All builds

I made Jev play chess against me. How did I do it? Jev can't generate a move. It's a classification model, so it only picks from options you hand it. That turns out to be easy to work around. Any position has a fixed set of legal moves, so I generate all of them in code and let Jev choose one. Then I built a small app with a basic UI and played it. First I wanted a number, so I ran it against 600

@GitHubNext Would be interested to hear how it stacks up? See https://t.co/2c2k0er2NK where I benchmarked Jev vs other models. Context is I used this to use it for realtime decision making to visualise the impact Jev has for these types of use cases. It's fat and still pretty intelligent https://t.co/J1zj0KDfmA

What if an AI that can barely do anything is exactly what makes it 200x faster? That’s JEV. It makes simple decisions in a fraction of a second and can reportedly be 40–400x cheaper. I built 10 of them. Comment “Agent OS” for the guide. https://t.co/a9SMgTRw6T

Spent a little while with Jev, a model which doesn't output the next tokens but it returns typed answers and probabilities your code can use directly. So what did I do? I asked claude for ideas on what to do with Jev. So it built "jgrep": grep where the pattern is a description instead of a regex. Asked both tools for network calls with no timeout. Regex returned 5 hits, 2 of them wrong — it can o

@typesafeai Jev decisions ! An original, inspectable life-and-civilization simulation. HTML, JavaScript and an authoritative Node.js server. A bundled isometric renderer works without a CDN; optional Three.js shows the same world. This is not an EA product or a reproduction of proprietary Sims assets. New voluntary actions require valid Jev decisions. Rendering, navigation, physical consequences a

I hooked up TypeSafe’s JEV to Snake. Every move is a live model decision. https://t.co/UXEUOKvDaQ

I built an AI inbox that sorted 200 emails in 5 seconds. Jev handled the obvious ones and left me with just 23 to review. 500 emails reportedly cost just 3.5 cents. The AI does the boring decisions. You handle the uncertain ones. Comment “Agent OS” for the guide. https://t.co/2g7Ei2vMwy

I let Jev choose the direction of the fabric in a music shader. It's sampled twice every second, reads the song's spectral analysis, and decides the flow over 1296 individual cells. Safe Return by Rob Simonsen https://t.co/iPGXOYqyN5 https://t.co/bsvLaz65mb

jev has changed computer-use forever. i built a computer use tool with jev that costs basically nothing to use, and has sub-second latency. it can even decide what to type, and doesn't need an LLM. open source. link below. no LLM required. https://t.co/eqLWQ3vAa7

1/4 TypeSafe AI has introduced a new "species" of models. The first is Jev: structured answers, not chat. I built HelloJev, a small PoC that sends data pipeline logs to Jev and asks what to do next: retry, fix, or investigate. Here is the demo. https://t.co/PUFQQN9XQG

Got access to Jev - not yet sure what use cases I have for it so experimenting for now 😄 Typing a half-remembered movie description into two things at once: Jev on the left, GPT-5.6 Sol on the right, both given the same 258 films. Every keystroke fires a new call to both. Jev returns a probability for every film in one ~400 ms call, so the wall re-sorts as you type. The chat model has to reread th

I built Svara, an open-source Mac app that helps blind and low-vision users browse the web using Sinhala voice commands and keyboard controls. Built with Astra in Codex from @OpenAI , Google Chirp 2, @GeminiApp , Gemini 2.5 Flash TTS , and @typesafeai Jev. It’s built around Sinhala, but developers can tweak a few settings in the code and adapt it for other languages too even languages macOS doesn’

Bought into all the Jev hype last night, to see what the fuss is about. Made a browser that you can talk to. Standalone exe. Local LLM for everything Jev can't do. Here it is booking a hotel and playing Wikirace at super human speed. https://t.co/NByn3KJZRT

been playing with @typesafeai Jev and honestly - insane. what a time to be a builder found a use case i couldn't stop building: instant model routing. why send every request to your biggest model? Jev scores the task in ~100ms, picks 1 of N, and returns a confidence + a needsReview flag. cheap calls stay cheap - only the uncertain ones escalate to claude/gpt. shipped it as a tiny 0-dep toolkit. je

jev + astra crushing candy crush. > python capture > opencv reads the board: 11ms > valid moves calculated: 0.1ms > jev chooses best move: ~242ms > python drags: 180ms full move: 433ms (without game animations but includes mouse drag). https://t.co/KzzLZc3NRB

i made X-Ray Anything with Codex. paste text, edit the labels, and Jev marks each sentence through classifier dot dev. one HTML file. here’s claim / evidence / guess / instruction. https://t.co/Tww5hQfvy3

Jev is pretty crazy. Classification this cheap makes me want to build all kinds of things. For fun, I used it to find subscription receipts in my Downloads. 1,070 files. 7.6 seconds of Jev time. ~2¢ estimated. It read further into the files it wasn't sure about. https://t.co/YAk5Vs5aY1

jev is insane 🫣 it makes realtime virtual try-on hauls possible. built this experiment for Drape with @typesafeai > i talk > jev reads transcript + what i'm wearing > picks from my closet > changes my outfit in realtime cost: $0.0011 per decision time: ~620ms per decision imagine getting ready like this:

Tested Jev vs GPT-5.6 for predicting my taste in books. I used my 1,000 Goodreads ratings as starting data and held out 100 to test. Then I asked both models to guess which would be 5 stars. Jev was slightly more accurate, 53x cheaper, and 25x faster 🤯 https://t.co/hrqzKMNPwZ

jev'i elimdeki gerçek projenin içerisine soktum bakalım ne yapacak diye bi projemde 200k ürün var, 20 property üzerinden 41 soru sorup eleme yapıyoruz. ben şimdilik kategorilere göre ayrılmış bi 9k'lık csv ile oynuyorum videoda 200 ürünlük bir işlem yaptırıyorum. 12,7 saniye sürdü, 319 istek attı, 0.012 cent yazdı. 200k'nın hepsini döksem 13 dolar filan. şuana kadar jev benim açımdan kendini kanıt

What if a model answers in ~200 ms and costs less than a pixel? You can put it in weird places. I put it in three experiments: 🌦️ A shader that reacts to text. 🔎 A table that filters itself in natural language. 🏙️ A city with a thousand npcs that react to what happens. thxs jev! https://t.co/5pJ8wOV80h

Jev使ってトロッコ問題解かせてみた。 命がかかっているのに即決されるし、選択は再現性あって面白い 車掌がレールを切り替えているのは独自の世界観ですw https://t.co/hilVF7h3Us

I asked Jev to research private AI second-brain tools. 47 seconds. 11 documents inspected. Findings linked to evidence. Built Jev Radar with @typesafeai: watch it choose sources and check claims live. Open source. Bring your keys. Demo at 1× speed. Repo below ↓ https://t.co/2V7RB1CqG7

Multi-second LLM classification is officially dead 💀 Just built a drag-and-drop agent router powered by @TypeSafeAI Jev. 1. Input comes in 2. Jev evaluates intent in ~15ms 3. Directs flow to the exact downstream agent Zero prompt parsing, What do you think? 👀 #BuildInPublic #AIAgents #TypeSafeAI #TypeScript

Experiment 1: I keep a living UI shop for Tiny Design Shop. Audited it against the live site this week using JEV. some Alerts exist only in the shop. The same green pill means beta in the spec and free on the catalog. classic drifts. Documented ≠ shipped. So, i asked @typesafeai's JEV to audit what is shipped and not documented and should i make it into components. it generated a very good audit r

Got access to @typesafeai Jev and the first thing I tried out was the realtime sentiment analysis and context-aware backchanneling with @pipecat_ai . How it works: - Jev evaluates the turn and sentiments. Sending many instructions in a single call, and it just returns immediately - Reads interim text every 100-200 sec while I'm talking and responds to what I say. It handles two things 1. react now

I used Jev (@typesafeai ) as the decision layer in a Pipecat voice pipeline. Three typed primitives, two calls per turn: Noul — a yes/no with a probability Choice — a labelled decision with your own criteria Score — a scalar on a rubric you define Call 1 runs before the LLM. Noul("is this a complete thought?") ends the turn semantically instead of on a silence timer, and Choice("which support flow

Built a profanity checker with Jev. Inspired by the profanity API that @joshtriedcoding built two years ago. He mentioned back then that using AI for this was too slow. It’s crazy that we’re now at the point where AI can handle it fast enough to be practical. https://t.co/c17R7Vf5gC

i built a Jev tool that reverse-engineered what actually goes viral in ai twitter. 18,000 posts. 31 seconds. 71 cents. the same run on opus 5 crawled through a few hundred and cost me ~$400. per post that is hundreds of times cheaper. viral analysis is the perfect Jev job. it is not writing, it is 14 yes/no calls per post: > does the hook open a loop > is there a real number in the first line > is

My toddler wanted to tap on the keyboard. Made this quickly where Jev picks up the next object. Things from @thiingsco https://t.co/kGbzW2to66

Your coding agent is one bad instruction away from running rm -rf ~/.aws. I built Auto-Guard with Jev (@typesafeai) to check every tool call before it runs. Here it stops a credentials delete but lets the build cleanup through. It's open source! Install in the replies 👇 https://t.co/lgcdoEZopv

Tried Jev from @typesafeai on AresSim environment. Overall the Rover's movements until first 50 steps was great. Had to tweak the rules abit to get it work. But it did a decent job (on par with Masked PPO I trained earlier). Making the State form more efficient would also make Jev work better,

天才！ llama.cppにjev互換サーバー組み込んだら、どんなモデルでも無改造でjevっぽくふるまえる！ Qwen3.5-2Bが画像に対する4つの質問に800msで答えている！画像は448x448 https://t.co/krN4OGJMRd

I tested Jev, @typesafeai's new classifier model, on 100 real banking support messages (77 intents, zero-shot): → 82% accuracy → 91.9% accuracy on the 74% of tickets where it was ≥90% confident → 329 ms median latency → $0.009 total Code: https://t.co/HtEjw82e7o https://t.co/zGGiGNGdJa

話題の新AI「Jev」でオセロゲームをつくったぴょん 盤面のうち正しそうなマスをガイドしてくれます AI vs AI戦も実装 強いかどうかは不明ですが、一回のAPI呼び出しはわずか250msで応答と高速。API使用料も表示していますが驚くほど安い！ Web版のURLをリプに貼ります。挑戦してね https://t.co/jW8twW2wT8

Made a GeoGuessr where the only clue is emoji. @typesafeai Jev reads a random point on Earth and scores all 254 emoji in my catalogue. You drop a pin. You're scored on distance and on how many emoji your pin shares with the target — so wrong place but right vibe still scores. https://t.co/RYUfnlMK3y

playing with Jev, i built a small @pidotdev extension that sends my latest Git diff to Jev and asks: → how strong is this change according to a defined rubric? it returns a 0–5 score with confidence: git diff → Jev score + confidence → review signal in simple terms, it's a simple code review, not a replacement for tests or human review, but this gives me a very fast signal for deciding which chang

i recreated the hot dog or not hot dog app from silicon valley with jev https://t.co/kcNpnlkwGT

Jev classified 1,004 TikToks & Reels for about $0.07 in estimated model cost 😂 i'd saved hundreds of videos that went viral or converted well. figuring out why was still a coin toss. so i built a dashboard around Jev. it labels every video across 8 dimensions: hook type, format, script structure, CTA placement, creator persona, and more. i can filter by hook and format, compare views and saves, an

I ran @LeanCodePl Marionette with @typesafeai Jev. One sentence: "win the game." A Flutter game it had never seen. 5 puzzles, rules only on screen. It read the code, flipped the right switches, tapped blue, red, green. 🤯 9.7 s. $0.0012. ⚡ No LLM involved. Marionette reads the widget tree and taps, Jev decides 🧵