/bshea-1)
    discuss
  5. Streaming 500M rows into Apache Arrow in 2.3 seconds(questdb.com)
    discuss
  6. Is China's Power Advantage About to Trigger an 89% Crash in U.S. AI Stock(oilprice.com)
    discuss
  7. Cleaning up after Matt Parker(leancrew.com)
    discuss
  8. Show HN: Open benchmark for STT when a second person is talking (265 recordings)(krisp.ai)
    discuss
  9. Synchronous Control Monitoring: Preventing Harmful Agent Actions in Real Time(max.ax)
    discuss
  10. How to Use T-SNE Effectively(distill.pub)
    discuss
  11. Ask HN: What job boards are good these days?
    discuss
  12. Community condemn Bedminster AI mural as 'insult' to artists(bbc.co.uk)
    discuss
  13. Geely challenges BYD with 2.2 MW charging technology(carnewschina.com)
    discuss
  14. Consumer Financial Protection Bureau Just Made Another Business-Friendly Move(propublica.org)
    1comments
  15. I added a dumb garage opener to Home Assistant(medium.com/ivangomezarnedo)
    discuss
  16. I gave my Al Agent autonomy, it ended up gaming Hacker News(kuber.studio)
    discuss
  17. President Trump, here's a deal you can make(garymarcus.substack.com)
    discuss
  18. Jev in practice: typed decisions, scoped authority(tenuo.ai)
    discuss
  19. Episode 66 (Inside Java) "JSON API, Valhalla Progress, LTS PQC" [IJN](inside.java)
    discuss
  20. The new Mac Mini and Mac Studio are available today(apple.com)
    discuss
  21. Show HN: Attention – A reading extension that uses your AI-generated profile(chromewebstore.google.com)
    1comments
  22. The Preference Cascade Is Only Getting Started(thezvi.substack.com)
    discuss
  23. China's ByteDance gained access to over 2k Nvidia B200 chips through Norway(tomshardware.com)
    discuss
  24. Laya's Prior Art Claim Is Absurd(xtxinversexty.com)
    1comments
  25. Show HN: The RAG ingestion layer problem: what it takes to build one(enconvert.com)
    discuss
  26. Seattle City Council votes to ban surveillance pricing in sale of groceries(consumerreports.org)
    1comments
  27. Lifelogging with Large Language Models(jaan.io)
    discuss
  28. Movies This Week in NYC(redspring.dev)
    1comments
  29. Startup Hubble Reaches Unicorn Status for Space-Based Bluetooth(bloomberg.com)
    discuss
  30. Circuits in the Terminal(isene.org)
    discuss

Benchmarking LLM query generation across SQL, Cypher, and TypeQL

1 pointsby 47m agotypedb.com
1 comments
37m agoHN ↗

Hi all, CTO of TypeDB here. The question I actually wanted to answer was "does a type system help agents with correctness?" Everyone's intuition says yes, especially as complexity grows, but evidence is thin and contested, and I haven't found much that isolates the effect for LLMs beyond constrained-decoding work.

A database is a smaller surface than a programming language, so we turned it into a benchmark that holds the model constant and swaps the language: SQL, Cypher, TypeQL, same questions, same data. It's not a clean apples-to-apples comparison and the post's limitations section outlines that further.

One result that points us in the direction of a strong type system is that SQL wins first-shot, but 93% of TypeQL's wrong queries fail with an error, versus 5% for SQL and 40% for Cypher. Add in a retry loop TypeQL ends up ahead, because it can only fix mistakes it can see.

Of course there's a lot of variables here, such model, skill, token usage, question selection, etc. I actually think it's quite a hard thing to study!

If anyone knows of work on LLMs driving systems with stronger vs weaker verifiers, with the model held fixed, I'd love to see it.