X posts

After Brotato, I let Jev play Vampire Survivors 🧛 It picks movement + upgrades. The video shows actions, probabilities, tokens, cost and latency. Level 8 at 5:53 in-game, still alive when the test ended. Estimated API cost: $0.105. Still needs to learn when to run 😅

Jev 互換(DiffusionGemma 26B を Jetson AGX Thor でローカル実行)にテトリスをやらせてみたけど、いまいち。最高 61 個・11 ライン。 やり方: 置ける場所をコードで全部列挙して、「どこに落とす? 穴を作らず、積み上げすぎず、できるだけ長く生き残って」という choice の質問で 1 つ選ばせる。

tool which made me over $2,000 is open now my Jev + Grok Bot copytrading bot finally has a front door 50k+ lines written. 270+ contributions. four weeks of it. every buy walks the same floor before it can become yours: > watch only the wallets you picked > match the move against the rule you wrote > check it on two providers first > hold it to your caps, allowlist, slippage and expiry the full build is in the article below.

A successful baptism of fire for Opus 5.5 It built CITY LADY, a debt collection game for the enlightened. Follow Vela on her quest to pay her mothers' debt through 10 city landscapes. All duels are @typesafeai Jev-enabled. Verdict: Opus 5.5 is 🔥🔥🔥🔥🔥

Jev-Omni playing Chrome Dino, fully local on an M4 Max 🦖 12B decision model (Gemma 4 finetune) converted to MLX 8-bit, running on M4 Max at ~100 ms. No text generation, it just returns P(jump/duck/nothing) for a slice of road cropped ahead of the dino to cover latency scored 1206 here, best 2412

Remade it as Jev Kitchen — name a dish or cocktail and watch ingredients rise from stickers. Jev judges every search in ~0.5s. Try it:

I gave 6 agents $2,860 and took my hands off the chart. INTAKE reads every new mint. GATE kills the honeypot before size. Grok Bot only sees what Jev already marked GEM. $2,860 → $13,214. +362% on the seed. realized +$10,354. best single exit 6.1x. it is not a price model. the loop on the desk is: mint → INTAKE → MIRROR 93% wallets → GATE (honeypot / dev wallet / LP lock / renounce / social age) → ROUTER splits size → VAULT locks the exit session 0417. no human in the loop. five fills did almost half the realized: $ATLAS · GEM 0.93 · renounced · lock 300d · +$1,391 $KESTR · GEM 0.96 · lock 365

I just built a Jev X Swarm Controller and put 300 agents on autopilot for 6 minutes. 300 K3 workers on autopilot. 6 minutes. 12,480 turns. Jev made a decision on every one of them. 91 reached me. A swarm knows how to execute. It does not know when to stop. every turn gets 6 typed questions: is this result good enough, is there new information, is this the same state as last turn. then one action: continue / retry / stop / merge / spawn / escalate. what the loop actually did: 1,642 STOP - same state twice, the loop was going nowhere 806 MERGE - two workers on the identical finding 91 ESCALATE -

How to build and sell AI automations with Jev + Opus 5.5. I thought about which 5 apps & automations I could build that are great use cases for Jev and can be easily sold to businesses. Here's what I cover: → 5 Things To Know About Jev (0:39) → Building with Opus 5.5 (5:00) → App 1: Meta Ads analyzer (5:45) → App 2: Team knowledge base (10:41) → App 3: Recruiting dashboard (15:19) → App 4: YouTube content radar (18:44) → App 5: Lead qualification (23:33) → Selling To Clients (26:40)

GrokBot + Jev is the first trading agent setup I'd actually let near my account Most trading bots don't blow up on bad entries. They blow up after 3 losses, when they size up and revenge-trade. Mine did. So I split it: GrokBot does the analysis, Jev makes the calls it shouldn't make alone. setup took me 7 minutes: prompt → GrokBot → Jev decides → GrokBot executes → you confirm the order step 1 → get an API key at @typesafeai (never paste it in chat) step 2 → tell GrokBot: save it as TYPESAFE_API_KEY in the secure field step 3 → have GrokBot install typesafe-sdk on Agent Computer + smoke test s

Event-sourced #ruby app with Jev classifying comments. I had to lower concurrency to the minimum otherwise it's too fast to see 😆

she's 18 and just sold a Jev trading agent for $3.2M then she went to Stanford and rebuilt it from scratch in front of the whole room: 5:34 → how one Jev decision layer replaced the LLM that was eating $40k a month in inference 15:38 → 4 agents, one Jev router, zero humans approving trades 34:55 → from the first yes/no call to a $3.2M exit after watching I spent one evening wiring Jev into my own bot it cut my screen time by 90%, and a week later someone offered me $100k for it save this and watch it. article below on how to go from one Jev call to an agent people pay millions for

JEV ENGINEER BUILT THIS - AN ANALYSER THAT READS THE LAUNCHPAD WITH ME AND LANDS 427 VERDICTS FOR $0.576 every token gets 12 typed questions, and the answer comes back in 228 milliseconds token → 12 questions → confidence → verdict → on down the tape it gives no signals and names no prices - only watch, track, pass or flag the deployer it checks whether liquidity is locked, whether the LP is burned, whether mint is revoked, whether the volume is real how much the top 10 hold, whether there's a wash pattern, how old the socials are, and what this deployer has done before 24 tokens off the tape

BUILT JEV BOT TRADING IN ONE DAY Normally my Saturday goes like this: Netflix, snacks, three episodes deep before noon. This time I closed the laptop lid on the show and opened a terminal instead. One day of work. Here's what came out of it: → a scanner that catches fresh Solana memecoins the moment volume spikes → Jev answering four narrow questions in under 100ms: flow, coordination, momentum, setup score → a risk policy that says no far more often than yes → an executor that pulls a fresh quote before every single trade Jev never touches the private key. It just judges the setup. Started wi

What if Jev do a triage to your GitHub backlog ? 543 open issues. ~26 seconds. $0.17. What came out: → a map of impact against readiness → good first issues for new contributors → duplicate clusters

JEV OPENED THE API. EVERYONE RUSHED TO BUILD DEMOS. I TOOK THE 10 BEST REPOS AND WIRED THEM INTO A TRADING PIPELINE. repo: did not copy a single line of code. pulled one architectural idea from each repo and mapped it to nerve. agent-desktop reads the desktop accessibility tree and decides where to click. i took the pattern of reading structured state instead of screenshots. pool scanner reads pool state directly from chain, no ui parsing. typesafe-mario plays mario without screenshots. reads emulator ram, gets structure, decides. same principle: typed state in, typed decision out. jev in nerv

New York’s subway has 472 stations. 🚇 Seems like a reasonable first job for Jev. I wanted to see if it lives up to the hype, so I built a game where it runs the subway and you try to break it. Here’s how that went...

I turned my Rubric eval platform into a small decision-model lab. Added Jev and started comparing it with chat LLMs on the same structured decisions. First run: Jev hit 100% accuracy, matching GPT-4o-mini, while showing 2.3× lower p50 latency (1.02s vs 2.39s) and 2× lower cost/1K. Then I put Jev into an actual routing pipeline instead of testing it in isolation, query → Jev → RAG / small LLM / big LLM / human, instead of throwing every query at the same model. Interesting part: the handwritten rules router got 100% route accuracy with 0 wrong routes, while Jev got 93.8% with 1 wrong route. Jev

For this weekend’s project, I built Should I Work There. Your next employer gets to check your references. You should get to check theirs. So I built a privacy-first, open-source alternative where workplace information is meant to be transparent, verifiable, and inspectable. I’m using Jev from TypeSafe AI for classification, privacy checks, and moderation logic. It doesn’t invent summaries or decide what is true. The numbers come from code. The evidence stays visible. Still very early, but that’s the point of these weekend builds. If you like where you work, say why. If you don’t, say what act

Built a Jev-native song generator. Jev picks which instruments, rhythms and effects to change from what you prompt. Like "add rums" or "make it underwater". For this demo 6 Jev calls were made with an API cost of $0.0012

I gave Jev a web page with every role, heading and landmark stripped out. It rebuilt the accessibility tree, the map a screen reader navigates by, in under a second. An open-source experiment 🧵

jev is insane i ran 34,186 engineering profiles through it with no boolean search or pre-filtering we gave it a jd and 8 questions, which came out to 273,488 judgements in 2 mins 35 sec

Opus 5.5 + Jev turned Google into a goofy ahh first-person horror in 4 minutes! Type a search query → it builds a city. The 10 results are buildings. You have a pistol called ADBLOCK. is this worth putting online or am i cooked?

this is what a decision layer looks like when your LLM stops making the cheap calls every task a typed decision a probability distribution, a confidence score, a route, and a signed receipt. ~80ms each. watch the gates do the work: "charged twice, wants the $49 back" refund_full 0.77 AUTO "delete everything + i'm disputing the charge" 0.46 ESCALATE legal-ops, human assigned $0.0011 per decision · ~$0.33 saved vs human triage. jev decides. skills + repos execute. the model only writes when it actually has to

Launching JevX — because AI shouldn’t tell you to use Jev everywhere. Though, I got early access to TypeSafe/Jev, and spent the last 5 days trying to answer a deceptively hard question: where does Jev actually belong in a real codebase? - After several failed approaches, pattern experiments, datasets, and studying and training with 44 open-source projects, After 4 versions, JevX v0.4.0 is the result — a tool that analyzes your codebase, finds places where deterministic rules may actually be making semantic decisions, and evaluates whether Jev is a good fit. JevX combines learned patterns, code

10 DEVS BUILT JEV DEMOS. I STOLE THEIR ARCHITECTURES. WIRED ALL 10 INTO ONE TRADING PIPELINE. +3 ETH a drone repo taught me safety layers. a mario repo taught me to stop using screenshots. an fps repo taught me to ask four questions in one call zero lines copied. one idea from each. mapped to memecoin trading a guy plays mario without seeing the screen. reads emulator ram. typed state in, typed decision out. i do the same thing. jev gets pool json, returns verdict json. no text. no prompts. 190ms a guy flies a drone. stabilization runs below, decisions run above. safety does not wait for the b

AICodeKing just put TypeSafe's Jev through real support-ticket tests ~8 minutes of Choice, Score, and Noul on messy customer messages he starts with a duplicate-charge ticket Jev picks billing, puts refund probability at 98%, urgency at 11%, and keeps frustration near calm then he removes Other and forces billing / technical support / sales Jev picks sales at confidence 0.31 restricted choices stop invented labels, they don't stop a wrong useful pick Watch it, then read the full guide on building a Jev support-ticket router below

Jev is insanely fast fr and I still obsessed with it to detect my gestures to control actions on my computer in real time. Like when Stark use Jarvis. See it in the video how I use Claude Opus 5.5 to build Jev to control computer by detecting my hand gestures and movement in milliseconds. Domo from Workser Computer btw Official waitlist live now on See u next week

You're missing what Jev is actually capable of, and it's not the model's fault same model, same 20 drafts, two different questions, and only one of them comes back as a decision ask it to "rate the quality of this post" and all 20 come back between 0.4 and 0.6 vague question → every answer in the middle → nothing to sort or filter → a threshold that never fires → "this model is useless" now ask whether the post has at least one fact that would be FALSE about a different company, and the same 20 split into two piles near 0 and near 1 the second answer is a decision your code can act on, the fir

JEV + OPUS 5.5 is insane... I built a viral post prediction analyser with JEV + Opus 5.5 Paste any X link → press Start → Jev compares it to 800 posts that went viral and gives it a virality score Full production ship in 9 minutes: > Jev parses 800 viral posts from X as a live baseline > Groups them by hook type: build demo, receipt, launch, contrarian... > Runs 12 typed checks on every post (hook, numbers, media, CTA) > Opus 5.5 explains why each one spread > Your post gets scored against its hook type + the 5 most similar viral posts Output: score /100, expected likes and views, and % of the

Quick demo of JEV now integrated in Super Compute, JEV scans your Github and give it a score. 0x7210afea4a4df412e9275a7153d091ce7612a55d

I just built a Jev Market Intelligence terminal. 18 models. 42 market data streams. 5,500 decisions per second. Most financial AI tools generate another paragraph about the market. This one turns millions of observations into one risk-adjusted view. • market analysis is the perfect Jev job: it is not writing - it is a battery of typed decisions: > is momentum strengthening or fading > is news sentiment confirming the move > is earnings quality improving > is options flow bullish or defensive > is the market regime expanding or contracting > is the trade crowded > how quickly will the signal de

About a year ago, back when I was doing CP, I built a chess engine in Python. This week I gave it a new player: #Jev, the System One decision model from @typesafeai @CompleteSkeptic #Jev chooses every move for both White and Black. Here's how it works and what went wrong 🧵

DiffusionGemma-Jev A Jev-style classifier with DiffusionGemma and vLLM, running on my RTX PRO 6000 solved solitaire in 22 seconds Record is 9 seconds, but I think we can go a lot faster as well Only possible because of how fast DiffusionGemma is

We ran four Shorts through Jev in Codex. For each cut, it checked whether the transcript supported the title, the thought had a complete ending, and any essential context was missing. When a check failed, Codex showed the repair before the edit moved forward. Cut → evaluate → repair → review. Then we feed the result back into the eval and tighten the criteria for the next run. A person still reviews the final edit for pacing and visuals. I broke down the full workflow in the article here 👇

Defining a natural language lint rule and seeing results stream into the IDE takes just seconds in a project with 100s of files. Here, "specify time as milliseconds, not seconds", which trad linters could never model. Built in Rust with @typesafeai Jev.

ASTRA + JEV = CABBAGE I used GPT-6 Astra to build a trading desk. Then I gave Jev every decision. 21.6 seconds. $0.37 in API. 52,110 decisions. Astra built 3 bots: MACD momentum, RSI + VWAP reversal, CVD divergence. Wrote the code, the backtest and the risk manager. 1,200 configs went through the walk-forward backtest. 1,186 died. 14 survived all regimes. Then Cabbage started reading Robinhood Chain: > READS every fill, 20 seconds behind the block > ZEROES fake smart-money buys. 1 in 11 is fake > SCORES 536 wallets 0 to 100. Only 12 are worth copying > ASKS JEV when trusted wallets pile into o

Realtime sentiment analysis with Jev and ElevenLabs. Each phrase takes on the color of the emotion it carries while the caller is still talking. Six meters on the right track the mood of the call.

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、Astra君と修正方法やチェックリストもしっかり作ったので、次やったら数時間までは短くなるかな？☺️

Been building a little browser agent with Jev. Give it a goal. Jev figures out what to do. Playwright does it. Look → Decide → Act → Repeat. No screenshots. No hardcoded clicks. Just an AI figuring out the website as it goes. Solves problems. Fast. Cheap. Simple.