Jev for data extraction — a daily feed

前にJevの記事で「ほとんどの場所で要らなかった」と書いた。 実は使いどころが2つ目あった。 決算の経費申請の仕分け。 毎年どこかで1日こもってポチポチやってた作業が、今期は1時間半くらいで終わった。 https://t.co/HKOKOzp6TY

レシート自動記帳スキルにJevの判定も追加してみた。 Gemini予測するようにしてるし結局自分で確認するからいらないかもなぁと思う。ちょっと使ってみて不要なら元に戻そう。 https://t.co/h5cOEUnDg4

JEV-STYLE TYPED DECISIONS WITHOUT TRAINING a new approach called mini-jev lets you extract structured decisions from a frozen Qwen3-4B without generating a single output token. how it works: → Turn each schema field into a lettered question → Read the logits for each option → Get the decision directly on closed enums, it matches grammar-constrained JSON accuracy. the speed gains are interesting to

I hate receipts, so I built something that does them for me. Gmail in → @Revolut business ready receipts sent to your revolut mailbox. Every receipt found, matched to the right expense, filed. Even the Uber tip that comes separately. Powered by Jev by @typesafeai . 5,400+ emails in, costs pennies. Also, makes Revolut's "auto-matching" much more accurate by providing additional information in PDF f

補正強化学習モデル「Jev」を用いたプロンプトインジェクション(PI)検出手法の提案。1回のAPI呼び出しで複数の評価確率を一括取得し、JailbreakやPI等を高精度に判定可能。従来のLLM-as-a-Judgeと同等以上の精度を1/63超の低コストで実現できる。ICLR 2027に応募中の論文。 https://t.co/g3LBwv5b8b

npm install jevtok-ts Open-sourced a tokenizer for @typesafeai’s #Jev. Exact token counting, fully local, with zero runtime dependencies. https://t.co/oAOSWJ8p0r

On a big extraction job, a cheap local model with Jev as the judge replaces Claude Opus. Not matches it in a lab. Replaces it where Opus is impossible to pay for: 100,000 company websites cost us about $650 this way, against about $50,000 with Opus. https://t.co/WZ9lwG4sbW

I am now indexing @solana tokens using @typesafeai JEV x @clawddevs x @supermemory IMO we have begun @solana super intelligence. https://t.co/emtJX0KHsq https://t.co/KskS9tOJdq

J'ai intégré JEV pour traiter le rapprochement entre les virements et les loyer pour surveiller qui paye quoi. Mais bizarrement dans mes tests, il est moins efficace que l'heuristique que j'avais déjà mis en place. Donc ce que j'ai fais, c'est que l'heuristique reste le premier passage et si elle n'est pas certaine, JEV passe derrière pour affiner. Sur la lecture des libellés non encore intégré da

@suidevelopers Just finished my first side project on Sui. Using Jev to help better understand transactions. https://t.co/KXcx1EU6wL

Timelapse - 0.5 (3 hrs) Didn’t plan a long session. Just poked at Jev and wrote a tiny BPE. - Read what a System One model is and tried jev in the playground. - Implemented a small Python BPE to see how a basic tokenizer actually works https://t.co/1ne7byQ2fP

codex 帮我整理报销票据，填写报销表格，antigravit 里面的 skills jev 帮我审查填写是否正确，完美闭环了，我需要做的只有一件事邮箱下载发票。当然可以给 ai agents 自己下载邮箱邮件附件，但我还是比较保守派，没肯给它邮箱，我自己动手我放心。 https://t.co/e92IOuqQLC

tab-jev: jev-like model + tabular foundation model = an in-context learner for your text & tabular data. https://t.co/x22clSWJ3L Many real industry datasets are a mix of tabular and text data: - Tabular foundation models like TabPFN learn from a few hundred labeled rows in context, with no training. But they can't read text. - Jev-style models read text and answer typed questions with scores. But

JevでbaserCMSのお問い合わせフォームに「営業お断り」バリデーションを実装してみた。 Jev、本当に早いな。 もしも誤検知での取りこぼしが不安なら、営業用のフォームに誘導するようにすればスパムだいぶ減らせそう #baserCMS https://t.co/mmzkU8Gkc0

First demo I built using Jev in Avalanche, identifies Canadian federal taxes document for ~1 second and less than 1 cent per document. Jev fits right into what we built Avalanche for, and I see myself using it in almost every workflow, often just as a smart if statement. It’s also available to use right now in Avalanche. Try Avalanche here: https://t.co/jamHnhvgKE

jev に文字起こしテキストと show notes 渡して、リンクのトピックが文字起こしの何分何秒にでてくるかを判定。3時間のエピソードで 0.5s, $0.003 でできる https://t.co/KdPA1MYQEX

luccafort さんのプロフィールを 8,000 文字用意して jev にデータを入れました。 AI で作ったので、レビューしてないです。 #byebye_lucca https://t.co/DaK7mtsp3a

but movies can be pre announced right? so its not that accurate you know what cannot be pre announced? death 💀. so we've gathered celebrities that left this world (shout out to wikipedia) and tested Jev on whether they are alive at some point this resulted in a much clearer signal, which is in fact end of January 2025 btw if you heard that jev never hallucinates, its not exactly true. Jev can in f

JEV is INSANE It read 6,030 ChatGPT answers to buyer questions about 1,005 YC companies and checked every company name in every answer. 37,636 verdicts. 9.1M tokens. 40 seconds. $0.38. Grading the same 6,030 answers with Claude Sonnet cost me $36 last month. 71% of these startups were never named once, even when the buyer asked ChatGPT for exactly what they build. Scan your site and see if AI ment

今度はTripoのP2.0からのエクスポートでのリアルキャラ。キャラ自体はぜんかいとおなじだけど、今度はP2.0データから他のことやりつつAstra君とまた詰めて。各種設定・リグ入れ、演出調整。メッシュがいいからリトポは基本無くて１日でここまで☺️😋 １動画目：日本語でリップシンク ２動画目：同じシチュで英語でリップシンク（声が同じ日本人なので少しなまってるｗ） ３動画目：DLSS5 Off（いやーこんなに違うｗ） ４動画目：元のTripo作成時映像 Jevでのリアルタイム感情抽出、Audio2FaceでのLipsync、その他体や目の表情、瞬きなども抽出した感情ベースで演出、マイクロサッケードもちゃんと入れて。仕上げにDLSS5。 まあだいぶ実用に近づいてきましたかね。まだ目の調整少ししたい。ここまでリアルにするのにDLSS5の意味がめっちゃあるし、Jevもいいですね。 かなり整理して、

Pub/Sub AI Enrichment with Gemini and Jev https://t.co/YdSqFmaUsR

Probé Jev 1.13, de TypeSafe, contra Haiku 4.5 en una tarea concreta: sacar datos de filas de texto. Lo curioso de Jev es que no escribe nada. Solo responde preguntas cerradas con una probabilidad. En 19 filas que ninguno había visto, empataron: 16 de 19 filas enteras bien cada uno (84,2 %), aunque fallaron en filas distintas. Campo a campo ganó Haiku, 96,8 % frente a 90,5 %. Jev salió unas 22 vece

When testing Jev for Science, ScienceBuddy parses uploaded GEO bulk RNA-seq matrices without manual coding. Every script run by #ScienceBuddy helps advance #AIforScience methods. https://t.co/O3RSmytMtg

Through Jev for Science, ScienceBuddy accepts uploaded study diagrams and hands back a clean table alongside full code execution logs. #ScienceBuddy https://t.co/Rz3MaelA7A

使用 #Jev 做了一个填写外链信息的演示。 眨眼的功夫就找到入口，并准确的把已知信息填写完了。 准备把之前的Submit Agent改成使用JEV模式。 关注github，在评论区 https://t.co/Ffwex3kxqS

I’m surprised Jev for tabular data isn’t a bigger deal. Every ai spreadsheet company should be racing to rebuild their product ground up with Jev. When we were building people search last year we spent months optimizing the search algorithm. Here is Jev processing 100k rows for $2.50 in under 60s, no optimizations whatsoever. Public demo at https://t.co/6kK0XAURtj try it out with your own csv. A f

When ScienceBuddy pulled protein annotations from UniProt for our target list, Jev for Science meant we could click open the trace and verify the inputs ourselves. https://t.co/K1Sh5xRdUL

受启发也搞了一个，用 Jev 识别剪帖板信息实现表单自动填充 https://t.co/pvYnvmHBGA

有報からtypesafeaiでTOPIX1000銘柄全部で定性スコアを多次元抽出して組んだL/SポートのFF3残差の時系列推移。これ作るのにFableにお願いして１時間+typesafe API代金５ドルだぜ、笑っちゃうね。 https://t.co/ZXD5CUPJcp

話題のJevを使って個人情報を一気に入力する機能を作ってみた。もう個人情報をちまちま入れることから解放された。ガチでAIの進化が凄すぎる。 遊んでる暇はない、AIに適応した人としていない人の差が、必ず顕著に現れる日がくる。世界が変わる。しかし、AIのしすぎで肩と腕がいたい。時間が溶ける。 https://t.co/oU28nsk8Cv

Jev is cool, but this is OWL: 24 free api based web scrapping agent running on a ~25mb embed model inside of indexedDB. EDGE AI in a minimal format: Robust planning/git cloning/price matching using only free API: https://t.co/81Bvg6ICF5

turned the Constitution into a knowledge graph in 4s, the Odyssey in 45s. - per sentence, deterministically generates possible subject, predicates and objects. - used jev to sift through it all. result is very fast and cheap triple creation. https://t.co/vfFawqABOZ https://t.co/pn0tphOAvI

never forget your favorite cup of coffee super proud of my submission for the all gas hackathon we've got: - @convex components galore (static hosting, auth v2, agent, aggregate, workpool, rate-limiter) - @firecrawl to scrape coffee roaster product pages, paired with a little jev action to classify the data - @openai's 5.6-luna in a convex agent to recommend your next bag of coffee by comparing yo

Curating clinical variables from free-text notes is tedious. General LLMs can help, but processing thousands of notes can be slow and costly. Inspired by Jev and the open-source community, we’re releasing MedJev to turn clinical notes into structured fields on consumer GPUs. A 0.8B model + 43 MB LoRA adapter, trained to extract 11 predefined clinical variables. On our benchmark of 2,895 held-out n

can @typesafeai Jev count the r's in strawberry? no. 47% says 3, 47% says 2. a coin flip, same as every LLM. On 70% on 168 test words, it undercounts doubled letters like everyone then I gave it the letters instead of the word: ["s","t","r","a","w","b","e","r","r","y"] 168/168. same model, same question, 260ms @CompleteSkeptic is this expected?

I built a tool powered by Jev that extracts any products features and business logic into a markdown repo Enter a product URL and you can copy features, ICP, messaging, pricing strategy, etc straight to a md file https://t.co/mS8CoBuk1x