What is being built with AI, how it is put to work, and what happens when it goes wrong.
Written in Spanish. A headless Chromium agent with Jev choosing the links: asked to get from Wikipedia's page for coffee to the one for artificial intelligence by following hyperlinks, it did it in 20 seconds, which the author reckons few people could match.
Mormon Negro
Acabo de terminar la implementación de @typesafeai + Chromium Headless para que mis agentes puedan navegar por internet a una buena velocidad!
En este ejemplo le pido al agente que entre a la página del término "Café" en Wikipedia y navegue por los hipervínculos hasta terminar en la página "Inteligencia artificial"
Logra completar la prueba en 20 segundos. Creo que pocos humanos podrían hacerlo en ese tiempo.
Vercel's own note on the same day, with the call as it reads in the AI SDK: a model, a state and a set of questions.
Vercel Developers
Jev from @typesafeai is on AI Gateway.
Build agents that decide, route, score, and stop in milliseconds:
𝚊𝚠𝚊𝚒𝚝 𝚎𝚟𝚊𝚕𝚞𝚊𝚝𝚎({
𝚖𝚘𝚍𝚎𝚕: '𝚝𝚢𝚙𝚎𝚜𝚊𝚏𝚎-𝚊𝚒/𝚓𝚎𝚟',
𝚜𝚝𝚊𝚝𝚎,
𝚚𝚞𝚎𝚜𝚝𝚒𝚘𝚗𝚜,
});
LangChain's engineering post, two days after launch, wiring Jev into an agent harness as middleware in two places: choosing the model per request, and checking a tool call for risk before it runs, the same job Vercel's fx safety reviewer does. Both are LangChain middleware.
Sydney Runkle and Hunter Lovell, LangChain, in langchain.com
A simple lookup doesn't need the same model as a difficult debugging task. Model-routing middleware lets Jev assess the request and choose a model based on criteria you define, so fast and inexpensive for straightforward tasks, more capable for complex ones.
[...]
AutoModeMiddleware uses Jev to check tool calls for risky decisions it may take, and block calls before the tool executes.
Jarrod Watts' bot has Jev decide buy or sell from a price feed and place the order on Kuru's on-chain order book on Monad, once every 300 ms block. The demo and the code are at the link.
Jarrod Watts
I built a trading bot with Jev!
Jev decides if it should "buy" or "sell", given the price feed of an asset pair, and executes real trades.
It uses Monad to place the orders on Kuru's on-chain order book in every 300ms block.
Demo link →
Lahfir runs Claude Code, Codex and opencode side by side and had been routing between them with rules in a system prompt. A Claude Code hook now asks Jev before delegating: mechanical tasks to Haiku, harder ones to Opus subagents, long implementation work out to another harness.
Lahfir
Jev solved local harness/model routing
I use a combination of Claude Code, Codex and Opencode as my local agentic stack and routing to other harnesses was always enforced in the system prompt/rules
With a deterministic hook that Claude Code can decide before delegation, Jev helps to route to the right harness/model based on the task, and it's pretty accurate based on the intensity/intelligence of the task
> Mechanical tasks get routed to Haiku
> Intelligent ones to Opus sub-agents
> Long-running implementation work to external harness
Guillermo Rauch is the CEO of Vercel. fx is Vercel's coding agent, its default mode is auto, and every command it runs is checked first by a safety reviewer. That reviewer runs on GPT Luna; the Vercel team benchmarked Jev in its place and found it up to 18x faster at p95 and more accurate, which Rauch says makes it the likely new default. The first production guardrail anyone named this week.
Guillermo Rauch
We're seeing extraordinary results from @typesafeai. Default mode in 𝚏𝚡 is auto, with a safety reviewer analyzing every command.
That reviewer runs on GPT Luna today. Jev is up to 18x faster (p95) *and* more accurate. It's coming to @vercel AI Gateway and likely new default.
Pranit, on the Vercel team, with the benchmark behind Guillermo Rauch's post: the fx auto-mode safety classifier on Jev against gpt-5.6-luna, five to eighteen times faster and more accurate.
Pranit
We benchmarked fx auto mode (safety) classifier with @typesafeai's Jev.
tl;dr: ~5-18x faster and more accurate than 𝚐𝚙𝚝-𝟻.𝟼-𝚕𝚞𝚗𝚊, our current top choice
Malte Ubl is Vercel's CTO. An existing classifier eval that had been running on Gemini 2.5 Flash Lite: Jev saturated it and answered six times faster.
Malte Ubl
Ran @typesafeai's Jev against an existing classifier eval that previously used Gemini 2.5 Flash Lite. It won both on quality (saturated the eval) and speed (6x)
Paarangat's is the shortest version: a language model reasons its way to a sentence, Jev fills in decisions you defined in advance, each with a probability.
Paarangat
this is the easiest way to understand Jev:
LLMs generate answers.
Jev makes decisions.
that sounds like a small difference, but it actually changes the entire use case.
say you give a normal LLM this:
“here’s a user, their account history, payment behavior, support chats, device data, etc.
tell me if this looks risky.”
the LLM might reason through it and return:
“yes, this looks high risk.”
maybe in JSON if you ask nicely.
with Jev, you define the possible decisions upfront:
risk:
* low
* medium
* high
manual review:
* yes
* no
and Jev returns something closer to:
risk = high (96%)
manual review = yes (91%)
that’s basically the product.
it’s not trying to be another ChatGPT.
it’s more like an AI-native if statement.
instead of:
if transaction > $10,000:
review()
you can start thinking more like:
if “does this behavior look suspicious?” > 95%:
review()
and that opens up a pretty interesting category of software.
a few assumptions I had at first that turned out to be wrong:
1. “so it’s just a classifier?”
kind of, but that undersells it.
the input can be messy real-world context, and you can ask multiple typed questions about that state at once.
fraud?
churn?
escalate?
eligible?
priority?
all from the same input.
2. “so it replaces GPT / Claude?”
not really.
I actually think the interesting architecture is:
Jev decides WHAT needs to happen
Claude / GPT reason or generate WHEN deeper intelligence is needed
normal code executes the deterministic stuff.
Jev becomes the routing layer.
3. “it can’t hallucinate?”
this one needs nuance.
if your allowed answers are:
LOW
MEDIUM
HIGH
Jev won’t suddenly invent:
“EXTREMELY HIGH 🚨”
the output structure is constrained.
but it can still be wrong.
HIGH at 92% can still be the wrong decision.
so “no hallucinations” doesn’t mean “always correct.”
4. “why not just force an LLM to return JSON?”
you can.
we already do this everywhere.
but you still deal with generation latency, schema validation, retries, weird outputs, confidence estimation and a lot of glue code.
Jev is designed around the decision itself rather than text generation.
5. “why should I care?”
because most software is ultimately a giant tree of:
if this → do that
if this → route here
if this → escalate
if this → reject
if this → ask a human
Jev is basically asking:
what if those if statements could understand messy human context?
that’s a much more interesting framing than “another AI model.”
I can see this being very useful for:
fraud / risk
support routing
moderation
PR / QA automation
lead scoring
compliance
workflow orchestration
agent routing
especially as the cheap + fast decision layer sitting in front of larger reasoning models.
early tech, obviously.
but the category itself makes a lot of sense.
Faadil Shaik, the day after launch, with the first game anyone outside TypeSafe put it on.
Faadil Shaik
got @typesafeai’s new model Jev to play Super Mario Bros.
fast inference + structured outputs makes it surprisingly good for real time use cases.
I'm excited to see what can be done with these new models!
Chubby, whose account tracks model releases, with the reaction most people had first.
Chubby
Sounds almost too good to be true.
Jev is an AI model built for decisions rather than text generation, introduced by Diogo Almeida and trained using a new method called RLCD.
Claimed to be 20–200× faster and 40–400× cheaper than comparable LLMs.
Costs $0.042 per million input tokens, with free outputs.
The trade-off: it cannot generate text. Demos include controlling a game and choosing Wikipedia links.
The idea is to embed fast, cheap AI decisions into software.
Harsha Gundala's answer to the two years in stealth, posted the morning after: a Qwen 2.5 1B set up to batch every key of a JSON schema in one pass and hand back probabilities over the categories, running on an M4 MacBook, on Hugging Face. His point is that any LLM can be made to do this without new training.
Harsha Gundala
They were building in stealth for 2 years, I was building in stealth for 2 hours…
Happy to open source Qwen-2.5-1B-RLCD, 5x faster on-device inference for JSON workloads that need to be type-safe.
⚡️Demo below on a M4 MacBook⚡️
every LLM has the ability to efficiently batch inference every key of a JSON at the same time and generate probabilities from a set of possible categories. No new training required, but it’s easy to optimize if you need!
On hugging face now!
Nathan Flurry is a co-founder of Rivet. This is the explanation most people ended up passing around, and it is worth reading whole: what Jev cannot do, what it can, and the shape of workflow it fits.
Nathan Flurry
hype-free explanation of jev:
jev does not replace gpt / claude
jev is just a *really* smart switch statement
like if 2016 ml classifiers got 2026 levels of intelligence
it's a new* type of tool that will make a lot of workloads insanely fast, cheap, and accurate
* = and by new, i mean rebranded
~~~
it needs a predefined set of options and it will tell you which one to take
it cannot:
- write code
- generate natural language
- reason step by step / show its work
- produce any output you didn't define in advance
- pick from more than ~255 options in one shot
but it can:
- classify, route, score, rank
- give confidence
- pick the right branch, tool, model, or sub-agent
- judge / verify / guardrail an llm's output
- label tons and tons of rows
~~~
i'd imagine a lot of workflows that look like:
llm proposes options → jev decides → code executes
and i see this fitting *really* well with code mode and mcp
~~~
implying this will lead to agi seems incredibly far fetched to me, but i don't want to discount the types of applications that this will make possible
