X posts

i killed my ai knowledge graph and rebuilt it as a news feed. the whole classification pipeline now runs on a decision model that can't generate text and it's ~10x cheaper. the old stack: neo4j + redis + a claude pipeline doing 8k-token extractions on every story. ~$40/week, and i shut it down because it was too expensive. the new stack: postgres, two tiny services, and jev (@typesafeai's system o

Made Jev from TypeSafe AI play 2048 At every turn, I give Jev the current board state and let it decide the next move: up, down, left, or right. No text generation. Just state → decision → action. A fun little test of System One models https://t.co/5Ps61plMid

Jev ile Jennefer ı birbirine bağladım. 8 hazır task seti üzerinden agent seçimi motoru olarak kullandım. Sonuçlar inanılmaz derecede doğru: Öncelikle LLM lerin uzun süren ve reasoning gerektiren karar alma süreçlerini unutun. Jev bu işlemi saliseler içinde yapıyor. Ekstra yazı, kod vs. üretmediği için harcadığı token yok gibi bir şey. 8 taskın kararı için harcadığı token: 2010 ve daha gülüncü bunu

同じテトリス盤面と候補手を、特化型判断モデル「Jev」と汎用言語モデル「Claude Haiku 4.5」に渡し、プレイ性能を比較する実験 私も試してみました 確かにJevの速度・コストは魅力ですが、渡すデータ工夫すれば元ツイートよりもうちょいHaikuは賢いっすね ※ Vercel AI Gateway 経由のJevです https://t.co/wo0Hh7hrN9

Astraとおしゃべりしてjevニューロン作ってもらった 思ったよりそれっぽくなってる シナリオが進んでいき、リアルタイムで感情が変化していく感が出せている https://t.co/jWsVf3bPqt

Jev vs Deepseek vs Luna on Connect 4 Jev beats Deepseek and ties with Luna But Deepseek faster than Jev 👀 https://t.co/vkvpv0rlbq

built a highlighter on jev paste any language → my code tokenizes → jev names the lang, colours every word, then says which of 9 lint rules fire and where nine rules in code. jev just answers. near instant https://t.co/RvPmRciFOd https://t.co/YFeYD9fyv8

Jevエモーン！ メールが26万通もたまっちゃったよー！ 重要なものだけ仕分けてよー できたよ〜メール整理ワークスペースぅー！ @typesafeai https://t.co/Us4YvnSef5

jevを活用してカメラを使った運転時の危険予知アプリを作ってみた。 とりあえずyoutube動画で試験してみたが思ったより精度が出ないなぁ… そもそもこの分野への理解がないからもっと学ぶところから始めなきゃかなぁ https://t.co/PluQ2qjboN

I jused Jev to create a personalized search engine. Give it some components, can be generic too. Jev analyzes the query and picks the right component A web model goes out to fetch the data Data rendered in component picked by Jev. The bottleneck is the web search, so you need a very fast model for this, but you could do this without web searches too. You could for example just display information.

This is a Jeff classifier made with Jev. It processed 20 images in 9808 ms, showing the live probability and raw result for every round. Built out of pure love for the “My name is Jeff” meme. Fast, goofy, and extremely Jeff. https://t.co/BbXC2pvcQ6

Another night of jevsomnia Inference this fast opens up new ways to consume content. Here Jev augments the reading experience rather than summarizing it. Jev reads the novel with you, marking what it finds emotional, live as you scroll, and you set how emotional you want it to be. Paired with a character map, so you can jump straight to any character's mentions. A working demo of interesting thing

ちょっとJevのテストするトロッコ問題サイトを作って、動いたのでもうちょい実用ベースのものを次はやるか https://t.co/u2Q68Qw98W

Built a construction plan-set classifier with Jev. Proq turns civil and building plan sets into bills of materials using an LLM pipeline we built on GPT-4.1. Jev classified an entire 26-sheet plan set in 2.9 seconds for $0.0052. It matched GPT-4.1 and GPT-6 Astra on 100% of sheet-level classifications while running 17–21x cheaper and 5x faster than our production pipeline.

Jev is WILD I gave it two launch tweets and fed it over 4000 demographic profiles of real survey participants Twelve seconds later, a simulated A/B test tied to actual personas voting on the best tweet you can just do things i made it 100% free (link below) https://t.co/UcN4Eg3h3X

Jevを使ったバイタル急変シュミレーター 重症患者ではモニターをつけてバイタル急変を予測しますが、アラームが機械的すぎてあまり役に立ちません Jevは確率がキャリブレーションされているという大きな特徴があり、急変確率判定に使えるのではないかと思いました。 結果として、想像以上に好成績をあげました。 ①左側のビデオが正常状態です。モニターアラーム・Jevともに正確に動作していますが、モニターアラームは体動も異常と判断してしまいました。 ②右側のビデオが徐脈患者です（高齢者でしばしば見ます）。モニターアラームは常になり続けて正常に動作しませんが、Jevは正確に患者状態を予測し続けています。

Battle TankのボスのAIをJevにしました。 かなりつよい。 ガキの頃友人に対戦でまったく勝てなかったのを思い出した。 ここで遊べます。感想ください https://t.co/0QeFMvhtz5 https://t.co/XUbsMYWjsi

話題のJevってどういうAIなの？の雰囲気がわかるサイト作りました！入力した言葉のイメージを判定して色で回答します。 高速で固定の型で出力するJevの特性。こんな簡易サイトでも今までのAIだとすぐ破産できたけどJevなら無料で公開したって大丈夫！という格安感が表現できていると思います。 この感じで使えるJevはゲーム開発とめっちゃ相性良いと思うんですがどうでしょう？ CloudflareでもJev使えるようになったのでJevもサイトもCloudflare一括管理で作ってみました。このURLから誰でも試せます！ https://t.co/LTFiDESEl5

给 Claude Code、Codex、Cursor、OpenCode 等编码智能体提供本地 MCP 代码质量评分循环。 https://t.co/UEKPaFRYOJ 一个本地运行的 MCP 服务器，基于 TypeSafe 的 Jev 评估服务为 AI 编码智能体提供结构化质量分数。通过 npx plugins add 直接从 GitHub 安装，支持 Claude Code、Codex、Cursor、OpenCode，运行时是本地 Node.js 20+ 进程，经 stdio 只暴露一个 jev_review 工具。

I made a Magic Jev Ball for code reviews 🎱 Click it on any GitHub PR and ask: "should I approve this?" It checks CI, diff size, and reviews, then lets @typesafeai Jev decide your fate in ~200 ms No more thinking. Just shaking. https://t.co/8wqdab2Axf

We let Jev, which can only choose and can't see or write, play today's Wordle. Its first guess never landed. Then it decided to search the web, and its text-writing sidekick typed "today's wordle answer". It never opened a result. GPT-6 Astra solved it in 3. https://t.co/qNAi61vrzr

Jev is incredible. I used it to vibe code a chrome extension that automatically detects AI slop on LinkedIn. I'm doing X next. If you want to try it let me know 🫡 https://t.co/Uybj2fsNCD

Jev にぷよぷよをさせてみた。 - 実行エンジンは https://t.co/NiVAe3js64 。ゲームはリアルタイムで進行 - 盤面とそれぞれのポジションにおいたあとの結果を与え、①今どのフェーズか ②フェーズごとに、戦略からどの結果がベストか の1+3つの質問をして、選択肢を選ぶ 画像は Random に勝つ様子。決断が早いぶん Random も強い。 最初は盤面だけ渡して、単にどこにどの向きで落とすか判断させていたが弱すぎたので、配置後の結果を選ばせるようにしたり、戦略を渡したり、判断を多段化した結果なんとか3連鎖もでき Random に勝てるようになった。

家庭内AI秘書に話しかけるときに元々 haiku で意図分類させてから後段の sonnet/opus/fable に流してたのを Jev に置き換えて playground を作ってみた。タイピング進む度に意図分類が変わっておもろい https://t.co/WORQ03oZvA

Built a quick thing with JEV that checks whether content was written by AI or a human. So now if you write with AI, you get checked by AI. Funny thing: Ran it on Typesafe's docs - yep, AI-written

現実を言葉で検索できる「現実ブラウザ」を作りました michiyomiのapi情報をjevを使って分類検索をしています https://t.co/jn9IMDdckJ

I’m in 1038 telegram groups, channels and bots for meme coin trading It’s impossible to look at them all So I had Jev and Codex build me a dashboard I can access on my computer and my phone to see calls from my groups in a nice feed I’ll open source this later if there’s interest! Quick little vibe coded project

Embassy firmware running a simulated esp32c3 it uses openrouter api for jev jev here plays snake! https://t.co/F3HQlvoT1K

Gave TypeSafe’s new Jev model 637 ancient religious texts and has it classify them by tradition and theological mechanic. 13 minutes, 5,096 classifications, ZERO errors. Every tradition was correctly identified. It found cross tradition patterns across Egypt, Persia, Sumer, Kabbalah, and Gnostic texts. Same mechanics. No contact between traditions. And it just indexed a corpus that would have take

32 AI agent decisions. 6 seconds. $0.0006. Watch all 3 👇 I asked a model 32 questions about my inbox, my leads and my X feed. Jev doesn't write a single word, and every agent in Orbit can call it now. https://t.co/55PTQXM6Qm

We tested Jev by @typesafeai on a real Ruka workload: 500 products × 801 possible master supplies. Done in 8 seconds. ~62 products/sec. Our current LLM pipeline: ~2 products/sec. 🤯 And we still have room to make this faster and more efficient. https://t.co/NKm8tZmvPQ

introducing jev for linkedin posts! paste a linkedIn post and see how much of it is a humble brag, ai slop, or a fake sob story. built with @OpenRouter @jjacky https://t.co/3wCAs3xXnZ

Built another small experiment with JEV from @typesafeai integrated into Synkora AI. 🚔 Talk Your Way Out You’ve been pulled over for speeding. The officer approaches your window. What do you say? I built the game around an LLM-driven conversation where your response affects the tension meter—every turn can make things better or worse. The interesting part for me was using JEV as a tool for the LLM

先日、配信されたKEITOさんのJevの動画見て、自分はVercelのAI GatewayからAPIキー作って、簡単な文章をJevで評価・判定するデモアプリを手探りながら作ってみました。わかりやすい動画いつもありがとうございます。 ▼KEITOさん動画 https://t.co/ZjXVh3Aip9 https://t.co/tSfTh45Wzd

Jevにブロック崩しプレイさせてみたけど下手な人のゲームプレイとしては逆に良いかも ちなみに「通信中も動かす」をオフにしてAPIレスポンス来るまでゲームを止めると普通にクリアできるようになる https://t.co/14olk1SQRl

I gave @TypeSafeAI's Jev a robot and one rule: keep the flower alive. 🔥 Fire? Put it out. ❄️ Frost? Warm it up. 🐝 Bee? Let it through. https://t.co/hX0N7SvXkU

给大家汇报下成绩，我用JEV做自动化智能交易，现在来看比其他Agent强太多了。 这家伙累计交易114笔，胜率在30%左右，还没算持仓盈利的数据。 我大概看了下，这家伙这轮行情一直做多，没有一个做空，所以后面还要观察会不会做空。 最让我惊讶的是这家伙NEAR的2.7多单一直拿着，从来没有卖出过。 还在继续测试，有啥新情况及时给打击说哈，越来越有意思了。

Been spending a few hours tinkering with Jev for that 'generative UI' example (in Steve's vid below) Only to realize it only works if you have - a pre-written set of UI - and the data for it (like names, etc) Again, Jev can't write. can't write names, can't write styling. you need those as pre-written set of options. This is a small experiment in @lunagraphHQ combining: - Jev - Pre-written styling

jev is live on halofy! build reusable jev plugins for your entire team using any agent with a simple prompt here a quick flight search plugin that i built using claude + jev try now at https://t.co/EUQ2Ny9Lnc https://t.co/rfhyamAxpl

Sprint Planning Sim with 4 Jev Agents, which actually works to connect to your Sprint ticketing system. https://t.co/mid4lnfKrg