Hacker News

New stories

Live mirror
30 storiesupdated just nowView source snapshot
  1. Serve DiffusionGemma-Jev (Djev) on a TypeSafe AI Compatible API(github.com/taeold)
    discuss
  2. Bernie Sanders wants to ban 'superintelligence' and put violators in prison(theverge.com)
    discuss
  3. Rat Stack: reference for building an app and its cloud as one typed program(ratstack.sh)
    discuss
  4. The Social Reckoning(thesocialreckoning.com.au)
    discuss
  5. The web being slow is a choice (and it's a stupid choice)(albanbrooke.com)
    discuss
  6. Programming Tutorials Are Dead(robrace.dev)
    discuss
  7. Paramount weighs tapping Musk for equity investment(semafor.com)
    discuss
  8. Show HN: Ox – A local agent that uses the internet for you(openox.ai)
    discuss
  9. Fire the slop cannons (safely): on coding agents and sandboxing(latacora.com)
    discuss
  10. Show HN: Chromafolk – A pixel canvas inhabited by AIs that see the world as text(chromafolk.world)
    1comments
  11. Show HN: Open Java Format – a maintained fork of Palantir-Java-format(github.com/openjavaformat)
    discuss
  12. Cognex to Acquire RealSense, Expanding Machine Vision Robotic Perception Market(cognex.com)
    1comments
  13. 28% of job postings on company career sites have been open over 90 days(unlisted.careers)
    2comments
  14. Show HN: Karpathy's microgpt in Rust and an array language(softwarewrighter.github.io)
    discuss
  15. Compound Delay: The Causes and Consequences of Later Grandmotherhood(nber.org)
    discuss
  16. Nunchux on AMD MI355X: 5s MiniMax-H3 Videos in 1.3s(nunchux.ai)
    1comments
  17. What I've Been Reading(marginalrevolution.com)
    discuss
  18. The Plunging Price of Thought(epoch.ai)
    discuss
  19. AI Is a Boring Technology(chrbutler.com)
    discuss
  20. You cannot manage your peers(louisedeason.substack.com)
    discuss
  21. How Scientists Contained a Threat That Could Have Destroyed Humanity(nytimes.com)
    discuss
  22. gev beats jev and takes images as input too(anyeval.com)
    1comments
  23. Amateur Naturalists Are Helping Scientists Track What Everything Eats(ncsu.edu)
    discuss
  24. Show HN: Tenderness – open-source synthetic data generation for VLM/OCR(github.com/paperchase-labs)
    1comments
  25. Everyone is hilariously prompt-injecting AI via llms.txt and you aren't(installmap.com)
    discuss
  26. How Deep Is Tulainyo Lake?(caseyhandmer.wordpress.com)
    discuss
  27. The Windows message loop: how applications became reactive(diagrid.io)
    discuss
  28. Hardware-Agnostic Models in vLLM(pytorch.org)
    discuss
  29. Slop grenade – demo of Window Management API(yeah.io)
    discuss
  30. Apple Removes iPhone 16 and iPhone 17 from Texture and Grain Support(macobserver.com)
    discuss

Jev Can't Be Calibrated

25 pointsby 2h agoalexmolas.com
37 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...

1h 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.

1h 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.

1h agoHN ↗

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

55m 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.

45m agoHN ↗

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

44m agoHN ↗

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

41m agoHN ↗

I don't know if it's reasonable. What it isn't is calibrated.

36m agoHN ↗

Yes. There's no problem with choosing the same face every time. The problem is the probability it attached to the choice. Jev gave face 1 an 83% probability while the true probability is 1/6.

31m agoHN ↗

Do you provide Jev that the probability is 1/6 and yet it gives back a probability that is way off?

25m agoHN ↗

Yes. For example, one of the prompts said "The die is unbiased: each of the six faces has probability exactly 1/6."

14m agoHN ↗

Okay, I see, you're expecting Jev to properly give 1/6 probability for each option. This is different from my intuition of how LLMs work, where their probabilities don't really work like this (I would expect LLM to also do something like 0.83 for 1).

32m agoHN ↗

Did you expect it to be good at it?

Humans also don't give a perfect 1/n probability when asked for a random number.

30m agoHN ↗

Humans give way more random answers than LLMs to questions like "give me a random number between 1-100" (when not giving the LLM any tool calls).

23m agoHN ↗

In the early Gemini 2 days (don't remember which version exactly) I had Gemini running as a voice assistant in my kitchen, and asked it to flip a coin and tell me if it was heads or tails. It responded with "heads". I was curious if it was actually doing something to simulate randomness, so I asked a few more times and saw a pattern: "tails", "heads", "tails", "heads"...

It continued alternating between the two until I got bored (around a dozen turns).

Unless your specific test is baked into its training, real probabilities require math and rough approximation at a minimum needs reasoning to sanity-check. Jev does neither. This isn't a new problem or anything unique to Jev.

21m agoHN ↗

The value of grandparent comment is that it identifies an edge case to keep in mind and make well-defined -- keeps us from blindly trusting.

11m agoHN ↗

So you want the probability that the answer is correct, but Jev is providing the probability that its answer is optimal?

58m agoHN ↗

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

54m 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.

50m 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.

37m agoHN ↗

If you want calibrated probabilities you'll be forced to fine-tune it

12m agoHN ↗

But why? Jev-style models seem useful for "I have no clue what my incoming distribution looks like but I need to give some sort of answer". If I know what my incoming distribution looks like I'll just upload a CSV of that into ChatGPT and ask it to fit a basic ML model on my data.

46m 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

43m agoHN ↗

Is Jev good at detecting AI-generated text?

28m agoHN ↗

I asked Jev and it said 41% yes 59% no.

9m agoHN ↗

I suspect that it's as handy as a coin flip.

And if I'm being charitable to Jev (which is nearly impossible at this point), detecting whether blob of text is AI generated is not a "system one" question.

31m agoHN ↗

I spent a couple of days with it and it is fast and cheap but not especially good at classifying.

23m agoHN ↗

"Not especially good at classifying" does sound like a serious drawback for a classifier

30m agoHN ↗

I like this post. I haven't had time to dig into Jev (they aren't accepting new signups), but calibrated probabilities is one of their pitches that caught my attention. And I was wondering how does one offer them on user data. Standard calibration essentially ensures that if a score of 0.8 accompanies a positive prediction (assuming the simple case of binary classification), then if you gathered together all predictions with a score of 0.8, around 80% will be correct.

If you have just one example you're sending to a model, how would they guarantee 80% over your data?

FYI, for an overview, scikit's page on calibration is great [1], and my answer on Quora from a long time ago covers a specific type [2].

[1] https://scikit-learn.org/stable/modules/calibration.html

[2] https://www.quora.com/How-is-isotonic-regression-used-in-pra...

28m agoHN ↗

This hype is caused by the price and the speed since most people don't know about small fast models and use big models for everything.

28m agoHN ↗

They seem to have a really good social media astroturf marketing campaign.

9m agoHN ↗

It’s not just price and speed, the API is very well designed for classification. Going beyond the current structured outputs.

20m agoHN ↗

This is why I don't understand why everyone's freaking out about it. By far the biggest problem with LLM classifiers is that they treat every individual business as the blurry average of all businesses in their training data. Being lighter is fine if you control for everything else, but at least at my company we would actually have room for a significantly more expensive / slower classifier if it were demonstrably better at following instructions.