X posts

Jev + on-device models = results in seconds with no LLM in the loop. Quick demo app to show the possibilities. Drop in an audio file: Ear detects the language, Voz transcribes it and Redact removes PII. Then @typesafeai's Jev makes about 20 decisions in one call in milliseconds, and picks which of our on-device models to run. Voice memo to to-do list. Meeting to redacted transcript. Podcast to clips.

This weekend we took part in the @PromptQL "Rebuild With Jev" hackathon and won! Thanks @tanmaigo and @rajoshighosh for a great event. In 90 minutes we asked one question: "Can Jev, introduced by @CompleteSkeptic from @typesafeai supervise a robot in real time?" Full write-up, with the @rerundotio recordings: P.S. From idea to experiments to the blog post we never left the @PromptQL chat window. The VLM and Jev calls, the @rerundotio viewer, @huggingface datasets, the writing, the video and the review agents all ran inside one bot's VM. That's the robotics harness we want for end to end agenti

18 turns + 1 fruit VS 1867 turns + 67 fruits. What's the difference? Both Snakes controlled by @typesafeai Jev. The one on the left provides Jev with a strategy to play snake and then asks what to do: left, right, forward. It plays 6-18 turns and eats at most one fruit. The one on the right, breaks down this strategy into 5 questions and plays for 1867 turns and eats 67 fruits. The questions per available action are: 1. Is this going to lead to a crash? 2. Does this bring us closer to food? 3. Does this move bring us closer to a dead end? 4. Can the head still get to its own tail? 5. Is the ro

JEV is INSANE. 4,000+ people applied for internship at @tryAlanAI. So we built an in house review console to find evidence of real work at scale. Every square is a real SDE intern application > 3,518 applications checked in 4:54 > 20 surfaced for human review > Estimated Jev cost: $0.4125 You can watch the shortlist change in real time ↓

My first "holy shit" moment with GPT + Jev: I put real-world personalities into a simulation as autonomous agents > Marc Andreessen: pushes techno-industrial acceleration > Ilya: derives mathematical boundaries for safe AI > Peter Thiel: attacks institutional and economic stagnation (this feels a lot like inception already) Agents have System 1 + System 2 thinking. Im seeing what looks like emergent personality + ideologies, with surprisingly consistent behavioral patterns None of this behavior is hardcoded Every action, discovery, and collaboration emerges autonomously and gets written into t

I GAVE CLAUDE A BRAIN THAT DECIDES HOW HARD TO THINK Not per session. Per step. Jev is a classifier that scores each move 0 to 1. Claude Opus 5 shifts between LOW, MEDIUM, and HIGH reasoning mid-run. Reading files. LOW. Twelve steps in a row. Then it cross-references modules. LOW to MEDIUM. Final analysis across 1,176 lines. MEDIUM to HIGH. 6 real bugs. 14 steps. $0.38. 160 lines of python, no framework, and one API beta you've never seen in a demo.

since jev (@typesafeai ) dropped i saw a lot of ultra fast computer-use claims but no one open sourced anything. i built a voice-to-computer-use app for mac using no llm. only macos native speech to text and jev to choose the next action out of all the possibilities. it works.

In one day, I went from complete fucking noob to... a Jev power user? (Read this, and you can too!) I've been playing with Jev, a new AI model from @typesafeai. I'm not a developer, so I wanted to find out what I could actually build with it. I made four increasingly crafty/interesting demos: An email sorter 2,000 test emails sorted into buckets in 4.4 seconds, for about 4 cents. It flags uncertain cases for a human to review. A customer-review playground Would this person buy again? What should we improve? How did they feel about the product? Ask all three and get answers your app can use. “S

What if Jev controlled your feature flags instead of static conditionals? I built a demo using Cloudflare Flagship and Jev that dynamically changes the users UI based on their personality. Configure flags in the Cloudflare Dash -> Jev semantically evaluates them based on the user's answers.

at @MetaviewAI, over the weekend we shipped @typesafeai's jev into every agent on metaview. candidate searches in our sourcing product that took minutes now take seconds. same accuracy, ~10x faster, meaningfully cheaper per search. but the speedup isn't the interesting part. jev let us treat intelligence like software: break each agent into its smallest semantic units, query each one, set our own thresholds, and fix bugs by adding a question — not by begging in a system prompt. that's a different way to build, and it's now our baseline. credit to the typesafe team. optimizing for intelligence

Hooked up r-1 to Jev and now I can ask a PDF what I’m looking for and watch it light up on the page in real time. r-1 gives us the text, layout, and where everything sits on the page. Jev finds the relevant bits. Parse once, then keep asking questions. No embeddings to generate, no search index to maintain.

Ross(@TheCoolestCool ) posted this and I could not leave it as a bookmark Jev by @typesafeai for SEO internal links cannibalization thin content intent keep / update / merge / kill redirects schema who AI actually cites same model I used to score 1000 jobs now for the short-term rentals bix pointed at pages (extracted from sitemap) I built a prototype today screencast attached it is early but it looks promising do you think Jev could disrupt SEO / AIO tools? or only the part where you stare at a crawl and decide what to do?

I built a free internal linking tool using @typesafeai Jev for classifying and selecting the links. BYOK or pay $1 to use mine. It works for up to 500 pages and gives you a CSV or JSON to pass to an LLM to implement.

652,200 people watched my jev news scanning demo last week. today it's a real product: the stories you can jump on every morning, the angle you can take, and the journalists already covering it. delivered to your email or team slack, every morning 7am. in my demo jev read 384 headlines in 24.9 seconds for 19 cents. opus 5, same feed, same clock, got through 4/384 at 390x the cost per headline. on friday i shipped a slim open source version already in v0.1.19. so why am i also shipping it again? three things i learned building the free one: 1. the news is shared every monitor reads the same new

I knew you all were waiting for this so here it is combining two of the most hot topics going on the tech twitter in one place when a Fly meets Jev in DOOM go ahead and go bonkers on this. it uses my Jev API key (so I would love to see those tokens go brrrrr) it is still a work in progress, and I'm still thinking in which direction should I take this project. also it is open sourced so you can go through the repo (in comments) and I would love to have feedbacks on this :) right now it is still in training and with very playout learns how to perform better. and a thing that I have noticed is th

8 Grok agents produced a full trade thesis. JEV was allowed to return only three values. In this concept simulation, the agents proposed $2,260 of DOGE, liquidity controls reduced it to $520 and the final router returned: decision: REJECT confidence: gated reason: stale_context Three fields turned a persuasive discussion into something the rest of the system could actually enforce. FENCE removed permission, RELAY sent nothing and the broker-request counter stayed at zero. That is the key difference behind Jev. TypeSafe built it around typed Choice, Score and Boolean outputs rather than unrestr

I let Jev play Brotato 🥔 9 waves cleared for ~$0.09 in API calls. 244ms median response. Watch it dodge enemies and make shop decisions, with its actions, probabilities, tokens and cost shown alongside the gameplay.

Jev + Grok Bot is the best agent setup I've built so far it's cheaper and faster than 95% of agent stacks i've seen, and the setup takes just 5 minutes: your prompt → Grok Bot → Jev decides → Grok Bot acts → result step 1 → go to @typesafeai and create an API key. don't paste it into any chat step 2 → ask Grok Bot to save it as TYPESAFE_API_KEY in the secret field step 3 → have Grok Bot install typesafe-sdk on its Agent Computer and run a quick system_one test with one Choice question step 4 → ask Grok Bot to build a small usage lab: router, dry-run mode, config and logs. or just clone my repo

Launching Jev on Command Code for free today. Jev x Command Code is super interesting. You can only use Jev in headless mode or Provider API. Also open sourced a mod "cmd-mod-jev-nudge" Use Jev for free to keep your agent running as long as the there's more work to do.

The racers in SCRAPS have more personality now thanks to JEV from @typesafeai. JEV helps them decide when to pass, defend or attack, while adaptive difficulty adjusts their pace and aggression around how YOU drive. Still learning? They ease off and give you room. Driving like a god? They push harder and look for chances to wreck you. The goal is to keep you challenged as you improve. Here’s a race with the Solana Degen truck, a flamethrower, @Pumpfun rear decal and @solana boosters. Really happy with how that boost trail looks and fades out behind the truck. I show the @fomo decals in the gara

GROKBOT + JEV JUST DID IN 1.7 SECONDS WHAT A LOOPING AGENT COULDN'T DO IN 53.8 SECONDS AND 24 TRANSACTIONS every chatbot agent on-chain has the same bug. tx reverts. it retries the same broken tx. burns gas. learns nothing. grokbot on pons.v2 just fixed it. not with a bigger model. a cheap policy layer called jev sits in front of every decision. demo one. 12 new launches ranked in 0.2 seconds for $0.00001. demo two. 24 candidates filtered cheaply, only top 5 contracts opened. 2.0 seconds, $0.00020. demo three is the one that matters. a $rugz buy tx reverts. loop-prone agent retries twice. burn

JEV FILTERS 40 BILLION DATA POINTS EVERY 15 MINUTES. AN LLM ON EVERY RESULT COULDN'T DO IT a research team pushes that volume through their pipeline. the first version put an LLM on every result. it worked, and it was unusable: millions of explanations nobody read, latency stacking at every stage, inference bills growing with the dataset. so they moved the LLM out of the hot path and put a judge in it. jev doesn't explain the data. it scores it and decides what the system investigates next. this is the telemetry from that redesign. p95 latency before and after, stage by stage: planner -24ms, r

If anyone ever says I don't ship, this is my SECOND launched product today! Introducing Built on the @typesafeai Jev model everyone's talking about. Input a CSV/Spreadsheet...ask a question about it, and Columns answers every row in seconds. Try it free. Shoutout to @AleksDoesCode for launch repo. I couldn't keep up with all the launches if it wasnt for that!

Jev is so fun! i've made a chrome extension that lets you tell it what kind of content you like and what you hate and it just removes it from every page in real time built it with the new opus 5.5 + jev model, called it wilwid (what-i-like-what-i-don't) we're so close to everyone having their own personally curated internet and honestly it's kind of wild that this wasn't a thing already

JEV MADE ME $3,031 WITH A 7-MINUTE AI AGENT SETUP Prompt → GrokBot → Jev decision → GrokBot execution → Result Jev decides what actually needs to run. GrokBot handles the execution directly on your computer. → Less wasted calls → Lower cost → Faster agent workflow Create a Typesafe API key, install typesafe-sdk and smoke test system_one. Build the usage lab with router, dry-run, config and logs. Add jev-usage-router before browser, research, retry or extra bot calls. Start in shadow mode and check the logs. Once you trust the routing, flip it to active. Now GrokBot follows the route. Jev decid

jev-1 from @typesafeai is a new and interesting model and it's lightening fast <150ms. it provides typed choice, probabilities and confidence out. people have been building some interesting stuff with it from agent web browsing, game playing, order booking, search, classification, binary decision making (spam/not-spam), multi-choice, scoring and more. i made this simple tic-tac-toe game where you can play against jev: it's fast, fun and cheap. jev is free on @vercel's ai gateway until the end of the week 🚀

Hacker News' "Show HN" is where builders launch to the internet's harshest crowd. ~98% flop. ~2% hit the front page and get seen by millions. I asked an Jev AI to guess which from the title alone — 5,200 real posts, $0.077, zero cherry-picking. It ranked them right... but got its confidence completely backwards. 🧵👇 #BuildInPublic #AI #Jev #TypeSafe #TypeSafeAI #Software #AIExperiments

Jev can pick a full shadcn design system in 250ms. Almost zero cost. I took the shadcn preset catalog… → turned all 10 parts into fixed option lists (style, neutrals, accent, charts, fonts, icons, corners, chrome) → wrote a plain-English description of every option so Jev knows what each one looks like → Jev answers all 10 in a single request and code encodes the preset. Nothing is generated. Type "luxurious" and the whole thing re-themes while you're still typing.

🎙️ You can now talk to GeoLibre, and it talks back. Voice command mode just landed in the GeoLibre AI Assistant. Click the mic, or hold the spacebar to push-to-talk: "Zoom to the counties layer." "Make the basemap dark." And it's faster. Simple map commands are routed by TypeSafe's Jev instead of a full LLM round trip, cutting multi-second LLM execution down to sub-second, 10–20x faster. Anything complex still falls through to the full model. Here's the demo, unsped-up: watch how quickly the AI Assistant responds to each command. Documentation: GeoLibre Web: GitHub: #GIS #Geospatial #OpenSour

Real-time answers to "Is this video OK for kids?" on YouTube with Jev, YouTube API, and MobileCLIP. It costs ~$.60 per 1000 videos with .66 second median eval time. Built with Claude Fable 5.1 Low in about 4 hours. Limitations: Runs locally on my Mac. Chrome extension (web only). API calls. Rubric is based on my preferences. Possibilities: 🤯

two builders put jev and claude on the same task, one slide deck, same brand guidelines, live on stream: "we were very depending on the mood of the model of that day" 16 seconds and 3 cents against 3 minutes and 50 cents. opus walked 10 self-checking steps and burned 70,000 tokens writing it, the typed path spent 3,000 in 45 minutes they show the split that fixed it: jev picks the layout out of a fixed set of options, the LLM only fills the blocks, and the guidelines stop depending on the model's mood that day "it is a decision model, it is not replacing Claude or ChatGPT, it will enhance it"

Claude Code ran out mid-task. Codex finished it from a 3 KB note: 7 of 7 checks passed. Same repo, same Codex, told only "continue": 5 of 7. The note is written by Jev, not Claude, because once your limit is gone Claude can't write it. Claude Opus 5.5 sorted the same 112 session messages for $0.86. Jev: $0.003. What baton hands the next agent: → every ask you typed, word for word, in order → which ask changes an earlier one → decisions and findings, narration dropped → where it stopped, git status, the last failing test Same output? Jev and Opus 5.5 made the same keep or drop call on 92% of me

GrokBot + Jev is the fastest agent stack I've run on my own machine My agent was paying a frontier model 4 cents and 11 seconds to pick a flight. Now a router makes that call in 0.4s for $0.00002 setup took me 7 minutes: prompt → GrokBot → Jev decides → GrokBot executes → you approve 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 system_one step 4 → build the usage lab (router, dry-run, config, logs) - or clone my repo below step 5 →

Jev launched last week to a lot of hype. We put it to the test and made our browser-use agents 3.5× faster over the weekend The experiment reinforced a broader view: fast, reliable agents will combine frontier reasoning models with specialized models 🧵

MY COPYTRADING BOT THAT MADE ME OVER $2,130 IN 6 DAYS HAS AN OPEN REPO NOW MORE THAN A MONTH OF WORK AND TESTING, AND THE RESULT IS FINALLY HERE. it's lintcha. one call to Jev per wallet move, and the answer comes back typed. everything in the repo is what happens after that answer: > follow a wallet and get told, with nothing prepared for you > copy its buys inside the caps and the expiry you set > reject anything outside those limits before it is signed there is no auto-sell path in the code. every exit is two actions you confirm yourself. zero runtime dependencies at the root, and the money

I find it excruciatingly hard to find people I want to on LinkedIn... so I built a Jev powered personal CRM from my LinkedIn contacts. I can literally ask it to find anyone in natural language and it'll find me all results in around 1s. Examples "Find me someone who runs user acquisition for a mobile game and knows the ad networks well" "creative production employees at mobile games" "seed investors in b2b software in seattle" "people in sales at companies with under 50 people in new york" and bonus.. each result has contact details! (powered by @treg_ai )

1,092 BROWSER CALLS. THEN JEV CUT IT TO 101. Same task. Same browser. 91% fewer calls. → the model doesn't see the whole page anymore — just a short, fresh action table → pick operation + target, that's the only decision it makes per step → the small LLM only wakes up when a field actually needs typing 25% lower median time. Not from a faster model. From asking it fewer questions. Zürich → London, 7.1 seconds, real time, no speed-up. Official verified run. → Full Jev deep dive ↓

Nothing is impossible with Jev now, I just proved it myself. I built a Jev X Feed Analyser, it reads my timeline while I scroll and judges every post on the screen Here is how it actually works: > every post gets the same 12 typed questions > not a summary, not a vibe check, a schema > hook_archetype, topic, reach_band, is_shill, reads_like_ai, claim_needs_source, regret_if_quoted, worth_reading, worth_reply, bookmark_worthy, opens_with_number, mute_author > each answer is a bool, a score or a choice from a fixed list > each answer carries its own confidence, 0 to 1 > 203ms average per answer,

Built Yap over the weekend: talk to Chrome, it just does it. Runs on TypeSafe's Jev. A nightly Claude routine reads telemetry and tunes the prompts and tests. This video? Opus 5.5, one prompt + a few tweaks. Self-learning apps are now hands-off.

Introducing ballot-jev-0.5b: typed decisions on your own hardware — 0.23s each on a CPU, no GPU, no API bill. ▎ It answers with a calibrated distribution, so you can route on confidence instead of trusting a label. #jev #kev #laya #typesafe