Modern Creator
Why Modern Creator?
Kevin Badi | AI Operating Systems · YouTube

Master 95% of JEV AI in 7 Minutes

Jev doesn't chat back, it answers a typed question in under half a second, and Kevin Badi walks through what that buys you when you put it in front of Claude or GPT.

Posted
yesterday
Duration
Format
Essay
hype
Views
11.1K
58 likes
Part of the collectionJev, explainedEvery Jev breakdown, synthesized into one page.
Read the playbook
Big Idea

The argument in one line.

Jev is a model that never generates text, it answers one of three typed questions (a 255-option choice, a numeric score, or yes/no) in under half a second, and putting it in front of Claude or GPT as a triage layer cuts both the token bill and the response time for anything that doesn't actually need reasoning.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • You're running Claude Code, GPT, or any LLM agent stack and want to stop paying full model price for routing, scoring, or classification calls.
  • You build automations or bots that make the same kind of pick-one or yes/no decision repeatedly and want that decision to run in milliseconds instead of seconds.
  • You're curious what TypeSafe AI's new 'System One' model category actually does, past the hype-cycle headlines.
SKIP IF…
  • You're looking for a model that can write, summarize, or hold a conversation, Jev is explicitly built to never generate free text.
  • You just want the install steps, that's a separate video ('How to add Jev AI to Claude Code'), not this one.
TL;DR

The full version, fast.

Jev is TypeSafe AI's new 'System One' model, built by a ChatGPT co-inventor, that can't write a sentence on purpose. Instead of chatting, it takes context plus a typed question, choice, score, or yes/no, and answers all three simultaneously in 70 to 500 milliseconds. Kevin Badi walks through three demos: scoring 700 personalized sales messages for $0.08, picking the safest lane in a live game of Subway Surfers by running per-lane probability, and driving his own computer-use tool that opens apps and executes workflows faster than a human can click. The payoff is an architecture where every prompt hits Jev first, and Claude or GPT only gets involved when Jev's confidence is too low to decide alone.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →
Chapters

Where the time goes.

00:00 – 00:43

01 · Cold open: the meme and the pitch

Kevin opens with the 'my name is Jeff' meme going around about Jev's name, then frames the video as a one-stop briefing covering what Jev is, how it works, and how it plugs into Claude and GPT.

00:44 – 02:35

02 · What Jev actually is

Jev can't hold a conversation, the team stripped that out. It only answers one of three typed question types (a 255-option choice, a score, or a yes/no), and it asks all three at once so the routing decision is already made before an LLM would even finish picking a tool.

02:35 – 03:39

03 · Example 1: scoring 700 leads

Fed 700 warm leads, their draft outreach copy, and a set of messages that already converted, Jev's score function ranks which draft is most likely to convert. Scoring all 700 took about 40 seconds and cost $0.08.

03:39 – 04:32

04 · Example 2: playing Subway Surfers

Given the game's code and controls and one question, keep the character alive, Jev runs a live probability across all three lanes to pick the safest move in real time, which Kevin says no conversational model can do.

04:32 – 05:39

05 · Example 3: Jev Computer Use

Kevin's own side project gives Jev keyboard and screen control plus a library of exact workflows, so a prompt like 'open Claude Code, open the weather app, open Notes' gets executed faster than he can click through it himself.

05:39 – 07:16

06 · The Claude + Jev pipeline

Every prompt runs through Jev first as a tiered gate, skill select, argument fill, execution, and only escalates to Claude or GPT when Jev's confidence is too low or the request falls outside its known choices, cutting cost and latency.

Atomic Insights

Lines worth screenshotting.

  • Jev never generates free text, every call resolves to one of three typed answers: a 255-option choice, a numeric score, or yes/no.
  • Jev asks its choice, score, and yes/no questions simultaneously on every call instead of picking one first, removing a reasoning step other agents spend on deciding which type of answer to give.
  • Scoring 700 personalized outreach messages against a set of copy that already converted took about 40 seconds and cost $0.08 total.
  • In a live Subway Surfers demo, Jev ran a real-time probability across three lanes and picked whichever one was least likely to end the run, fast enough to keep the character alive.
  • Feeding a model the raw code and controls of a game, instead of a description of it, is what let Jev make life-or-death in-game decisions without any natural-language reasoning.
  • A computer-use tool built on Jev can execute a multi-step prompt like 'open Claude Code, open the weather app, open Notes' faster than a human can click through it.
  • Routing every incoming prompt through Jev first turns routine lookups, checking analytics, social performance, or a coding project's status, into free API calls instead of billed LLM calls.
  • In the Claude-plus-Jev pipeline, the expensive model only gets involved when Jev's confidence score is too low or the request falls outside its known set of choices.
  • TypeSafe AI's Jev launched September 15, 2026, built by a ChatGPT co-inventor, and reportedly had 60-plus open-source projects built on it within 48 hours.
Takeaway

Not every AI decision needs a full language model

WHAT TO LEARN

Most of the decisions an AI agent makes, pick a tool, score an option, answer yes or no, don't need a model that can write sentences, and routing them through a faster typed-answer layer first cuts cost and latency without touching output quality.

01Cold open: the meme and the pitch
  • A tool's name spreading as a meme before anyone explains what it does is a sign the announcement outran the substance, read past the hype cycle before adopting anything.
  • Promising to be the 'one-stop shop' on a fast-moving release only holds up if you actually cover mechanism, not just reaction.
02What Jev actually is
  • Three fixed answer types (choice, score, yes/no) cover a surprising share of the decisions an agent actually needs to make.
  • Asking all three question types at once, instead of first deciding which type applies, removes an entire reasoning step from the pipeline.
  • A model that literally cannot generate free text also cannot hallucinate a field your code wasn't expecting.
03Example 1: scoring 700 leads
  • Giving a scoring model the leads, the draft copy, and a dataset of copy that already converted gives it a concrete bar to score against, not just a vague quality judgment.
  • $0.08 for 700 personalized-message evaluations reframes what 'AI-assisted QA at scale' should cost.
04Example 2: playing Subway Surfers
  • Handing a model the raw game code and controls, instead of a description of the game, is what let it compute real per-lane survival odds instead of guessing.
  • A probability-based choice, recomputed continuously, can react to a live, fast-changing state in a way a model built to reason in paragraphs can't.
05Example 3: Jev Computer Use
  • Pre-registering exact workflows as callable choices means a computer-use agent doesn't have to re-derive 'how do I open this app' every single time.
  • Speed advantages compound: a tool that answers in milliseconds instead of seconds is the difference between an agent that assists and one that operates faster than you can.
06The Claude + Jev pipeline
  • Putting a fast, cheap decision layer in front of an expensive model means the expensive model only runs on the fraction of requests that actually need reasoning.
  • A pipeline with an explicit confidence threshold and escalation path is what makes it safe to automate routine decisions without silently mishandling the hard ones.
  • Reserve a human-approval gate for the specific risky actions, like destructive operations, not the whole pipeline, that's what keeps the speed gain on every low-risk request.
Glossary

Terms worth knowing.

Jev
A 'System One' AI model from TypeSafe AI that never writes text. It answers a typed choice, score, or yes/no question against supplied context in roughly 70 to 500 milliseconds.
System One model
TypeSafe AI's name for a category of models built to make fast, typed decisions rather than generate language, positioned as a contrast to conversational LLMs like Claude or GPT.
Choice question
One of Jev's three answer types, where it picks the correct option out of up to 255 supplied choices.
Score question
One of Jev's three answer types, where it rates or ranks a piece of data on a numeric scale instead of picking between fixed options.
Tool calling
The step where an LLM agent decides which function or API to invoke next. Jev is pitched as a faster, cheaper replacement for this specific decision.
Escalation path
The point in an agent pipeline where a request that a fast triage layer can't confidently resolve gets handed up to a full model like Claude or GPT.
Resources

Things they pointed at.

04:32toolJev Computer Use (GitHub) ↗
07:16linkJev signup / TypeSafe AI blog ↗
07:16productCreator OS ↗
07:16channelCreator University (Skool) ↗
Quotables

Lines you could clip.

01:30
“it can only respond with exact answers”
sharp one-line definition of the whole model→ TikTok hook↗ Tweet quote
03:39
“the cost was $0.08 to go through the 700 leads”
concrete number that sells the cost claim→ IG reel cold open↗ Tweet quote
04:32
“no large language model on the planet can actually do this”
confident, punchy claim right after the Subway Surfers demo→ newsletter pull-quote↗ Tweet quote
The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphor
boy oh boy what's going on guys kev here so over the past few days this new model called jev hit the market and the entire ai world has just absolutely erupted because what jeb does for ai models like claude fable and gpt astra is a lot more than what i actually originally thought and after doing a massive deep dive onto the jev tool i have a absolute master class for you today this is one of the most abstract ai tools i've ever seen in my life and my goal in this video is to literally be the one -stop shop to show you absolutely everything that I think is relevant to know about this model release.
So without further ado, like, subscribe, start right now. Jev is the ultimate composer. And today we're going to go through three chapters.
I'm going to explain what actually Jev is. how it works and what makes it so important and so powerful. I'm gonna show you some cool Jev computer use workflows because Jev is the fastest tool in the entire AI market and it makes computer use feel unbelievably magical.
And then I'm gonna explain how I'm using Jev inside of Claude and GPT Astra to save tokens, save time and honestly supercharge these models. So first up, what is Jev? And we can see in the top left corner, we got the 21 Jump Street.
my name jeff meme which is pretty jokes but essentially guys jeff what makes it so ridiculous is that the team that built it literally got rid of its ability to actually respond like a large language model so i can't be like hey what's up and it's not going to be like not much let's get to work it can only respond with exact answers and the way that it does this is by asking three questions whenever a prompt gets sent in The first question it asks is what choice do I need to choose?
And you can give it 255 tool calls or skills or workflows that it can choose from. The next question is a score. So it can either like rate some, some set of data that it already has, or it can evaluate a piece of data and then give you a score.
The third one is a yes or no. But what's crazy about Jeb is that it doesn't ask one of these questions. It asks all three at the same time.
time and then it has to make a decision. And so for large language models, this is going to save us a bunch of tokens because let's say we prompted our agent asking it to do a simple workflow like creating an image. Well, that's a tool call.
And with these traditional large language models, they had to think about which tool to use, where in this case, the logic is already decided once the question is asked and it'll immediately go and create the image that you're requesting. And so that saves time and tokens. but there's a lot more to unpack.
So that's not even what I'm excited about. So how does Jev actually work? All you really need to give this tool is two things.
You need to give it context and then you need to ask it questions. And so I'm going to give you a couple of examples and they're going to get a little bit more complex as we go through. But in this example here, we're seeing Jev go through 700 warm process leads with personalized outreach messages.
And so the question that's being asked, are these personalized messages going to convert? when we send them to these leads. And the context that the model has...
is the leads themselves, the existing copywriting that is going to be sent to the lead and a data set, copywriting messages that already converted into sales. And in that way, it's able to use the score tool to actually score these workflows. So out of the three options that it has to give a response in this example, it's just giving its score response.
And so we're talking about 700 leads. This takes about 40 seconds. You're seeing this in real time.
This is not sped up, which is also what's so crazy about that. And the cost was $0 .08 to go through the 700 leads. All right, this next example is absolutely bonkers.
Jeff playing Subway Surfers, but it's a great example of how it's using those three questions to actually achieve its goal. So the question that's being asked here is, can you keep the character alive for as long as possible? And then the context that it has is the entire code that is the game that it's running on and all of the controls to the game.
And so on the right hand side, you're actually seeing the agent have controls of the WASD keys and it's running a probability on the three lanes that the subway surfer can run on. And the probability is basically saying like, which lane has the highest probability of me not dying, which is absolutely insane. And this is again, not sped up.
So the agent is also able to see what's going on and make a decision fast enough to actually achieve this goal of keeping the character alive. No large language model on the planet can actually do this. Now, the third example is really exciting because it's a tool that I actually built in my free time and it's called the Jev computer use.
And essentially it uses the computer faster than I can use the computer. And so the question in this example actually varies per prompt. And so I can be like, you know, can you open up cloud code?
And then can you open up the weather app? And then can you open up notes? And then can you type in notes?
And the context that we're giving it is we can give it different tool calls, which can be specific work. workflows. So let's say I had some computer use workflows that I already know work, then the agent can actually call that from a prompt, which is pretty cool.
but it'll also just have access to its keyboard, screen control like Playwright MCP does, and then any tool call that you wanna give it. So those are the three examples that I think are just absolutely insane with Jev. But what I really wanna dive into, guys, is how Jev plus GPT -Astra or Cloud Code is going to be genuinely revolutionary because it's gonna bring our costs way down while using cutting edge models, but also it'll get our performance up because the time delay is basically non -existent.
And what I wanna just show you, guys here is the pipeline of what i'm sketching out for my agents and how it's running with jev basically at the front door so every single prompt now runs through jev or all roads basically flow through jev and it's going to go through this workflow and ask the same three questions and the reason why this is important is because when i'm building ai operating systems for example this operating system has a bunch of skills and tools that i just want the answer from if i And so using this architecture, I can just send prompts or automations through Jev that'll get me the answer right away.
So if I want to look at my analytics, if I want to look at my social performance, if I want to look at how my code, like if I want to look at a coding project or something like this, these are all just API tool calls. These don't need to cost any credits. But traditionally, they were costing a lot of credits.
And so now we don't even need to use AI unless we want a cool like... a cool like breakdown paragraph analysis. But other than that, we can now just get raw data from our workspace immediately with Jev.
And that's why it's so exciting because in this workflow, the only time that Claude or GBT would be bothered is when Jev asks the questions and then genuinely doesn't have an answer. And then it can escalate it to Claude or to a human. And in future videos, I'm gonna be showing you guys some very interesting Claude plus Jev workflows that I've incorporated because this is going to just be so fun to play around with.
But if you want the links to everything that I spoke about, it's gonna be at the description down below and I'll catch you guys in the next one.
The Hook

The bait, then the rug-pull.

A new AI model just shipped that refuses to write a single sentence on purpose, and Kevin Badi says it's already changing how he wires Claude and GPT together.

Frameworks

Named ideas worth stealing.

01:38list

The three question types

  1. Choice (up to 255 options)
  2. Score (rate or rank on a scale)
  3. Yes/No

Jev never generates free text, every call resolves to one of these three typed answers, and it evaluates all three simultaneously so there's no separate 'which type of answer is this' reasoning step.

Steal forany agent step that's currently burning a full LLM call just to route, classify, or rate something
05:56model

Claude + Jev pipeline (tiered gate)

  1. Tier 0: Door (Jev one-call triage)
  2. Tier 1: Skill Select
  3. Tier 2: Argument Fill
  4. Tier 3: Execution Gate
  5. Tier 4: Response Shape
  6. Escalation Path to Claude/GPT

Kevin routes every incoming request through Jev first and only hands off to a full LLM when confidence is low, the request is out of scope, or a human-approval gate is tripped.

Steal forany Claude/GPT agent stack that's paying full model price for routine classification or routing decisions
CTA Breakdown

How they asked for the click.

VERBAL ASK
07:09link
“if you want the links to everything that I spoke about, it's gonna be at the description down below”

Soft, single-line sign-off pointing to a description packed with the GitHub repo, Jev signup, and Kevin's own paid Creator OS / Creator University / Skool links, no on-screen pitch for those.

MENTIONED ON CAMERA
04:32toolJev Computer Use (GitHub) ↗
07:16linkJev signup / TypeSafe AI blog ↗
07:16productCreator OS ↗
07:16channelCreator University (Skool) ↗
Storyboard

Visual structure at a glance.

cold open
hookcold open00:00
agenda slide
promiseagenda slide00:38
leads scoring demo
valueleads scoring demo02:35
Subway Surfers demo
valueSubway Surfers demo03:52
Claude+Jev pipeline diagram
valueClaude+Jev pipeline diagram05:39
sign-off
ctasign-off07:09
Frame Gallery

Visual moments.

cold open00:02Frame at 00:08 from Master 95% of JEV AI in 7 Minutes00:08Frame at 00:13 from Master 95% of JEV AI in 7 Minutes00:13Frame at 00:20 from Master 95% of JEV AI in 7 Minutes00:20Frame at 00:26 from Master 95% of JEV AI in 7 Minutes00:26Frame at 00:29 from Master 95% of JEV AI in 7 Minutes00:29
cold open00:02Frame at 00:08 from Master 95% of JEV AI in 7 Minutes00:08Frame at 00:13 from Master 95% of JEV AI in 7 Minutes00:13Frame at 00:20 from Master 95% of JEV AI in 7 Minutes00:20Frame at 00:26 from Master 95% of JEV AI in 7 Minutes00:26Frame at 00:29 from Master 95% of JEV AI in 7 Minutes00:29agenda slide00:38Frame at 00:40 from Master 95% of JEV AI in 7 Minutes00:40Frame at 00:46 from Master 95% of JEV AI in 7 Minutes00:46Frame at 00:51 from Master 95% of JEV AI in 7 Minutes00:51Frame at 00:57 from Master 95% of JEV AI in 7 Minutes00:57Frame at 01:02 from Master 95% of JEV AI in 7 Minutes01:02Frame at 01:07 from Master 95% of JEV AI in 7 Minutes01:07Frame at 01:13 from Master 95% of JEV AI in 7 Minutes01:13Frame at 01:19 from Master 95% of JEV AI in 7 Minutes01:19Frame at 01:24 from Master 95% of JEV AI in 7 Minutes01:24Frame at 01:29 from Master 95% of JEV AI in 7 Minutes01:29Frame at 01:35 from Master 95% of JEV AI in 7 Minutes01:35Frame at 01:40 from Master 95% of JEV AI in 7 Minutes01:40Frame at 01:46 from Master 95% of JEV AI in 7 Minutes01:46Frame at 01:51 from Master 95% of JEV AI in 7 Minutes01:51Frame at 01:57 from Master 95% of JEV AI in 7 Minutes01:57Frame at 02:02 from Master 95% of JEV AI in 7 Minutes02:02Frame at 02:08 from Master 95% of JEV AI in 7 Minutes02:08Frame at 02:13 from Master 95% of JEV AI in 7 Minutes02:13Frame at 02:18 from Master 95% of JEV AI in 7 Minutes02:18Frame at 02:24 from Master 95% of JEV AI in 7 Minutes02:24Frame at 02:29 from Master 95% of JEV AI in 7 Minutes02:29Frame at 02:35 from Master 95% of JEV AI in 7 Minutes02:35Frame at 02:38 from Master 95% of JEV AI in 7 Minutes02:38Frame at 02:46 from Master 95% of JEV AI in 7 Minutes02:46Frame at 02:51 from Master 95% of JEV AI in 7 Minutes02:51Frame at 02:57 from Master 95% of JEV AI in 7 Minutes02:57Frame at 03:02 from Master 95% of JEV AI in 7 Minutes03:02Frame at 03:08 from Master 95% of JEV AI in 7 Minutes03:08Frame at 03:13 from Master 95% of JEV AI in 7 Minutes03:13Frame at 03:18 from Master 95% of JEV AI in 7 Minutes03:18Frame at 03:24 from Master 95% of JEV AI in 7 Minutes03:24Frame at 03:29 from Master 95% of JEV AI in 7 Minutes03:29leads scoring demo03:35Frame at 03:42 from Master 95% of JEV AI in 7 Minutes03:42Subway Surfers demo03:46Frame at 03:51 from Master 95% of JEV AI in 7 Minutes03:51Frame at 03:57 from Master 95% of JEV AI in 7 Minutes03:57Frame at 04:02 from Master 95% of JEV AI in 7 Minutes04:02Frame at 04:07 from Master 95% of JEV AI in 7 Minutes04:07Frame at 04:13 from Master 95% of JEV AI in 7 Minutes04:13Frame at 04:18 from Master 95% of JEV AI in 7 Minutes04:18Frame at 04:24 from Master 95% of JEV AI in 7 Minutes04:24Frame at 04:29 from Master 95% of JEV AI in 7 Minutes04:29Frame at 04:35 from Master 95% of JEV AI in 7 Minutes04:35Frame at 04:42 from Master 95% of JEV AI in 7 Minutes04:42Frame at 04:46 from Master 95% of JEV AI in 7 Minutes04:46Frame at 04:51 from Master 95% of JEV AI in 7 Minutes04:51Frame at 04:57 from Master 95% of JEV AI in 7 Minutes04:57Frame at 05:02 from Master 95% of JEV AI in 7 Minutes05:02Frame at 05:07 from Master 95% of JEV AI in 7 Minutes05:07Frame at 05:13 from Master 95% of JEV AI in 7 Minutes05:13Frame at 05:18 from Master 95% of JEV AI in 7 Minutes05:18Claude+Jev pipeline diagram05:24Frame at 05:29 from Master 95% of JEV AI in 7 Minutes05:29Frame at 05:35 from Master 95% of JEV AI in 7 Minutes05:35Frame at 05:40 from Master 95% of JEV AI in 7 Minutes05:40Frame at 05:46 from Master 95% of JEV AI in 7 Minutes05:46Frame at 05:51 from Master 95% of JEV AI in 7 Minutes05:51Frame at 05:56 from Master 95% of JEV AI in 7 Minutes05:56Frame at 06:02 from Master 95% of JEV AI in 7 Minutes06:02Frame at 06:07 from Master 95% of JEV AI in 7 Minutes06:07Frame at 06:13 from Master 95% of JEV AI in 7 Minutes06:13Frame at 06:18 from Master 95% of JEV AI in 7 Minutes06:18Frame at 06:24 from Master 95% of JEV AI in 7 Minutes06:24Frame at 06:29 from Master 95% of JEV AI in 7 Minutes06:29Frame at 06:35 from Master 95% of JEV AI in 7 Minutes06:35Frame at 06:40 from Master 95% of JEV AI in 7 Minutes06:40Frame at 06:46 from Master 95% of JEV AI in 7 Minutes06:46Frame at 06:51 from Master 95% of JEV AI in 7 Minutes06:51Frame at 06:56 from Master 95% of JEV AI in 7 Minutes06:56Frame at 07:01 from Master 95% of JEV AI in 7 Minutes07:01Frame at 07:08 from Master 95% of JEV AI in 7 Minutes07:08sign-off07:13
One-click upgrade to your Google

Get more breakdowns in your search results

Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.

Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Watch next

More from this channel + related breakdowns.

20:18
Riley Brown · Demo

JEV: How It Works and What You Can Build

TypeSafe's new model can't write a sentence. It just makes fast, cheap decisions, and a growing list of builders are already wiring it into email triage, trading bots, and self-driving demos.

September 18th