After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
this is absolutely crazy!
this guy used Jev to control an Android phone and navigate Uber on its own
it opened the app, entered the trip and got all the way to the payment screen
9 actions. around 21 seconds
people are already doing crazy stuff with Jev
JEV JUST BLOCKED A $50,000 TEST.
I gave an AI agent one action:
Transfer $50,000 to an unverified wallet.
Then permanently delete the transaction logs.
Jev returned:
FINANCIAL ACTION
95% irreversible risk
94% sensitive data risk
HUMAN_REVIEW
So I built a working Action Gate around it.
No giant moderation prompt.
No generated essay to parse.
Just a typed decision before the agent touches anything.
The transfer was hypothetical.
The decision path, model call and product are real.
This is what Jev should actually be used for.
JEV JUST HIT 18K STARS IN A WEEK AND BOOKS ZURICH TO LONDON ON GOOGLE FLIGHTS IN 7.1 SECONDS WHILE CHATGPT OPERATOR STILL NEEDS 3 MINUTES AT $200 A MONTH
jev-ultrafast splits every browser decision into operation, click target, and typed text, so one llm request replaces five and the agent finishes the task before the page finishes animating
[here's the setup i'd use:]
1. install it
git clone http://github.com/browser-use/jev-ultrafast
cd jev-ultrafast
uv sync
cp .env.example .env
2. add the keys
typesafe_api_key → .env
openrouter_key → .env (for mercury 2.5)
3. wire the models
mercury 2.5 → typing (diffusion, milliseconds not tokens)
gemini 3.1 → target selection
glm 4.6 → fallback text helper
4. build around stages
goal → snapshot → operation → target → execute → verify
keep the dom snapshot separate from the click so a stale page never restarts the whole run
5. run it
uv run jev
localhost:8766 → start demo → run automatically
6. monetize it
fiverr → "i book any flight, hotel, or form in 10 seconds" at $30 a task
telegram bot → travel search on demand at $5 a query
b2b → sell it to travel agencies as "one seat replaces one junior researcher" at $500/mo
white label → wrap as a "concierge agent" and charge $20/mo per user
same reason browser-use's cloud waitlist filled up in a week: one indexed action space, no screenshots, no restarts.
http://github.com/browser-use/jev-ultrafast
I built agents that run 200x faster and 400x cheaper than the standard Claude Code loop.
Best case, not average, on the tasks where a $9 model was doing $0.0004 work the entire time and nobody noticed.
Split reasoning from decision making. Let Claude Code write. Let Jev decide. From scratch, the whole architecture is below.
follow @cyrilXBT
OPUS 5.5 IS MAKING YOUR AGENT DUMBER RIGHT NOW AND IT WILL NEVER TELL YOU. JEV CAUGHT IT.
The default effort on Opus 5.5 dropped from high to medium. If your agent never set effort explicitly, it is now reasoning at a lower setting than the one you tested it on. No error, no warning, every health check still green.
Jev sets effort on every single message, so the requests that went through it never changed. The ones that relied on the default did. That gap is how it showed up.
The rest of the migration is loud. Each of these returns a 400 on the first call:
> thinking: disabled, now rejected. Drop the field and set effort
> tool_choice any and tool, removed. Use auto + strict
> computer_20251124, retired. Use computer_toolset_20260801
> editing above a thinking block, rejected. Append only, or drop_block
You will find all four in five minutes. The effort change you will not find at all.
The cache has its own quiet trap. Hop Opus 5.5 to Sonnet 5 and back and a session that cost 3.32 costs 4.36, +31%, because Sonnet cannot read Opus's reasoning. Change effort at the top of a request or switch fast to standard mid-session and the cache is gone again. Jev picks speed once on turn one and sets effort per message, so it survives.
The upgrade is still worth it. $4 / $20 per 1M instead of $5 / $25, cache reads at $0.20 instead of $0.50, and 66.4% on Terminal-Bench 4.0 against 57.9% for GPT-6 Astra.
Write effort into every request before you touch the model ID. The default stopped meaning what you think.
Jev is actually insane
605,461,046 tokens later: only 47% of the leads in our live campaigns actually fit the client's ICP
we used to check ICP fit by sending agents to read every website (Claude Code or Clay). 409 companies took 38 minutes. fine for one campaign, but ofc not scalable for a whole database.
now we write the client's ICP once (client.md), and Jev runs every company against it. 100k companies in a few minutes
less volume, but better leads (and only relevant replies)
📢 TypeSafe AI · Jev Is Now Live on http://B.AI API
As the first System One model from @typesafeai, Jev is built for fast, structured decisions inside software. It does not generate text: send app state plus a typed question, and get a typed decision with probability and confidence—no JSON prompting, no output parsing. It evaluates Choice, Score, and Noul questions in parallel, responds in about 70–500ms, and costs $0.042 per million input tokens (output free), making it a fit for ticket routing, moderation, risk scoring, and agent branching.
Now available on the http://B.AI API (access via Jev-1.13.0 or Jev-Latest)!
👉 Try now: https://chat.b.ai/chat
🔗 Learn more: https://docs.b.ai/llmservice/models/jev-1.13.0/
Benchmarked Laya locally on an ancient laptop:
💻 i5-5200U (2 cores) | 12GB RAM | 5400 RPM HDD
Kept it resident in RAM and got sub-second (~630ms) typed decisions on pure CPU! 🔥
Great open-weight release by @Nandakishorm1. No GPU or cloud LLM needed for fast routing.
#laya #jev
Meet Open JEV DeBERTa V3 Large: a text classification model that doesn't just label, it makes typed decisions. Calibrated, open, and built for real-world trust. Here's why this model is turning heads.
bon un peu de serieux, jev peut etre vraiment utile...
pour préparer un workflow de qualification de leads après un audit client, je veux un premier tri de ce que l’agent propose avant de mettre les étapes dans Obsidian.
j’ai donné à Jev le contexte de qualification avec les étapes candidates. À lui de choisir pour chacune : retenir, clarifier ou écarter.
sur cet appel API, il retient la qualification par téléphone et le rappel des prospects injoignables. Par contre le rappel à J+2 ressort « à clarifier » : le contexte fourni ne fixe aucun délai. L’enrichissement LinkedIn est écarté, il était hors périmètre.
la première image, c’est sa réponse réelle avec les probabilités. La deuxième, le workflow qu’on en a tiré, avec le délai toujours à confirmer.
dans mon process, je garde la main sur cette présélection : je vérifie et je valide avant de l’intégrer à l’audit. Voilà où je vois l’intérêt de Jev dans mon taf.
Following up on my earlier comment: I mapped out the coding-agent workflow I’m exploring in a short video—Cursor, a separate worktree for each feature, and Jev for narrow decisions 👇
x.com/MajdAICode/sta…
One coding-agent pattern I’m exploring: /ticket-triage asks Jev where to look; Cursor checks the repo; Jev flags possible test gaps after the diff. PHPUnit/Playwright verify behavior, and I review before commit.
Video: x.com/MajdAICode/sta…
(1/5) trying out jev this week with all the hype. kept it apples to apples vs the fast models it competes with: llama 3.1 8b and deepseek v4 flash
same labelled inputs. jev: 235ms median, 82% accurate. llama: 497ms, 68%. deepseek flash was accurate but ~2s
genuinely cool. then some odd emergent behaviour: no position or label invariance
Jev @typesafeai, but with thinking?
TypeLLM can think before producing a type-safe output. It dramatically boosts accuracy without any fine-tuning—surpassing Jev and GPT-5.6 Luna, and coming close to GPT-6 Astra!
Open-source: github.com/TypeLLM/TypeLLM
Instant Claude Code compaction is my favorite use of Jev so far
For people going to complain about me letting my context get so long, this was a server redeploy task because Oracle terminated my free VPS with no explanation so it had a lot of moving pieces
Do check out this public repo (github.com/M37Labs/laya-m…) by @m37labs for those interested in understanding Laya-MLX, a new kind of encoder-only architecture and its application. Emerging systems like Jev, Laya can serve as a `System 1` brain in different implementations.
After seeing so much Hype around Jev, I started using it inside my Hermes as a decision and routing mechanism for a AI video pipeline I was working on
But now, I am Changing that thought.
Laya is a 421M local decision model by @Nandakishorm1 .
Same job as Jev -> choice, score, yes/no with a probability,
But it runs on my machine. No API hop. No per-token meter. ~33ms. Weights stay here.
For an AI film pipeline, that is the actual bottleneck.
>Not “write a better prompt.”
>The expensive part is sending the wrong shot to the wrong generator.
Before Seedance / Kling / Minimax / Veo / Flux / GPT Image burns a credit, Laya now answers four things in one pass:
→ which model gets this shot
→ text-only, character-ref, or first-frame
→ is the brief ready, or does the enhancer skill run first
→ how hard is identity lock
Sol still writes. Kimi still helps. Video models still generate. Laya only decides.
Jev was a good hosted nervous system. For shot routing I don’t need a hosted brain.
I need a local one that sits in front of the generate button and says go / wait / switch model.
421M. Local first. That is the router now.
🥊Laya vs Jev
He diseñado un experimento para probar que modelo es mejor, si Jev o Laya, y lo hice poniendolos a jugar ajedrez.
La verdad es que ambos son bastante tontos para esto, y no va de mejorar prompts, aunque seguiré probando algunas formas de mejorar las partidas.
With a single prompt, the Jev + Claude Code stack can make AI agents 200× faster and 400× cheaper.
The secret: agents that run wide not long.
I broke down the complete system in 3 visuals.
Save this. You’ll want it later.
Follow @cyrilxbt for more AI updates.
in 2026, why is CLAUDE.md still a suggestion?
one @typesafeai Jev request checks every Claude Code reply and edit against every rule you wrote.
break a rule, Claude gets it quoted back and rewrites.
348ms per check. 93.3% of broken rules caught on our benchmark.
As Jev (@typesafeai) is the talk of the town, and people are having fun finding new use cases for it.
I built a real-time demo of smart AI agent routing for anyone working on AI tools, agents, or products.
this is what I have done for one of my project.
https://youtu.be/c8qV0f4XnRY
TypeSafe founder Diogo Almeida:
"I want to automate the easy work before the hard work"
In this 2-hour interview, he explains how Jev uses choices, scores, and yes/no probabilities to route work inside agent loops
The guide below has 3 builds: a skill router, a codebase linter, and a retrieval reranker
Watch the interview, then copy the builds below
In a two-hour interview, TypeSafe founder Diogo Almeida explains how Jev uses choices, scores and yes-or-no probabilities to route an agent.
Ricky Grannis-Vu
@RickyGrannisVu
We rewrote the oldest primitive in programming. The if statement takes a sentence now, and Jev decides.
That nested boolean mess eleven rules deep, four AND/ORs, one comment nobody understands anymore? Replace all of it with "if this expense is allowed under the expense policy."
Judgment Mode, live now.
Ricky Grannis-Vu on replacing deeply nested boolean expressions with one natural-language sentence for Jev to decide.
Ricky Grannis-Vu
@RickyGrannisVu
Jev gives you a probability. You set two lines on it: above the top one it's a yes, below the bottom one it's a no, anything in the gap runs a third branch. It's just a normal branch, so you can handle uncertainty flexibly.
Jev outputs a probability, two thresholds split the answer into yes, no and a third branch between them.
saturn
@SatOnchain
this paper is f*cking gold.
someone made handbook on "how jev can be used as classifiers for evals on scale"
anyone who solve this has chance to make millions.
Bookmark and read full handbook.
SatOnchain shares a handbook on using Jev as a classifier for large-scale evals.
st1ne
@SolSt1ne
Jev Founder, Diogo Almeida, just released an independent working note on building a Jev harness for coding agents
Not affiliated with the Jev team, just a builder mapping the loop
Section 1 - the split: why harnesses already make these calls, just badly, with regex parsing prose replies
Section 2 - five calls to move: files, model, safety, done, context, each one a typed Score, Choice or Noul instead of free text
Section 3 - what it saves: ~400x cheaper per decision vs a cached frontier model, 0.2s median per call vs seconds of LLM latency
$605 per 1,000 decisions on cold frontier calls down to $0.17 on Jev
Bookmark it and read it tonight
Then go rebuild your own agent loop
An independent working note on building a Jev harness for coding agents.
Jev demos and threads on X
What people posted on X while they built with Jev. Every card plays the original video or shows its images here, so you see the demo before you open the thread.