Hacker News

New stories

Live mirror
30 storiesupdated just nowView source snapshot
  1. OAuth Token Theft Through Microsoft's Front Door(huntress.com)
    discuss
  2. Apple Reference Image Explained Through Anti-Doping(medium.com/the-quantastic-journal)
    discuss
  3. Maynooth university: MU researchers build world-first DNA computer(maynoothuniversity.ie)
    discuss
  4. ComfyUI launches OpenRouter for generative media(comfy.org)
    1comments
  5. Show HN: PicoLM v1.0-rc2 ("Yura Kana"). Run an LLM on Digital Unix(github.com/whoreson)
    discuss
  6. A Startup Wants to Power Data Centers with 'Supercritical' Carbon Dioxide(wired.com)
    discuss
  7. We open-sourced an event ticketing platform(evnelo.com)
    1comments
  8. Functionally Zen(testdouble.com)
    discuss
  9. Trump reveals millions of dollars' worth of share deals in big tech and AI(bbc.com)
    discuss
  10. OpenAI nabs key Patreon execs ahead of upcoming announcement(theverge.com)
    discuss
  11. Open-weight models now carry 56% of production tokens and 14% of the spend(fromtheterminal.substack.com)
    discuss
  12. Show HN: Volum – An open-source visual library for 3D model files(didac.dev)
    discuss
  13. iPhone 4 "Antennagate" Q&A – Steve Jobs, Tim Cook and Bob Mansfield (2010) [video](youtube.com)
    discuss
  14. Containers Are No Longer a Security Boundary(depthfirst.com)
    discuss
  15. Data Centers Are Theft(machinesociety.ai)
    discuss
  16. Is AI killing the forums and wikis your IT team relies on?(spiceworks.com)
    discuss
  17. The science of Monkey Island: can grog dissolve a metal mug that fast?(jgeekstudies.org)
    discuss
  18. Top post in R/gamedev: "AI models have caught up with Unity dev."(reddit.com)
    1comments
  19. We ported the original Doom to SQL(cedardb.com)
    1comments
  20. Show HN: A browser word game about composing Chinese characters(chunqiuyiyu.com)
    discuss
  21. What's Inside Fauci's Diary [video](youtube.com)
    discuss
  22. Systemd v262 Released(github.com/systemd)
    1comments
  23. AI anxiety is sweeping the globe – and cutting across party lines(politico.eu)
    discuss
  24. Linux Implements Dynamic Bash Tab Completion(salivity.github.io)
    discuss
  25. I Couldn't Build Jev at OpenAI – Diogo Almeida, TypeSafe Co-Founder and CEO [video](youtube.com)
    discuss
  26. Performance and safety of a multi-cancer early detection test(nature.com)
    discuss
  27. Ethics of Agent Exploitation(cygankiewicz.com)
    discuss
  28. Build a devtool once. Mount it anywhere(devfra.me)
    discuss
  29. Troubleshooting a weird keyboard on a Pentium III(rubenerd.au)
    discuss
  30. Show HN: a Rust crate to mock concrete type without using trait(github.com/xtsoftwarelabs)
    discuss

Jev Can't Be Calibrated

9 pointsby 1h agoalexmolas.com
12 comments
1h agoHN ↗

I tested Jev with a fair die 400 times without telling it the die result. The true probability of face 1 is 1/6, but Jev always chose face 1 and the probability it returned was about 83%. I also tested with a fair coin 200 times and got 0.92 probability.

I did several tests and I think Jev is good at problems with a correct answer but weak at problems about actual probabilities whose answers can't be known at all.

Write-up: "Jev Does Not Play Dice" https://kantahayashiai.github.io/posts/jev-does-not-play-dic...

57m agoHN ↗

But "problems about actual probabilities whose answers can't be known at all" are exactly the problems where calibration is important. Since calibration is one of the big claims about Jev I'd expect it to perform well in these problems.

28m agoHN ↗

I agree. I think it's odd behavior too. Jev should be good at actual probability problems given the phrase "calibrated probabilities" TypeSafe uses for Jev. Maybe the reason is the data used in their training method (RLCD). If all the data consists of problems with a correct answer, I think this kind of odd behavior could happen.

15m agoHN ↗

The reason and the word you're looking for is mode collapse. It happens during RL, it's not a dataset problem, and it's the reason for many issues with LLM writing and all sorts of issues in all types of models. The result you got is 100% predictable and unsurprising, as it's a neglected problem nobody really rushes to solve as current big model training is laser focused on one thing and one tool.

23m agoHN ↗

Echoes a bit of a philosophical distinction with a long history: "Knightian Uncertainty" versus "Probability".

15m agoHN ↗

Hah! I did the exact same tests as you! I found that if you give it the choice to say "not sure", it picks that 100% of the time. But if you pin it in a corner, then yes it does these weird things. Also yes, the continuous options were much more accurate than the choices. Not sure why that is.

5m agoHN ↗

If you instead offer probabilities as answers, it picks the right one with high credence.

4m agoHN ↗

Maybe I’m confused here, but it’s perfectly reasonable to just guess the same dice roll every time right?

18m agoHN ↗

In future, we will see intiatives similar to OpenStreetMap for Textual data or Web similar to high quality non-contaminated steel.

15m agoHN ↗

I'm really looking for a multi-modal image capable version of Jev.

If we could get machine learning type results on images without training, that would be fantastic.

10m agoHN ↗

Fine tuning LLMs has turned out to be mostly not worth the effort, but I wonder if fine tuning Jev-style models will turn out to be a whole lot more useful.

6m agoHN ↗

I’m using Jev to classify a blob of text I see in browser with an extension. Calibrated? No. But it’s handy enough. If a large blob of text is likely AI generated, I’m very likely to skip it