All builds

Here’s a very alpha demo of a browser-use agent I hacked together today. It’s built around @typesafeai’s new Jev model, working alongside GPT 5.6 Luna for reasoning. A lot of the work is still being handled by Luna, and I need to figure out how to offload more of it onto Jev to get the iteration speed up. But it actually works. Right now, it’s basically a duct-taped Frankenstein of web search, Lun

TypeGPU + ruNNtime + Jev @typesafeai is a very fun combo :D ruNNtime gives me efficient local inference, TypeGPU lets inference and rendering share GPU resources directly with zero copy. That’s 3 separate NN inferences plus rendering, all happening in realtime Since we control the pipeline, Jev can just sit in the middle and add the semantic bit. camera + mic → Moonshine + YOLO26 + DepthART → Jev

Jev v1.13 dodges rockets with probability calculation 🚀 @typesafeai's new non-LLM model returns decisions instead of text so we had it calculate a safe tile every 330 ms while rockets fell and it survived 25 of 26 for under a cent Run Jev via API -> https://t.co/RbcCOIgVkj https://t.co/f2VmOEPFnO

hear me out… why are we sleeping on @typesafeai 's Jev for retrieval? tested 34 questions from QASPER. Jev went 17–3 against pgvector + OpenAI text-embedding-3-small, with 14 ties, on gold evidence coverage. small-to-medium doc RAG where accuracy matters might be Jev's playground

Built a bit of a proof of concept using @typesafeai jev. The most prosaic of enterprise AI use cases, a chat assistant for navigating your terrible mobile UI. Interprets your question and finds where you hid the menu option. No hallucinations. https://t.co/ylm3zycwKS

I got early access to Jev, so obviously I gave it $1,000 and let it play blackjack by itself. It made 243 hands in ~191 seconds, briefly reached $1,080, then ultimately depleted the $1,000 bankroll to $5. Turns out Jev is fast. Just not a very good gambler. @typesafeai https://t.co/NarPF9o3jB

I used @typesafeai's Jev to label spam comments on my YouTube channel. For the test, I picked 70 real comments, 12 of them spam. Jev missed 3. GPT-5.6 Luna missed 2. Jev: 25 seconds, $0.00114 GPT-5.6 Luna: 79 seconds, $0.00267 That makes Jev 3x as fast and about 2x cheaper. https://t.co/0BBpjKhlOY

2/ Cua Driver + Jev playing 2048. In this measured 2048 run, Jev was 5x faster and about 1,000x cheaper than Astra. That result is specific to this run, not a general claim about either system. Video attachment: 30 seconds, silent, 1920 x 1080. The final frame shows Jev at 44.9 seconds and about $0.00108 API-equivalent cost, versus Astra at 294.9 seconds and about $1.45. Demo by @injaneity

i made a @typesafeai jev+chatgpt instance race against a purely chatgpt instance on two super simple tasks to see the difference, 2-3x faster! :o on first-time tasks chatgpt will write a question set for jev on the fly, and then subsequent tasks get super fast https://t.co/gMBtV3ou3c

So I did a thing with Jev. Saw a few posts about training Jev to play video games. I figured, why not try to get Jev to play a song? Found a piano version of Bella Ciao. Put it through a piano transcription model. That gave 1,353 notes with their timing and loudness. The timing and loudness became the fixed level layout. The pitch itself was hidden from Jev. The task: At each note, Jev had to pick

I said I’d come back when I was done. I’m done. 😂 I took the Jev-powered experiment, rebuilt it with Jev + Grok Build, and turned it into parody. You can now use it to create parody versions of posts, comments, and basically any social interaction using the identity/context of an existing account. Search for an account → write your parody → generate it → share it. The whole point is to make fictio

Jev PoC on an iOS device remotely. @typesafeai great work, OCR is the bottleneck in speed. https://t.co/hwH4nTyJZu https://t.co/fGwQKh4SFq

Now we're cooking. Got @typesafeai's Jev + AXe controlling the iOS simulator now ultra fast at a fraction of the cost of using an LLM. This is game changing! https://t.co/JK5r23ySoI

Can Jev play Mario? Made this quick demo to check it, and from 10 tries it was able to get to the finish line only once! Most probably the issue is within the prompt that I am sending, as it is not being able to properly calculate the statistics from the information given. Will polish it further....lets see!

i used @typesafeai 's jev in a wine app i made, to check web search results. you ask where to buy a wine nearby. search finds "Steep Ridge Zinfandel" when you wanted "Ridge Zinfandel". Jev just answers: same wine, yes or no? 33 tricky ones. got all 33. little animation of it deciding 👇

I built a chat bot with jev, no LLM at all! Responses are instant, no hallucinations. I hooked it up to web search, wikipedia, weather, todoist and home assistant. Jev decides what tool to call and what args to use based on the prompt. Instant answers cite sources as well! https://t.co/owgeOwMFbZ

In my demo, Jev can play 2048 but makes some beginner mistakes without an advanced harness. It's so fun to watch https://t.co/HBqjIPNxqX

I built a trading agent with the all new @typesafeai Jev. You can try it here : https://t.co/e7bLjnCfTj. Just put your own API key and start playing or clone it from https://t.co/7nJeuB2H8t and test it locally. To my surprise the agent was able to take a trade of 2.405 $ETH and took a profit of roughly 0.23%. It is very low but testing a new model in town is genuinely crazy. @CompleteSkeptic did y

had to try something with jev so i tried what you'd expect from me: graph extraction - score every word with semantic significance (1-5) - also tag relevant words with ID - unique list of high scoring words - graph of high scoring words https://t.co/oJtI6py84H

some UI things this morning, character expressions driven by screen state and Jev (typesafe ai) https://t.co/LTzPvET9Fx

I gave Jev 3,282 of my X posts across 100M views and asked it to find what actually works for growth. 4,252,330 tokens $0.1282 for the full 8m 34s run! Each post got 8 questions about the topic, hook, tone, whether it teaches something, etc. How-to posts got 150 median likes vs the average median of 44. AI and coding was a 1.9x multiplier topic compared and SEO, despite recent posts, was right at

made a Jev-like classifier model + pathfinder to control my enemy AIs in my capture the flag minigame built in @threejs https://t.co/GkZujHExNG

Introducing jev board, short for jevin keyboard. Fast, reliable models unlock much subtler experiences for building ambient intelligence into interfaces. A small exploration at the edges of GenUI. With done @typesafeai and @CompleteSkeptic https://t.co/C5D6lojjGH

I made a toy CLI for talking to Jev npx @fernandoviac/judged https://t.co/cGTFFqj4xO

Jev for soccer ⚽️ Built a Chrome extension that watches live football with me. It scrapes the game page's commentary + stats, and every time something happens it sends Jev (@typesafe_ai) one batched request with 11 questions at once: who scores next, how it ends, is a goal/penalty/red card coming, who has momentum. Next step: adding an LLM on top to push the analysis further.

I gave an evolution simulation to Jev, a small model from @typesafeai that answers typed questions with odds instead of writing text. Two species with opposite DNA. 14 generations. An ice age. 432 creatures. Jev decided who survived, who mated, and what killed each one. https://t.co/Law8fyfy6d

🧵 I gave Jev $10,000 and let it trade BTC again. But this time, I gave it everything a trader would look at: market data, derivatives, macro, on-chain data, news and sentiment. 30 days https://t.co/6STcUk4ml1

Added a live eval system to my webmcp native component library, which I have been using for generative ui and consulting. It evaluates the component as soon as it's generated and if it's not valid, regenerates instantly. Left side is a chat agent with Claude, right side is the Jev Eval loop, webmcp tool calls.

Jev-style "typed decisions" without training anything. On a frozen Qwen3-4B: turn each schema field into a lettered question and read the option letters' logits. No output tokens. Same accuracy as grammar-constrained JSON on closed enums. 4× faster on short inputs, up to 2.4× on long ones with a shared-prefix cache. Strings and numbers still need generation. Write-up, figures, teaching bench: http

Built a pumpfun grading bot with Jev! Every new launch, Jev grades its chance at graduating. If it's over 10% it buys. Starting with a $10k paper account. https://t.co/4t9Xa6pAaL https://t.co/gJ6kUdPnLD

First quick test of Jev from @typesafeai Install via @vercel gateway Sending evaluation (input validation passed): { "model": "typesafe-ai/jev", "state": { "character": "2" }, "questions": { "kind": { "type": "choice", "instructions": "Is the character a number or a letter?", "criteria": { "number": "The character is a single numeric digit from 0 through 9.", "letter": "The character is an upperca

a binary event contract's price IS a probability. 0.62 = the market says 62%. so i pointed a @typesafeai jev at @dreamDEXSomnia to disagree with the book. @Somnia_Network reactivity pushes every event, whole loop under a second. 70k events tracked in 10 sec with offchain websocket. 0.212 brier vs the book's 0.311. testnet. mainnet next.

I built real-time Clippy with Jev. It quietly watches how you use the product and only wakes up when it thinks you’re struggling. Hesitating? Confused? Stuck? Clippy knows. Even its reactions are controlled by Jev. 👀 https://t.co/EgMhw8FPi1

I build a SQL extension that turns plain english into a WHERE clause with jev @typesafeai WHERE jev(people, 'could work from home') Every row gets judged individually, no index and no embeddings needed. Try it out (don't burn my wallet pls) https://t.co/4nYZVH9iXs

Can Jev bluff itself? I spun up two Jev agents, gave them 1,000 chips each, hid their cards from each other and let them play heads-up poker until one went broke. 20 hands. 153 Jev decisions. Biggest pot: 1,760 chips. Both got caught bluffing. Red Jev took all 2,000. The entire match cost $0.02. @typesafeai

Jev made a coin. Jev is TypeSafe's System One model. It answers in types, not prose. So its first coin is typed too. One schema, validated once, sent to two chains in the same request. $JEVCOIN Solana, on Pump CA: A66FqchgzB8PCv9smWYvDctB4tWBefoEMnCecLvfpump Robinhood Chain, on Pons v2 CA: 0x8C0B9EaE5a2aF9680968c4d46fB43266491681cC

open sourced jev-lint, the linter i run on my own markdown wiki. one run over 99 pages: 16 flags for me to read, about 2 cents in input tokens. it never edits anything. agents wrote most of it, i set the constraints. #AIAgents #Obsidian https://t.co/5mrgk9AT8R

Dowse is a terminal-native web browser + answer engine. Search the web, read pages, follow links and generate cited answers — without leaving your terminal. I’ve added @typesafeai @CompleteSkeptic Jev as an opt-in System One layer in the answer pipeline: Search → Jev → LLM Before generation, Jev makes fast, structured judgments over the retrieved sources — relevance, usable evidence and prompt-inj

Jev only outputs probabilities. So I gave it the Snake board as text and asked one question every frame: up, down, left, or right? No training, no game code inside the model. It played fine for a while, then boxed itself with nowhere left to go. https://t.co/e6lGyVTrlk

I taught Jev to paint! Jev predicts a probability for each color of each cell, then we visualize the probability distribution: high confidence = big, flat strokes; low confidence = fine brush to layer different possibilities https://t.co/L99f5txwha