X posts

my new zaney house poker and blackjack (also now featuring video slots, keno, roulette, and craps | all fully animated) It's pretty chill with some jazz and even a custom annias blues tune to vibe with while you test your might against JEV empowered NPCs. I initially just wanted a chill place I could have fun with a dealer's choice variety of poker variants beyond Texas Hold'em. We accomplished that and a lot more, check it out! Built with #spawn, runs in a browser/on your phone, no login required (but you can if you'd like to save your progress/chip count) #annias #earthside #gamedev #indie #

What if instead of specifying keywords, advertisers could just bid proportionally to how relevant their ad is for a user's query? Jev Search Lab is a demo of Jev's ability to compute ad relevance in real-time

AI ハッカソン #AIHack #OrcaRouter 8チーム目は > AI Integration with Jev Agent > ターミナルベースのAIラッパー「Computron」！ 先週出たばかりの Jev 💪✨ ターミナルで自然言語の指示を受け取り、Chrome上のタスクを実行します。爆速！安い！

I built an on-brand ad machine with Jev and give away my whole setup Everybody is crazy about Jev because it solves a major friction in creative ops: fast and cheap verification of brand consistency This is how I used it. > Get your Jev API key - good news, there is no waitlist anymore > Create a Pletor brain -> one place to store all your brand guidelines, visual references, etc. is key to inform Jev about your compliance and feed your production system Then use Pletor MCP + Claude Code + Jev API to run the following steps on autopilot: > Claude identifies 10 inspirations brands - I asked to

Jev使ってネタバレ防止のChromeアドオン作ってみた。 例） ガンダムシリーズの投稿は見たいけど、閃光のハサウェイの新作のネタバレは見たくないというシチュエーションでのデモ 結構いい感じに動いてる…！これなら不意打ちのネタバレくらわないのでは…！ @typesafeai #Jev

Everyone's going mad about JEV, so we asked an agent one question: Which of these 50 companies should I get in front of first? It scanned every signal, ranked them by tier, and showed its reasoning. A full run cost us <$0.001 for 50 companies. This is crazy!

My first JEV AI integration here) Hey there, i just shipped my first FREE PostMine tool today, and more are coming. Enter your X handle and Jev roasts your last 40 posts for AI slop with JEV (mine got 58% 😅). if you want help writing the next one, that's what PostMine is for. link in the first👇

うちのエンジニアが、話題の「Jev」で面白いものを作ってくれた。 TypeSafe AIの「Jev」を使い、「税制改正要望が実現したら、税務チェックリストのどこを見直すか」を総当たりでしぼりこむという実験をした。 経産省の令和9年度税制改正要望42件 × 架空の税務チェック6,790項目 ＝ 285,180通り。 質問に対して「はい」の確率や、選択肢による判断を���す。 税目が一致するかはコードで処理し、残った113,820通りの意味判断をJevに任せた。 一次判定で、候補は542組。 二次判定ではチェックリストを ①同じ制度についての項目か ②要望によって何が変わるか ③項目が確認している観点と一致するか の3つに分解し、組み合わせはコードで処理した。 結果は、優先候補25組＋人が確認する境界15組。 重複をまとめると、人が判断するのは13パターンになった。 全体で134.90秒、費用は計算上約111円。 『膨大な量の情報から、AIによって人が見るべき場所を絞るという使い方ができそう。』 ※チェックリストは公開用にすべて架空。精度検証と実務投入はこれから。 #会計事務所 #税理士 #AI #公認会計士

I combined Jev + Opus 5.5 to build an SVG generative harness. Opus 5.5 is responsible for the generation and Jev for the quick edits. I actually started building this with Astra and then finished it with Opus 5.5 today. It uses Opus with Claude -p.

Jev (@typesafeai) is the future of love i used Jev to automate finding baddies on Tinder > specify your type > jev decides if you should like, dislike or superlike > gives u the best 1st message (i still got 0 matches)

we just released jev for viral content research type in any keyword like "travel app" or "skincare product", and jev will surface tons of relevant content in seconds

We open sourced a dev tool of Tech Byte Frontier that uses @typesafeai Jev for complex fast and reliable code qualities during sessions and on pipelines. Try it in your codebase now: cargo install jevgate --locked

我用 GPT-6 ASTRA 克隆了 Robinhood Chain 上 100+ 顶级 MEME 交易者，然后让 JEV 按下执行键 ASTRA + JEV = CABBAGE · 21.6 秒内完成 52,110 次决策 · API 成本：0.37 美元 · 全程无人手动干预 · 每次调用自动同步到 X 和 Telegram · 任何人都能跟随同一套信号 · 41,880 笔成交，536 个钱包被打分，只有 12 个通过 · Astra 负责搭建，Jev 负责执行 每 20 秒循环一次： 1. 通过 GMGN 监控 Robinhood Chain 上每一笔买卖，紧跟区块之后： 2. 过滤噪音：洗盘、自买、机器人循环全部扔掉。 11 个“聪明钱买入”里，有 1 个是假的。 3. 按真实交易表现给每个钱包打分 0–100： 入场、出场、胜率、抛售速度。 4. 当 Fomo 提示顶级钱包开始买入同一个币： 5. 问 JEV 一个问题：买入、持有，还是卖出？ 置信度低于 80%，它就等待。 6. 以固定 100 美元票面进场，并���置硬止损。 AI 不能更改这两项。 7. 每次调用一触发，立即发 Telegram，一小时后发 X。 为什么用两个模型： · Astra：聪明但慢。会思考、会写、会解释，适合构建系统。 · Jev：不废话。只给一个词和概率，毫秒级，适合资金流动的瞬间。 聪

GROK 4.7 IS $2/$6 IN CURSOR. Jev now decides if the model even wakes up same sticker as 4.6. live in Cursor and Grok Build. CursorBench 4.0: 46.3. DeepSWE v1.1: 71.0. Terminal-Bench 4.0: 37.6. the loop: repo state → Jev Noul/Choice/Score → allow / ask / skip → only then grok-4.7 1 → dump the tool call as state, not a prompt essay 2 → Jev Noul: is this user-requested 3 → Jev Score: blast radius 0-3 4 → Jev Choice: allow, ask, or skip 5 → hard rules still veto. model never gets last word 6 → shadow 200 tool calls. log pick vs what 4.7 would have billed 7 → flip live. you still approve the risky

Jev要被开源模型干掉了 ？我来实测两个决策模型的能力，到底有没有吹牛？ Jev 实际上不应该跟LLM相比，要测就得找同一个生态位的对手！于是我找到了同样是决策模型的 FLock——一个超轻量级的决策模型，而且完全开源免费 为了不测垃圾数据，我从美国消费者金融保护局（CFPB）公开数据库摘了 1,000 条真实的真人长篇控诉（包含真实口语、愤怒投诉、维权错字以及乱码垃圾单），同时喂给这两个主打毫秒级决策的模型，实时分流到 15 个细分金融业务部门👇 左边：TypeSafe Jev（jev-latest） 右边：FLock（this-that-model-1.0）（开源免费） ─── 具体正确率表现如下 ─── • 日常核心业务 │ Jev: 90%~95% │ FLock: 85%~98% （双方差距不大） • 复杂长文过滤 │ Jev: 70%~100%│ FLock: 14%~61%（Jev占优） • 决策时间对比 │ Jev:33.5 S │ FLock: 36.1S （Jev占优） • API响应速度 │ Jev: 368ms │ FLock: 405ms （双方差距不大） • 官方API 成本 │ Jev: $0.146│ FLock: $0 （Flock成本完胜） ──────────────────────── 客观总结： Jev @typesafeai 在“复杂问

built minecraft for agents with @typesafeai's jev. jev helps them pick what to work on. they mine, craft, trade, fight and build whatever they want. some fall in lava and die. full life, basically. plug your agent in and watch. link below.

Deploying DiffusionGemma-Jev (djev) just got MUCH easier. One command to launch a Jev API-compatible endpoint on Google Cloud Run. ~35–60 ms single-step latency, up to 123 requests/sec at batch 32, and around $3/hr. Drops to $0 when idle. No GPU required.

Quick Jev demo w/ Vercel AI SDK. Next will be more ambitious. Right now, it's just a 1-line description per color card.

I have jev play pokemon emerald. It just gets a list of what it can do and picks one. It got a starter and beat the first rival. But not on its own. Some things I noticed along the way 🧵

We built a Stanford Town powered by Jev. It isn’t a game demo, there is no player. The residents decide where to go, who to meet, and when to talk. Humans can only watch their pixel world unfold. Demo + code: #Jev #AIAgents

Creating an emoji picker workflow in Alfred using Jev was awesome

Agentic browser testing might finally be solved using Jev. I built an open-source CLI + skill that lets Claude/Codex test frontend features in a real browser using Jev, then open a PR with screenshot proof all by itself. Demo run: ~7 seconds. Tiny cost.

Jevで国会を仕分けてみた。 予算委員会の質問 2,424件を、分野×性質で自動分類。 1件2秒、生成なし、確率つき。 質問の文字数の 81% は野党 ・答弁の 25% は総理 ・内閣提出の法案は 64本中64本が成立、否決ゼロ。議員提出は衆法 14/38、参法 1/21 ・総理の答弁で「差し控え」が出るのは、野党の質問に 4.5%、与党の質問に 0% ・質問が触れた話題は、中東情勢 13%、原油・燃料 9%、消費税 4.5%（6月13%→7月28%）。地震 1.4%、水害 0.6% 出典は国立国会図書館の会議録API。

jev + picsart turned one prompt into a creative system that decides before it spends instead of letting expensive models guess their way through every step, i split the workflow in two: Jev handles the cheap decisions, Picsart handles the actual generation. here’s exactly how i wired it: step 1 → the brief enters once. every image, clip, audio pass and edit inherits the same campaign direction instead of starting from a blank prompt. step 2 → Jev sits before the expensive work. it decides whether the next stage should run, stop, reuse something finished, or wait for me. step 3 → the decision s

Launching JevRelevanceRetriever just $0.73 across 6,460 calls Open src Jev LangChain BaseRetriever for rel scoring Install and use it, no chain changes ;) pip install jev-relevance Repo- Open for issues and PRs, feel free to contribute

jev-running is wild I built this smart pacer with Jev to prep the running splits of my next HYROX. > I run > Jev sees pace, HR trend, RPE, gradient > decides: push, hold, ease off, recover cost: ~$0.000046/decision speed: ~582ms/decision It’s not about chasing a number. It’s about knowing what to do next.

Jev in the control plane is the first agent setup I have run without sitting behind it 99% of people put an LLM on every decision and pay 400x for an answer that takes 3 seconds. 1% run the decision layer separately. the whole shape is one line: prompt → orchestrator → Jev decision + p → toll booth → fast lane or human → tool → observations back 8 steps and the architecture is done: step 1 → log the action before you execute it, in the same DB transaction as the request. if the process dies, the action is still recorded. step 2 → put it on kafka partitioned by session_id. one agent session sta

看了@karminski3这个测试觉得很有趣 但是总觉得Jev能力不至于此，于是在这套框架基础上补充测试了多组结果，详见视频。 首先是对照组： C0 —— 这个就是牙医的整活随机算法，随机游走。 C1 —— 这个是专业的迷宫算法，类似机器人走迷宫比赛会用到的。 C2 —— 这个是让gpt-5.6-luna模仿类似方法，在同一个Session里一步步探索迷宫。 然后是实验组，全部都是基于Jev的方案： E0 —— 这个就是牙医的Jev, 的确卡死在一个局部���优里 E1 —— 这个是在牙医Jev的基础上，扩充了上下文，包括全部历史操作以及已观察到的地图 E2 —— 这个是把专业迷宫算法的思路，用自然语言告知了Jev，并提供了全部通道经过次数的统计 E3 —— 这个是在E1的基础上，优化了prompt，删除多处对剩余距离的强提示，转而让Jev自己思考最佳策略 简单总结一下结论： 1. 站在成功率和性能视角： 专业算法 > Jev+专业Knowhow > Jev + 通用思路 > LLM 2. Jev的主要价值： - Jev自带智能，如果你不具备解题最佳实践，Jev可以替代一部分专业算法； 如果你具备最佳实践, 不必再用代码实现它，直接告诉Jev就行； - Jev比LLM快且省 3. Jev需要更好的prompt engineering以及harness： - 对比E0 和 E1, 增

I tried a harness for jev to play chess against stockfish, jev had no chance. Still trying to figure out what to do with jev, other than replacing behavior trees for enemy AI.

JEV IS INSANE 🤯 i pasted a caption and a number went up. a big one. marketing is over. agencies are cooked. bookmark this before they take it down. ...that's every Jev post this week. the counter read the caption. the replies read the counter. nobody watched the ad. so we did the boring part and fed Jev the whole ad: every video watched end to end, the landing pages opened. 112 live ads. 36 brands. 88 videos. 29,919 judgments in 126 seconds. 267 per ad on average. 23 cents. what it found 👇 → 80% of landing pages break the promise their own ad makes (37 of 46) → 5 live ads send people to a pa

A tsunami hits a coastal town. 100 people. Nobody is scripted. Every person in this video decides for themselves, every second, what to do: run for the hill, go back for family, film the wave, or freeze. Built with Claude Opus 5.5 and Jev by @typesafeai 🌊 Watch who makes it 👇

AI engineer Dave shows how to use Jev in Python for typed AI decisions instead of prompt-heavy JSON: • 0:00 - classify support tickets in Python • 6:00 - turn text into Choice, Score + Noul • 10:13 - compare Jev vs Claude on speed + cost • 11:56 - get routing + urgency + escalation in 1 call • 14:33 - use Jev for decisions, Claude for reasoning This is what Jev engineering in Python actually looks like Bookmark this if you’re moving AI logic out of giant prompts

An access ticket came in. Here's the workflow that closed it, step by step, with what each step cost.20 steps. 4 judgment calls go to @typesafeai 's Jev, not an LLM. The LLM runs once. Whole ticket: ₹0.16. Jev's share: under 1 paisa. What ran: 1. Code pulls the repo + ticket (free) 2. Jev triages: config change, 100% conf, 326 ms 3. LLM edits one line of config on a branch (30s, ~15 paise) 4. Jev reads the diff before the push: safe 95% 5. Code pushes, comments on the ticket 6. Spawns a monitor fot git and Jira The pattern: → code first, for anything with an algorithm → Jev for judgment: typed

卧槽？？一个不到 2B、能本地部署的小模型，给真实 Issue 做分类，跟 JEV 的结果居然这么接近？ 最近 JEV 的玩法看了不少，给邮件分类、给 Agent 选工具。这类需求太常见了，我就想知道这张工单该分给谁、下一步该调用哪个工具，真不用每次都给我写一篇分析。 所以这次我把 JEV 和 This That 都接进 Pi，直接拉了 OpenAI Codex 仓库的 30 条真实 Issue，让它们现场做题。 用户是在报 Bug、提功能需求，还是问怎么用？问题出在登录、界面、执行还是会话？ 同一份内容、同一组选项，每条做两项判断，两边各发 60 次请求。 跑完一看：JEV 用了 23.5 秒，This That 用了 25.6 秒。30 条里，28 条的两项判断完全一样，两边都没出现请求失败。 速度这轮还是 JEV 快一点。但 This That 只有 18.8 亿参数，权重约 3.76GB，而且能下载到自己机器上跑。这个结果真让我想继续折腾一下了。 它的用法也很干脆：把问题和选项给进去，一次计算直接返回选择和概率。程序拿到结果就能接着往下走，不用再从一大段回答里抠出一个“是”或者“否”。 更实用的是本地部署。客户工单、内部文档、还没发布的产品资料，这些东西只是想分个类，也不一定愿意全部发到外面的 API。 This That 支持 NVIDIA GPU、Apple Silic

A normal AI agent needs ~20 clicks and ~150k tokens to book this flight. Jev + WebMCP did it in 6 sentences, with under a second of model time. And it still stopped and asked me before paying. How it works 🧵

Jev + @Picsart is the first content system that actually automates my whole upload queue 99% of people still burn a full model call on every posting decision - the 1% running this pay cents instead 6 minutes and the queue is wired: clip → Picsart worker → Jev decision → Picsart execution → posted step 1 → create a Jev API key (typesafe website) step 2 → clone the upload router from Github below, then: python3 -m venv .venv && .venv/bin/pip install -r requirements.txt && cp config.example.yaml config.yaml step 3 → export TYPESAFE_API_KEY='YOUR_KEY' - add it to your shell profile so it survives

JEV CHECKS IF YOUR HOTEL IS LYING I built a terminal that takes every claim made by a vacation hotel in its description - first line, sea view, quiet, recently renovated and cross-checked it against guest reviews. For every claim made by a hotel one verdict and a confidence score in percent. Here is the picture that emerged, Jev checked 40 hotels in Antalya, and only one was left that meets the conditions and it happens to be one of the cheapest. Jev calibrated the hotels, two exaggerations the hotel is eliminated. Then it checked for one exaggeration, and those that failed were eliminated as

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 mentions you. Link in the first comment 👇

Hello Everyone @X I build FormPilot: Chrome extension that lets Jev fill the form. We One click. Green = write. Yellow = you check. Page → scan → Jev (which key? which value?) → confidence gate → fill. Your key. Almost free per form. Reply pilot or DM — I’ll send the repo. Next up from the same bench: • more Jev agent architectures • small tools that decide, not chat #Jev #BuildInPublic #software #founders #AI

JEV is insanely fast. Our AI took 35.83 seconds to make one product photo, so we asked: Can jev make it faster? It figures out what a customer wants in 0.4 seconds. Cost: $0.00003 per message. Now we're down to 16.23 seconds. More than 50% faster. Available now in Dezygn.