X posts

I COMBINED JEV TRADING WITH 166,700 FRUIT FLY NEURONS, THIS IS THE MOST UNHINGED THING I’VE BUILT I took a reconstructed fly connectome and wired its activity into the same live market state Jev was already reading price moves → Jev scores the state liquidity shifts → fly network fires both agree → risk gate opens they disagree → nothing happens the first hour looked completely useless then the fl

We gave @typesafeai's Jev 1,968 live skincare & wellness ads and asked it 33 questions about each one. 64,944 typed answers. 36 seconds, real time, no cuts. 17 cents. Hook type, selling angle, creator-led or not, health claims that would need proof. Every ad, every question. What would you ask it?

i built a text box that turns into whatever(some ui) you type. powered by jev https://t.co/HH0P51033y

I TESTED JEV against Gemini 2.5 and 3.7 on a banking-feed categorization task from our REAL PRODUCT, already serving MANY customers. 612 rows. JEV finished in 6 SECONDS. Cost / time / category accuracy: Gemini 2.5 Flash — ~$0.088* · 95s · 37.3% Gemini 3.7 Flash — ~$0.180* · 78s · 42.5% JEV 1.13 — $0.023 · 6s · 36.1% Gemini 3.7 scored highest. JEV? Roughly 4× cheaper and 16× faster than our Gemini

Codexリセットこないにゃ(-ω-；)ｱﾚ? …まぁいいか🤣 昨日からずーーっと調整してたAIによるナレッジベース自動整理が完成にゃ😸🎉 毎月AIが自動起動→情報整理→Jevで分類→自動処理。僕の判断が必要な時だけ🔔へ 文字だけじゃつまらんので、パックマンみたいにゴミを食べて整理するUIも実装にゃ😹 正常なら100%で静かに終了✨

Integrated Jev in my trading setup. I think I chose the wrong day to test this out. Missed all the fun today. Was able to finally make it all work by 2pm so only one trade. Will try tomorrow, lets see how it goes. Wish me luck https://t.co/nAAVh5Pk6B

last time, I made Jev play Clash Royale. this time, I gave it an opponent: Laya, an open-source alternative. Jev runs through @typesafeai's API, while Laya runs locally on my Mac. Each controls a separate emulator and sees only its own game screen. Both use Qwen (via cerebras) for battlefield vision and OpenCV to read cards and elixir. From there, Jev and Laya decide what card to play and where to

A really cool use case of how we use Jev to understand how people talk about Upstash This is 100% open-source too. Quick demo 👇 https://t.co/KlwpWJYsAH

Built a deal-review demo around Jev. It recommends what comes next: approval, more documents, or sales, finance and legal review. One deal can need several teams. Jev maps that out; people still make the call. #Jev #EnterPro #EnterArtifact https://t.co/thtHBETtYv

Made a real-time presentation coach with Jev. It listens as you present, follows your slides and notes, and keeps track of what’s still left to say. https://t.co/trPsxOmEXk

Introducing Jelight v0.1.0 A code highlighter lib backed by Jev. Fully Open Source with no external dependencies. Works with both @vercel & @typesafeai api gateways. https://t.co/MliJSuL8GX

I jumped on the Jev hype and had it play my own game, Chess Defence. The brief still isn't very mature, but after a bit of work it started playing decently. You can see the results in the videos. #indiedev https://t.co/ecNxwmBit6

Jev × インティメート・マージャーの Audience API で、ユーザー属性に合わせてコンテンツを動的に出してみた。 ブラウザに登録された IM-UID から属性を取得 → Jev に渡す。 色味、デザイン、文言、パーツ配置はあらかじめ用意しておき、属性に応じて組み替える。 全部ゼロから生成するのではなく、判定を高速・低コストで回して組み立てる感じ。 このスピードと費用感なら、ハイパーパーソナライズな体験は十分現実的だと思う。

jev helped me build my own ad blocker - apple safari-inspired thanos snap … INSANE! w/o any hardcoded filter lists. looks at each element and decides: ad or not in one article: 16 ads gone, ~20k tokens, $0.0008 cached for next visit open source, byok, repo below @typesafeai https://t.co/qkuzE6Wx1N

#aimeetup #Jev Jevに自作ゲームを遊ばせてみた 〜チャージ・ラン（1キーで遊ぶ避けゲー）〜 手のひらサイズの端末（Cardputer）向けに作った避けゲーを、TypeSafeのJevに遊ばせました。 ルールは「障害物をジャンプでよける」「空中の帯に長くいるとゲージが溜まり、満タンで無敵になる」の2つだけ。 比べる相手として、ローカルで動くOpenJev、状態を見ずに動くだけの相手（毎回大ジャンプ、ランダム）、自分で判断手順を書いたプログラムにも同じコースを走らせています。 【Jevを使っている部分】 ・着地するたびに、いまの画面の状況（自機と障害物の位置など）を渡す ・「待つ / 小ジャンプ / 大ジャンプ」の3択から次の動きを選ばせる ・判断を待つ間はゲームを止める（1回の判断に0.6〜0.8秒） 【結果】（点数は20コースの中央値） ・位置と速度だけ渡す：大ジャンプを一

基于FastJev 跑通了 Jev Ultrafast 的官方浏览器任务。 单张 RTX 5090，本地 Qwen3.8-27B EXL3 连续完成 Lisbon 搜索、Design 筛选、免费取消和打开 Casa Flora：5 个浏览器动作，第 6 次决策返回 DONE。 单次录制端到端 10.375 秒，决策请求延迟中位数 1.388 秒；FastJev 0 输出 token，录制中 0 次 TypeSafe API 调用。 源码 revision、决策 trace 和录像都在这里： https://t.co/S7SpS13066

Jev-powered Agent #Jev Built a local resume screening workbench for HR teams and early-stage founders. You define the bar → upload resumes in bulk → it checks every condition → every verdict has a quote from the resume to back it up. Not "AI thinks it's a fit." More like: which conditions passed, which failed, and exactly where in the resume. 👇 demo

Jev Engineering can make a https://t.co/kHgTxlkUry monitoring system faster and cheaper by keeping expensive models out of routine decision loops up to 201x faster and 456x cheaper in tests these are reported benchmark results and measured TRENCHNET performance figures here's how I applied this idea to tracking traders: observed trade → structured wallet history → graph context → Jev classifies th

I turned Jev into a Flappy Bird player and I won against it (sorraay @typesafeai). Can you beat it? Let's see your score: https://t.co/93Bl3QQsFq Tech: - Results stored in @neondatabase Postgres - @nextjs App Router 16 with @DrizzleORM, @shadcn and @tailwindcss - pdx1 region on @vercel gets the lowest latency to Jev's API

JEV BOT ON GOD-MODE there is no catch and that scares me Elon Musk posted an idea that hit 19 million views: "the win is priced in before it happens" that's basically what Jev Bot does now - I gave it $65 and told it to earn its keep or get shut off $50 → $8,730 in 24 hours still running, still compounding - nobody's touched it since it started, it just sits on its own machine in the cloud while the markets move somewhere else every 8 minutes it: > scans every open market before the crowd catches on > checks how far the real signal has pulled ahead of what the posted odds still assume > only s

Found an open source repo that plays Clash Royale on its own judgement. I rebuilt it for iOS and handed the controls to JEV. JEV is doing a brilliant job so far. It needs more training before it plays like a pro, but it is getting there, one match at a time. https://t.co/KwInPxXVsG @typesafeai #OpenSource #AIAgents #iOSDev

Local AI just beat the API. I built a movie search that runs Laya on my machine and Jev over the API at the same time. Same query. Same posters. Same quality. Local: 392 ms API: 676 ms Hardware: AMD Radeon Pro 5500 XT Then I flipped the toggle: Laya vs Jev deciding the results. Same answer. One of them never left my GPU. Faster. Same quality. 100% local. #LocalAI #Laya #Jev #AMD #OpenSource

@sbusso "JEV" PR branch merged to main, here it is playing 2 player PacMan by itself, powered by laya and MLX-Serve 💪 Thanks ! https://t.co/sD0DBZB9S7

#aimeetup #Jev #flywire Jevにハエの脳を運転させる 〜飛んでくる球をよける〜 【Jevを用いている部分】 ・進む方向を選ぶところだけです。少し先まで直進するか、左右に曲がるか、止まるか跳ぶかの7つを並べて、そこから1つ選んでもらっている感じです ・候補ごとに先の球との近づき方はコードで計算して表にして渡すので、Jevは表を見て選ぶだけになっています 【Jevを用いていない部分】 ・眼はflyvisの視葉、脳はコネクトーム由来のニューロンで、視葉から脳へは実際の結線で流しています ・跳躍の反射は脳の値を直接見ていて、Jevの選択より先に体へ届きます。閾値を超えた窓では、そもそもJevの出番がないです ・3D空間とハエの体はMuJoCo ・歩く動きはflygym同梱のもので、脳とはつないでいません。Jevの旋回も左右の駆動の差になるだけです 【所感】 ・当たるコースと

Most people are burning API credits on reasoning models doing dumb grunt work If your smartest model is spending tokens deciding what NOT to read, your architecture is broken. Simple fix: Routing layer (Jev) + Reasoning layer (Grok). Recent test on 3,412 leads: - 20,472 filter decisions via fast binary checks - Grok only opens the high-signal leads - Runtime: 15.7 seconds - Cost: $0.41 (down from

I disappeared from X for a bit. I was building. Jev? Déjà vu. Browser Use + Jev: 7.1s Google Flights. dejevu + plain Llama 3.3 70B: 5.6s. 10 model calls vs 17. 5.6× fewer tokens. 1 API key. No daemon. Every run code-verified. Traces public. https://t.co/oQ6IhcRhjE https://t.co/WwYpu1U8nU

Hype aside, I switched to Jev as the classifier for RedReplier. Ran it against the 2M dataset, and the quality is BETTER than my previous setup with DeepSeek. Saves me ~$700/year, and it's ridiculously fast - roughly 5x faster. https://t.co/khGk7inFKF

I built Jev-feed-analyser that reads my X timeline while I scroll Every post gets 12 typed questions, answered in ~300ms each > is it shill, does it read like AI, is it worth a reply > each answer comes with a confidence score > one verdict per post, keep, reply, skip or mute the author No summaries, no prose, just decisions that Jev makes for me, then I check it myself. I ran the same 12 question

Another good use I found for @typesafeai’s Jev: automatically switching Claude Code models based on what I’m working on. I built JevShift for this. It looks at the task and recent context, then picks a model. Here’s a quick demo 👇 https://t.co/0ADRsyeTFd

#aimeetup のJevハッカソンでプロトタイピングしてた音声ペアプロ環境！ マジで結構すごいから見てほしい、動画ではなんにもツールが入ってない環境から通話の指示だけでブログ環境構築して新しい記事作ってる 個人的には「チャットに送ったやつ開いて」とかいうとそのまま開いてくれるのが良さげポイント

Turns out JEV is great at reading content super fast and scoring what’s worth highlighting. So I built a smarter Cmd+F, one that understands context instead of just matching words. You enter a topic, and JEV scores the content and highlights the relevant passages with high confidence, without needing a larger model. I can see this being useful for students skimming material, getting a TL;DR, or bo

JEV BOT MADE MY BALANCE OVER THOUSANDS OVERNIGHT I put Jev between a trading agent and its wallet, funded it with $50, then left the loop running 22 hours later: $50 → $39,319,82 every five minutes it: > reads the latest market state > measures the signal against the posted odds > calculates whether the gap still clears fees > routes each check to the cheapest capable model > escalates only when t

Diogo Almeida, founder of TypeSafe and former OpenAI researcher: Jev Engineering. Imagine your entire agent team costs $500 a year. You used to budget $19,000 for the same workload. The savings are simple: Jev from TypeSafe handles routine decisions. The expensive model only gets called when a deeper analysis is needed. Here’s how I divide the roles in TRENCHNET: > Jev decides: skip, flag or investigate. > The graph checks which wallets regularly move together. > Astra handles the deeper analysis. > Telegram delivers the alert with transaction links. A wallet buys. Another follows. Both buy ag

Built @matchcndev because I think coding agents should search before they generate. You describe the UI in plain language, and matchcn finds the closest component, explains why it matched, and gives you the install command. No guessing component names. No jumping between libraries. No rebuilding something that already exists. Right now I’m indexing 1,783 components across 9 shadcn registries, classified across purpose, motion, density, interaction, data, and decoration. The classification and matching layer is powered by Jev through classifier, so every decision comes with confidence instead o

I just built a Jev X Self-Improving Agent. 12% → 88% success. 29 seconds. $0.02. Zero fine-tuning. Same agent. Same model weights. 50 attempts at a hostile checkout. after every failure Jev asks 6 typed questions: which step broke, what was the root cause, will it happen again or was it noise. 8 rules kept. 16 failures thrown away as noise. ignore the green "BUY NOW" banner - +16 pts captcha → audio challenge - +12 pts the agent was never bad at checkout. it was bad at not clicking ads. Read this article on Jev Engineering below and turn your ideas into reality.

it's absolutely insane JEV + NEW GROK BOT + RH API = ANOTHER $760 IN 13 HOURS the gap between a signal and an order is where copy bots die, but here is what sits in mine. > Jev turns a wallet move into one answer. > the harness acts on it. > every matched buy gets simulated twice. > the second run is against the limits you set yourself. > whatever survives both reaches the market exactly once. once, stopped, watch. three endings, nothing else on the floor. nobody has been at the keyboard since yesterday. am i should public the repo of this crazy tool?

I built a voice game of 20 questions without LLM. You ask out loud, speech to text transcribes the question, then @typesafeai Jev answers it in a few hundred ms. Really fun to play! (the video plays at 1x speed, turn the sound on) The code is below ↓

i gave Jev + Astra 6 a dead tiktok account and told it "make money or i'll delete you" 11 days later it's booked $4,200 in brand clips and i still haven't opened the app this isn't a wrapper. i trained the model in the video myself. 22,045 weights, 6 blocks, it learns which 4 opening frames hold a thumb. jev labels its food, astra 6 wrote its training loop every 30 minutes it: → pulls the 800 to 1,200 clips posted in my niche in the last 6 hours → labels every one of them with jev: hook type, pacing, why it held or lost. 1,200 clips costs 8 cents → keeps only the ones that beat their own chann

I put TypeSafe’s JEV in control of G1 humanoid in my virtual kitchen. The mission: reach the stove with obstacles on the way. A VLM and depth information helped map the kitchen. A path planner produced possible ways around the obstacles, and JEV chose which move to make next. A pretrained walking policy handled locomotion, with calibrated controls for turns, forward movement, and curves. The robot reached the stove in about 64 seconds of simulation time. This video shows the highlights. Still a prototype, but pretty cool watching perception, decisions, and movement come together.

Jev is INSANE when you give it an inbox. 20,000 messages. ~$0.0025 estimated model cost. I built an open-source Instagram + WhatsApp inbox map with Chatfuel SDK. Questions. Objections. Buying intent. Click a pattern. Read the messages. Code + installer below ↓