frontpage.
newsnewestaskshowjobs

Open Source @Github

fp.

Open in hackernews

Jev in 25 Lines of Python

https://www.nobodywho.ai/posts/jev-in-25-lines/
57•bashbjorn•1h ago

Comments

heaney-555•42m ago
Latency and compute comparison needed.
thephyber•20m ago
Is benchmarking Jev still a ToS violation?
ricardobeat•18m ago
Was it? That would make it unusable in any corporate setting.
dhsysusbsjsi•41m ago
Whilst I do like reading these things for technical know how, I can sympathise with the creator of jev who now presumably has to apply an order of magnitude effort to explain why the 100 smaller things done better than this add up to a much better product.
jpnc•30m ago
Replace 'explain' with 'sell'. Don't forget that it's a gold rush. There's no reason to sympathize with corporations in their rush for the slice of the pie.
no-name-here•36m ago
Beyond the missing latency and compute comparisons that Heaney commenter mentioned, also nothing about its error rate compared to Jev (nor if it even always outputs in a format the app can parse, not sure how solved that is).

But then at the end it says it’s parody. Maybe HN title should say it’s a joke.

est•29m ago
latency and compute comparisons highly depends on your local setup.

you can swith to a better model for lower error rate.

ricardobeat•21m ago
Which massively slows down the output. Doing this with Qwen 9B already takes you into seconds per answer territory, and Jev is supposedly frontier level intelligence.
sigmoid10•32m ago
Going directly for the logprobs is always icky when you use a chat model as base, because they are trained to write prose as output. So your "choice" tokens and thus their probabilities might get diluted in whatever else it wanted to say. If you have to do it in the same way as this post, at least add clear system instructions and a carefully worded beginning to the assistant output section of the prompt to lower the chances of it wandering off immediately.

I've found that using structured outputs solves this problem much better. Instead of letting a model generate only "A", "B" or "C" and looking at the probs, have it directly generate "Legitimate", "Spam" or "Phishing" or any other pre-defined option from a set of multi-token sequences. Behind the scenes it boils down to something quite similar, but you're not running into the risk that the model actually wanted to say "A phishing attempt seems likely, so answer (C) is correct.", which would lead "A" to have the highest probability in the first token. You can even use a reasoning budget this way either via inherent reasoning or a free-form part preceding the remaining output structure. You can also have it assign probabilities (either in words or numbers) using more complex output structures, but I would not rely on them much more than the token logprobs (they can still be quite good though).

_flux•22m ago
Seems like all normal english words could risk the same, so would using short but random strings be even better?

Actually to me it sounds it could be benchmarked if this kind of effect exists in the first place.

sigmoid10•17m ago
Best option would be reasoning + clear system instructions + constrained output. That is, if you have to use a chat model. Which works well enough to be sure, but hey I haven't tried raising millions of dollars when I did that 3 years ago. But perhaps I was the stupid one.
_davide_•
teaonly•31m ago
The principle is this.
ricardobeat•23m ago
Now, can you do it in <200ms for 45 questions at once, have 0% malformed output, and any kind of meaningful benchmark? We’ll wait!
onion2k•23m ago
It's fast.

If you're comparing with something, you need to state 'fast' in relative terms. Jev is definitely fast, and if this Python takes the same time to get a decision then it's also fast. If it's 100* slower than Jev though, you shouldn't be calling it 'fast', because relatively speaking it's really, really slow.

_davide_•16m ago
By design it can't be significantly slower than Jev: the prompt processing (AKA PP) is exactly the same on both and will take most of the time. Then you can process every single "question" in parallel, just predicting one or two tokens (if an answer is ambiguous with a single token) per each question, again in a single batch.

So, fast in the LLM space and comparable with Jev.

brap•9m ago
What I don’t understand is, why would you not want “reasoning” in a classifier?

Speed and cost are obvious reasons, but isn’t this a tradeoff?

cupofjoakim•4m ago
I wonder if this could be a good stepping stone to write a local prompt router to optimise what model get what prompt. I.e. if the prompt is just a lookup, send it to haiku, if it's reasoning, send it to opus and if it's implementation send it to sonnet.
iLoveOncall•1m ago
Nothing I hate more than bullshit articles claiming X in Y lines of code, only to use libraries abstracting hundreds of thousands of lines of code.
21m ago
Agreed, it's a real issue, but it can probably be vastly reduced by having the schema in the system prompt and by giving the model an expectation of a fixed value: no decent modern would pick a prose ligament over a provided value.

To completely squash the issue, a few cheap LoRa iterations will do the trick just fine.

wongarsu•6m ago
Sure, you can fix that in a couple lines. Then a couple more lines for evaluating multiple questions on the same answer in parallel. Then a couple more lines for the confidence score (which is trivial to compute from all we have, but missing regardless). Then a harness to fine-tune an existing model to perform better on this specific task, and a collection of training data to use for that

I think we can all agree that Jev is not rocket science. It's a good idea executed well, with marketing that might have been a tad too bold

ainch•18m ago
In my experience as well using logprobs to try to quantify uncertainty, LLMs are a poor fit. Neural nets in general struggle with 'calibration' --- ie. if a prediction is truly 50/50, neural nets are often prone to predicting overconfidently [0].

I ran some tests using GPT-4 to do some basic classification a couple years ago. On ambiguous options which had to be escalated to a human, the LLM would regularly output something like a 99.8% probability, compared to 99.99% for a correct answer.

0: https://arxiv.org/pdf/1706.04599

nautilus50•18m ago
+1, llama.cpp has a --grammar parameter which you can pass a BNF style grammar file to constrain generation. It can be used in Python llama.cpp wrapper

https://til.simonwillison.net/llms/llama-cpp-python-grammars

Jev in 25 Lines of Python

https://www.nobodywho.ai/posts/jev-in-25-lines/
58•bashbjorn•1h ago•24 comments

GPT-6 Sol and Luna

https://openai.com/index/introducing-gpt-6-sol-and-luna/
1512•OfficialTurkey•14h ago•713 comments

Claude Opus 5.5

https://www.anthropic.com/claude-opus-5-5
1518•km144•16h ago•935 comments

Transit rewards

https://waymo.com/blog/2026/09/transit-rewards/
127•raybb•5h ago•128 comments

OpenAI GPT–6 Astra breaks Enigma message that has resisted solution since 2005

https://www.cryptocellar.org/bgac/the-mvueh-break.html
652•sohkamyung•18h ago•385 comments

Data-only attacks are easier than you think (2024)

https://www.usenix.org/publications/loginonline/data-only-attacks-are-easier-you-think
48•segfaultbuserr•4h ago•14 comments

'We hacked the FBI:' Hackers say they have data on all FBI employees

https://www.404media.co/we-hacked-the-fbi-hackers-say-they-have-data-on-all-fbi-employees/
591•spenvo•15h ago•417 comments

Show HN: Npunlock – Run custom C kernels for Intel NPUs

https://github.com/hsfzxjy/npunlock
18•hsfzxjy•19h ago•3 comments

Microsoft killed FoxPro in 2007. Anyway, here's FoxPro revived

https://foxscript.org/
303•boredjohnny•11h ago•181 comments

ReBarUEFI: Resizable BAR for almost any UEFI system

https://github.com/xCuri0/ReBarUEFI
143•nateb2022•2d ago•45 comments

What California is learning from solar panels built over irrigation canals

https://www.kqed.org/science/2002033/heres-what-california-is-learning-from-solar-panels-built-ov...
208•Jtsummers•1d ago•357 comments

How did AMD Ryzen get 50% faster in two years?

https://lemire.me/blog/2026/09/18/how-did-amd-ryzen-get-50-faster-in-two-years/
320•ibobev•4d ago•122 comments

SAML: A fractal of bad design

https://blog.trailofbits.com/2026/09/21/saml-a-fractal-of-bad-design/
232•aray07•13h ago•137 comments

WordPress: Unauthenticated path traversal leading to conditional RCE

https://github.com/WordPress/wordpress-develop/security/advisories/GHSA-7hp8-65ch-5whp
189•vntok•16h ago•97 comments

Claude Opus 5.5 Intelligence, Performance and Price Analysis (Max)

https://artificialanalysis.ai/models/claude-opus-5-5
286•theanonymousone•15h ago•89 comments

No Easy Fix for Bogus Respondents in Online Opt-In Polls

https://www.pewresearch.org/methods/2026/08/27/no-easy-fix-for-bogus-respondents-in-online-opt-in...
26•luu•1d ago•7 comments

Pentagon says overreliance on AI contributed to missile strike on Iran school

https://www.bloomberg.com/graphics/2026-iran-school-attack/
629•devonnull•13h ago•316 comments

Unreal Agent

https://unreallabs.ai/blog/unreal-agent/
177•trollied•14h ago•105 comments

How often do you think about the 1893 World's Fair?

https://www.thebirthofacapital.info/chicago-worlds-fair-tataria-ware-larsen/
8•bryanrasmussen•3h ago•1 comments

People hooked on vapes try a new way to quit: cigarettes

https://www.bloomberg.com/news/articles/2026-09-18/to-quit-vaping-some-are-starting-to-smoke
175•alephnerd•1d ago•193 comments

The softness of metal

https://psyche.co/turning-points/his-frailty-made-ozzys-final-gig-true-heavy-metal
30•NaOH•1d ago•11 comments

The current balance of power in open models

https://www.interconnects.ai/p/the-current-balance-of-power-in-open
86•gmays•10h ago•29 comments

Obscura: VPN that can't log your activity

https://obscura.com/#faq-technical
130•Flimm•13h ago•106 comments

Grammarly will send unhinged messages to all your users if you try to cancel

https://www.reddit.com/r/sysadmin/comments/1wjdpgx/psa_grammarly_will_send_unhinged_messages_to_all/
127•ksec•4h ago•32 comments

Show HN: JevBench, a reproducible benchmark for typed decision models

https://benchmarkheaven.com/jev-models
102•florianstandhar•19h ago•25 comments

OpenAI is well positioned to fast-follow Jev

https://arcturus-labs.com/blog/2026/09/21/will-openai-eat-jevs-lunch/
292•JohnBerryman•18h ago•208 comments

Side-stepping the Secretary Problem, unwittingly

https://www.evalapply.org/posts/side-step-secretary-problem-hiring/index.html
79•pvdebbe•1d ago•13 comments

Delta: Highly available, strongly consistent storage using chain replication (2022)

https://engineering.fb.com/2022/05/04/data-infrastructure/delta/
25•grep_it•1d ago•1 comments

Markdown in /src

https://htmx.org/essays/markdown-in-src/
131•perrygeo•1d ago•69 comments

Apple has added persistent 'ads' to iOS, and it's driving users crazy

https://www.techradar.com/phones/iphone/i-wish-apple-would-just-stop-that-crap-apple-has-added-pe...
711•MC995•18h ago•518 comments