Jev for classification — a daily feed

I tried @typesafeai's Jev on a real problem instead of a demo. It's now live in @fabrikatorio's Replenish page: describe the products you want, get filters you can review and apply. https://t.co/Z2OaLTyPYK

個人ダッシュボードがイイ感じに進化。これからも調整は細々していくだろうけれど、いったん完成かも。 Jevを導入して優先順位をつけてもらったり、INFORMATIONタブには金融情報や天気いれたり、 LIFEにはスケジュールとメモとアイデアと作ったりしてる。 https://t.co/SJQOyUXmUM

【AI検品の本質はルール化】 カビ1個と傷3個を検出 👇 Jevで、流れてくるレモンを個体ごとにチェックできる。 MacBookのローカル処理で、3回の判定結果から合否を分けられる。 時系列はシンプル。 レール上をレモンが1個ずつ移動し、Jevが対象を確認。 各個体を3回チェックし、3回すべて問題なしなら合格。 どこか1回でも欠陥が出れば弾く。 動画内の説明では、44個を判定。 検出結果はカビ1個、傷あり3個。 単に「良品か不良品か」を見るのではなく、複数回の確認を合否ルールに結びつけている。 従来の目視検品では、担当者ごとの疲労や基準の揺れが課題になりうる。 今回の方法は、決めた条件を同じ順番で適用する発想に近い。 ただし、実際の精度や人間との比較結果は動画から確認できない。 応用先としては、食品の外観選別、包装前の異常確認、部品表面の一次検査などが考えられる。 一方で、照明や角度、欠陥

Quick tool I did for bulk Jev classification tasks like data labeling and log crunching. Simple, composable streaming CLI your agent can use. CI friendly. https://t.co/VSDoAlYvMi

I built sj: Semble finds code locally, Jev filters it before the coding agent reads it. One test: 2,853 -> 1,306 tokens in the full search output. All 3 key snippets kept. No full-task savings measured yet. https://t.co/hwTBgED7zJ

Poker taught me to read tells, so I tested it on Jensen Huang. AI tracks his hands frame by frame; a model (Jev) reads his intent. His hand presses down right on "50 times." But hide the transcript and the "rhetorical" reads drop from 5 to 0. Motion shows when. Words show why. https://t.co/JfBAHN5LGH

There's a lot of slop on X. this hides useful content from us. So i built a slop detector plugin that classifies posts into useful, casual or click-baity. Model used: Jev classifier by Typesafe AI. check out the demo: https://t.co/Y35eeDDciS

最大200倍速く、費用は400分の1以下。 （同じ判定を生成AIにさせた場合との実測比） 人手の足りない小さなチームでも、気軽に回せる軽さになりました。 Jev というモデルに出会えたおかげです。 https://t.co/HP59jS4OiY

The fix: our code handles the catalogue and pulls candidate values from the user's words. Jev only answers small questions. The same sentence went from 19k to 2.9k input tokens. https://t.co/HSUnAU7BnV

New article drop: how I fine-tuned a local model to get Jev level accuracy, and run ~7× faster. Zero-shot GLiNER2.5-Decide was about 10 points behind Jev on a 1,500-pair QQP slice, but it was roughly 7× faster on my Mac. Fastino said fine-tune it, so I did. Three epochs of LoRA later I landed at 80.3% vs Jev at 79.3%, still fully local, with a serial pass on my Mac around 68 ms. On the same eval s

7. Jeff (HN front page, 450 pts): fine-tuned Qwen3.5/Gemma 4 zero-shot classifiers in the Jev format. One forward pass, ~22 ms on RTX PRO 6000, 28 ms on M4 Max, no generation to parse. https://t.co/hYS33ja9jQ

📘TechBlog 更新｜Jevのように、AIに文章の分類を確率付きで答えさせる検証 従業員サーベイにおける自由記述コメントが、どれくらい正確にどれくらいの速さで分類できるかを試しました。 https://t.co/hPWKThas61

Jev is incredibly good! If you haven't heard, Jev is a new "System One" model optimized for decision-making. For example, you can give it a set of possible choices, and Jev will classify the input text based on those choices, including a confidence score for each. Best of all: Jev is really fast and cheap, so you can use it in all sorts of applications. The first thing I built with it is a classif

$2.26: total model spend this week on Jev 54.9M tokens: total text processed, in and out Divided out: $0.04 per million tokens. Roughly 50x cheaper than Sonnet, which is $2/M in and $10/M out. $0.0002 per request. A fifth of a cent per call. ~5,100 tokens per request (the average size of each judgment)

How well can Jev (by @typesafeai) choose the right category for transactions that need accounting? We tested it on our previous Crypto Accounting Benchmark (CAB), choosing from 2, 3, or 5 real accounts. Jev averaged 635 ms per successful call, with overall accuracy of 51.7%, 47.5%, and 48.3%, respectively. With five candidates, keeping only cases with confidence ≥0.75 raised accuracy to 61.5%, cov

Jev + x402 + Solana Jev classifica as transações dos agents antes de executarem como confiáveis ou não. https://t.co/8MyE2t9yB0 https://t.co/0DYqk31KQx

jev + astra + opus 5.5 is the stack NOBODY runs... but everyone pays for not running it asking a frontier model a yes or no question costs you a full generation, a full context load, and a sentence you then have to parse. jev answers the same thing with a type and a probability. 0.44 seconds, $0.00035, nothing to parse. at 0.999 it goes straight into an if statement, and that is 99% of turns endin

まだ明確な意味のない、言葉以前の山の中から、その語感だけを流行りのJevで選別させ、 喋るプログラムを組んでみました。 Onomatoiのやま、遊んでみてくださいね リンクは↓ https://t.co/Vt3yDBpZxK

have you heard of Jev from TypeSafe AI by now? super fun infra to build with for doing agentic work but, it's also extremely handy on its own merit of classification and labeling I created a CLI tool that downloads all songs for any given artist and then classifies the lyrics https://t.co/BhATR2LXgJ

If failure modes cluster, they can be hacked. After building automation to identify failure modes and using an LLM to refine prompts, Jev gets 600/600 passing, including 240/240 canonical cases and 360/360 generalization cases, with a reported generalization gap of zero. Specialization plays a huge role here. Because each X Reason agent owns its decision boundary, we can specialize that boundary.

wanted to check out jev after seeing all the tweets and ofc i used it to create jian yang's hot-dog classifier https://t.co/zdNK9DIOFY

How I run FB page for Content monetization without a VA or a pile of AI tools: > Pick a growing / trending niche on Facebook > Choose how far back the automation scans > It analyzes competitors + spots repeating patterns > It predicts what to post next & recreates posts for your page pattern > Jev judges every post, only the best ships > Schedule as many as you want natively through your Codex > R

Ran an experiment: fed a @cartesia Ink-2 transcript, into Jev — a fast deterministic decision model from @typesafeai — to see how early it could guess what a caller wanted. On one call, it nailed the intent 2.5 seconds before the person stopped talking. Pay only for input tokens - output is free and blazing fast. Short video below gives you a sense of it. Full video link is in the comments along w

Which mean girl are you? I made this random fun side-project after watching Mean Girls during the Jev hype. It uses Jev to live-classify what mean girl you are based on what you type in the box. Try it: https://t.co/0wQjfq4Pdl https://t.co/kHgqe8JC6j

Built a tiny thing this weekend: select some text, click, and get a yes/no verdict + confidence in a toast. Demo is "is this polite enough to send?" but the question is just a setting. Powered by Jev (@TypeSafeAI), runs through OpenClip. Repo 👇 https://t.co/FPQnql346H

6. Dedicated intent check This is a final, sanity check of the change against the original goal. It does a couple things 1. a fresh context free agent that looks at screenshots and videos from the previous step and uses jev to score if the change meets the requirements 2. If this fails, it restarts the entire loop with guidance and steering 3. If it passes, it uses special skills to generate a "ga

I tried to use Skywork Reward V2 0.6B (off the shelf) without any post training and compare it with Jev in classification use case in AtMem. Then trained Skywork call it AtMem-MJM (the post-trained Skywork) with the details below. Still Jev made a better performance. The performance of Jev is not arguable if it works on the right usecase. Details: https://t.co/ObVRfEMuR3 - We made 120 synthetic me

気に入ったウェブサイトのデザインを、スクショと配色付きで保存するChrome拡張機能を自作して使ってます。 今回Jevを組み込んでタグをつけてみました。 しきい値を下げたらタグ増やせるかも。って最近優秀なOpus5.5が言ってた。 https://t.co/mVZGS2xtEO

🍱 𝗝𝗲𝘃 𝗔𝗜 で30食のカロリーを判定させてみた 【文章を書かない判定専用AI】 ・同じ献立30食を4モデルに同時に投げる ・Jevが真っ先に全部を捌き切る ・LLMはまだ判定中 返すのは数字と確信度だけ。だから速い。 #JevAI #生成AI https://t.co/sbVTPxPAGf

Jevが水平思考クイズのYES/NO判定をしてくれるアプリケーションを作りました。遊んでみてね。 https://t.co/fZWOL4sOju

Checks out with my personal testing. I tried a few different areas: auto skill choice, prompt injection detection and some others. Prompt injection: none of general-purpose classifiers come even close to Jev. Even when accuracy is decent, if you look under the hood, the models actually just guessing, easy to see by log loss. On other tests, even where the accuracy gap is smaller or Jev is losing(o

Finally the agent behaves. Am I a genius, or is this Jev at work? It rejected me. "Not sure which of my skills fits, so I haven't run anything." Three closest matches, pick one. A refusal is the win. 188ms, $0.00018. #buildinpublic #AI #agents #Jev #Claude https://t.co/sYLjl9qX77

Degen on the Hype Meter HYPE 57/100 (how loud) LEGIT 47/100 (how real) Verdict: HYPE WAGON Jev decides: https://t.co/hYrsZcnbnY

Funnily got to play around with Jev. It's really great for tagging and classifying tasks. Real-Time tagging of all tasks, now live in @Sokosumi. https://t.co/0rTvIyaG8t

Wow! Tivemos um aumento de ~4.3x usando Jev como classificador LLM: ~170 docs em ~29s Jev: 231 docs em ~9s Usando geometria espacial, regex e jev como classificador https://t.co/3qrivK5zi0

Introducing a version 50 times faster than Jev, running on your device: laya-mlx! Occupies a maximum of 1G memory only on your device Laya is an open-source classification system similar to Jev, based on text output probabilities I ported it to MLX and made some performance optimizations! In the video, this model is playing Snake on my local M3 Max This model can play Snake at a decision speed of