Why Modern Creator?

Lukas Margerie · YouTube

Testing TypeSafe's Jev: A Pixel Painter, a YouTube Classifier Extension, and Voice-Controlled Subway Surfers

A hands-on tour of TypeSafe's new decision-only model, remixing five community projects from X into a Chrome extension that scores YouTube videos, a tool that turns any website section into a rebuild prompt, and a voice-controlled game.

Posted

September 19th

yesterday

Duration

10:49

Format

Demo

educational

Views

28.8K

241 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 'System One' model built to return fast, cheap structured decisions rather than generate open-ended text, and pointing a coding agent at other builders' open-source Jev projects on X is the fastest way to build something real with it.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You build internal tools or Chrome extensions and want a cheaper, faster alternative to an LLM call for classification, routing, or scoring tasks.
You're curious what a 'decision model' actually is and how it differs from a chatbot-style LLM.
You use Codex, Claude Code, or a similar coding agent and want ideas for remixing other people's open-source demos into your own tools.

SKIP IF…

You need a model that writes or explains things, Jev explicitly can't generate text or reasoning.
You're not willing to sit on a waitlist, access to Jev isn't self-serve yet.

TL;DR

The full version, fast.

TypeSafe released Jev, a 'System One' model that returns structured decisions, choices, scores, probabilities, in 70-500ms instead of generating text, making it far cheaper than a normal LLM call for classification-style tasks. After getting API access from TypeSafe's waitlist and installing an agent skill in Codex, the creator finds five community-built Jev projects on X and has his coding agent remix each one: a pixel painter that predicts colors pixel-by-pixel, an X-post classifier rebuilt for YouTube, a website-section-to-rebuild-prompt Chrome extension, a product-launch-video-to-Diffusion-Studio pipeline, and a voice-controlled version of an existing Jev-powered Subway Surfer bot. The through-line: the fastest way to learn a new model is to copy someone else's open-sourced demo and ask an agent to adapt it.

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 – 01:10

01 · Intro + X examples

Cold open surveying live X demos of Jev in the wild: a tweet-scoring classifier, an ad-account research tool reading hundreds of news stories in seconds, and a competitor ad-analysis tool.

01:10 – 02:02

02 · What Jev actually is

Jev doesn't generate text, it evaluates the whole input at once and returns a decision (yes/no, category, score). Fast (70-500ms/pass) and cheap because there's no text output to pay for.

02:02 – 02:54

03 · TypeSafe waitlist + dashboard

How to join the typesafe.ai waitlist (took the creator 4-5 hours to get access) and a tour of the dashboard: API keys, usage, playground, and walkthrough demos.

02:54 – 04:03

04 · Codex skill + API key

Copies TypeSafe's agent prompt into a new Codex workspace, generates an API key, and asks the agent what it can build. Gets back concrete ideas: smart inbox, creative brief matcher, transcript clip finder, lead prioritizer.

04:03 – 05:25

05 · Pixel painter remix (Anshu)

Finds Anshu's open-sourced 'Jevinci' project on X, a pure-JS canvas where Jev predicts every pixel's color in parallel from a text prompt. Points Codex at the GitHub repo and replicates it live.

05:25 – 06:34

06 · X classifier to YouTube Chrome extension (Peter Wang)

Finds Peter Wang's open-sourced X-post-labeling Chrome extension (three cents per 1,000 posts) and has the agent rebuild it for YouTube instead, producing real-time video labels while scrolling.

06:34 – 07:27

07 · Site section to prompt Chrome extension

Builds a new Chrome extension from scratch: hover over any website section, lock it, and get back a detailed rebuild prompt plus a reference screenshot to hand to Codex or Claude Code.

07:27 – 09:14

08 · Product launch video to Diffusion Studio

Asks the agent to scan a product launch video for its visual components and rebuild them inside Diffusion Studio, ready to re-animate with new copy and colors.

09:14 – 10:48

09 · Voice-controlled Subway Surfers

Riffs on an existing Jev-powered Subway Surfer bot by adding a microphone: saying a place name changes the game's environment live, and phrases like 'make it harder' adjust difficulty.

10:48 – 10:49

10 · Outro

Sign-off.

Atomic Insights

Lines worth screenshotting.

TypeSafe's Jev can't write a single word or explain its reasoning, it only outputs a decision like true/false, a category, or a score.
Jev runs a full decision pass in roughly 70 to 500 milliseconds because it skips the token-by-token text generation a normal LLM call pays for.
SuperX uses Jev to run a draft tweet through 61 questions pulled from over 50 million real tweets before you ever post it.
One ad-research workflow reads 384 news stories in under 25 seconds and matches them to which of 15 brands should pitch which story.
A pixel-painting demo has Jev predict every pixel's color in parallel from a text prompt, painting the canvas with pure JavaScript and zero image-generation models.
A Chrome extension that labels X posts as clean, engagement bait, or promo costs about three cents per thousand posts to run on Jev.
Rebuilding that same X-labeling extension for YouTube took a coding agent about eight minutes, because the classifier logic transfers and only the platform's data fields change.
A second Chrome extension turns hovering over any website section into a detailed rebuild prompt, capturing real typography, spacing, colors, and layout for a coding agent.
Jev can scan a product launch video for its visual components and hand them to Diffusion Studio ready to animate, replacing manual storyboarding.
The fastest way to learn a brand-new AI model isn't the docs, it's finding someone's open-sourced demo on X and asking a coding agent to replicate it.
A voice-controlled Subway Surfer remix works because 'change the environment to Tokyo' and 'make it harder' are both just short-phrase classification decisions, exactly what Jev is built for.
Getting off TypeSafe's waitlist took the creator about four to five hours.

Takeaway

A model built to decide, not write

WHAT TO LEARN

Jev proves that most 'is this true / which category / how urgent' problems don't need an LLM to write anything at all, just to render a fast, cheap judgment.

01Intro + X examples

Jev is a specialized 'System One' model TypeSafe built for fast structured decisions, not a general chatbot.
Real X demos already use it for tasks like scoring a draft tweet against 61 questions from 50 million real tweets, or reading hundreds of news stories in under 25 seconds to match brands to stories.

02What Jev actually is

Jev can't write a single word or explain its reasoning. It only evaluates the input and returns a decision like true/false, a category, or a score.
Each pass runs in roughly 70 to 500 milliseconds and costs far less than a normal LLM call because there's no token-by-token text generation to pay for.

03TypeSafe waitlist + dashboard

Access currently requires joining a waitlist at typesafe.ai; the creator's own approval took four to five hours.
The dashboard looks like a standard developer platform (API keys, usage, playground) but adds walkthrough demos and quick-start docs aimed specifically at decision-model use cases.

04Codex skill + API key

TypeSafe ships a copy-paste agent prompt for Codex, Claude Code, or Grok that teaches the coding agent how to call Jev, plus an API key you generate from the dashboard.
Asking the agent 'what can we do with this' inside an existing project surfaces concrete starter ideas: a smart inbox sorter, a creative brief matcher, a transcript clip finder, a lead prioritizer.

05Pixel painter remix (Anshu)

Community projects on X are a faster on-ramp than docs: pasting someone else's GitHub repo into your agent and asking it to replicate the idea produces a working demo in minutes.
In the pixel-painter example, Jev predicts every pixel's color in parallel from a text prompt and paints it with pure JavaScript, no image-generation model involved.

06X classifier to YouTube Chrome extension (Peter Wang)

An open-sourced Chrome extension that labels X posts (clean, engagement bait, promo) for three cents per thousand posts was rebuilt for YouTube by pointing an agent at the original repo and asking for the same idea on a different platform.
The rebuild took about eight minutes and updates its labels in real time as new videos scroll into view, showing how much of a 'new' tool is really just swapping the target platform.

07Site section to prompt Chrome extension

A second Chrome extension lets you hover over any section of any website, lock it, and get back a detailed reconstruction prompt plus a reference screenshot to hand to a coding agent.
The generated prompt captures actual typography, spacing, colors, and layout closely enough that the rebuilt section matches the source one-to-one.

08Product launch video to Diffusion Studio

Jev can scan a product launch video and identify its visual components (text, buttons, drawings) automatically, replacing manual storyboarding of a rebuild.
Selecting the detected components and building them opens a ready-to-animate file inside Diffusion Studio, and stating a style preference upfront ('precise, minimal, dark') steers what gets built before it starts.

09Voice-controlled Subway Surfers

A separate open-source project already used Jev to play Subway Surfer autonomously, controlling movement at low cost and high speed.
The remix swaps autonomous play for voice control: saying a place name changes the environment live, and phrases like 'make it harder' adjust difficulty, because both are just short-phrase classification decisions.

Glossary

Terms worth knowing.

Jev: TypeSafe's 'System One' model. Given text, it returns a structured decision (a choice, score, or probability) instead of generating written output.
System One model: TypeSafe's category name for a model built to make fast, cheap judgments rather than write open-ended text, unlike a conversational LLM.
TypeSafe: The company that built and released Jev, currently accessible only through a waitlist at typesafe.ai.
Diffusion Studio: An AI-assisted motion design tool used in this video to turn a set of detected visual components from a video into an animated project.
Astra: A reasoning mode of OpenAI's Codex coding agent, used throughout the video to write the Chrome extensions and demos from plain-English prompts.
Section Lens: The custom Chrome extension built in this video: hover over any website section, lock it, and get back a detailed reconstruction prompt plus a reference screenshot.

Resources

Things they pointed at.

00:00toolTypeSafe / Jev ↗

00:43channelOri Silver / Maxfusion breakdown video ↗

07:33channelGPT-6 & Diffusion Studio video ↗

00:00linkRob Hallam's SuperX tweet scorer (X post) ↗

00:00toolSuperX tweet tester ↗

00:43linkOri Silver's Maxfusion ad-research post (X) ↗

10:49linkWeekly10x newsletter ↗

Quotables

Lines you could clip.

01:15

“Jeff doesn't generate any text at all. It can't write a single word or even explain its reasoning.”

the clearest one-line definition of what makes the model different→ TikTok hook↗ Tweet quote

01:06

“It operates in roughly 70 to 500 milliseconds per pass.”

concrete, quotable speed stat→ newsletter pull-quote↗ Tweet quote

10:06

“Miami. New York City. Rio de Janeiro. The moon. Mars. The sun. Tokyo. The ocean. The Sahara Desert.”

fast, fun montage of the voice-controlled game working→ IG reel cold open↗ 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.

Hey guys, now TypeSafe just announced this new specialized system one AI model called Jev. And there's some really cool use cases that I found on X, like for example, this one from Rob Hallam, where you add a text for a potential tweet, that text then goes into a classifier. And Jeff then asks each post about 61 questions all at once based on more than 50 million tweets.

Or this other example by Elvis on Twitter. In less than 25 seconds, it read 384 news from this morning and told 15 brands which stories to hop onto today. So it helps you get press coverage.

It helps you find trending stories. It helps you find journalists covering specific stories. Or we have this example by Ori Silver, who is one of the founders of Max Fusion.

And I have a great video on that. I'll dig it down below for you guys to check out. But he's using Jeff to make a competitor research tool and basically analyze thousands of ads in just 19 seconds to analyze your customer journey stage, ad style, and also cover a full account deep dive.

And so this model outputs fast and probabilistic decisions for software rather than just generating open -ended natural text like... normal LLMs would do. It's also very fast and very cheap.

It operates in roughly 70 to 500 milliseconds per pass. And it's cheap because Normally, traditional LLMs would charge more because the output is usually to generate a lot of text.

But Jeff doesn't generate any text at all. It can't write a single word or even explain its reasoning. And so it just evaluates the entire input state at once and then instantly takes a box on what it thinks to be true.

Like, for example, is this urgent? Or what category does this fall under? And now, you know, since you just learned about this, it might be...

hard to understand this all, but don't worry. The more that you get into this, the more that you poke around and test things out. the more you'll understand.

And now I know that more than 90 % of you that are watching this video are not subscribed to my channel. So if you like these types of workflows, consider subscribing. It would make my day, honestly.

Now, the first thing that you want to do is you want to go to typesafe .ai. And if you don't have an account already, you want to click on join waitlist. And then you want to just add your email and click on submit.

Now, for me personally, it took around four to five hours to get access to this. So don't worry, you're not going to wait a lot of time. And other people that I know that have also signed up to this waitlist say the same thing.

And now once you're in, you're going to get something looking like this. It looks like a traditional developer platform on the left -hand side. You can see you have your API keys, you have your usage, you have your playground.

But here in the middle, everything looks a little bit different. You have a few quick demos that you can play. You can check out their docs.

You can use their playground. which is also interesting to test out these different walkthrough lessons that they have available to you. But in order to get this started, all we have to do is go over here to this quick start and you can just copy this agent prompt and then you can open up any agent you want.

In this case, I'm going to be using Codex, but you can use Cloud Code. You can use Grokbot as well. And you want to open up a new workspace.

In my case, I named it Jev. And we just want to... paste in that prompt that we got from that website and then hit submit.

And also another thing that you have to do is you have to generate an API key. So you can just click on this over here and create a new key. And then you can copy and paste that to like a simple plain text file where you store your API keys for that specific project.

Now, once we have that installed, I always like to ask this, like what... What can we do with this? And usually this is helpful when you're doing this in a workspace that has other projects.

So this agent can kind of give you suggestions as to what you can do with these other projects or based on these other projects with this new skill that I gave it. We get this, right? We can build apps that understand language and turn it into decisions.

Choose an action, rank items, or estimate whether something is true. So it basically gives us the technicalities behind Jeff and a few quick ideas like a smart inbox, sort incoming requests, or create a brief matcher, or a transcript clip finder, which is something that would be useful for me, for example. Now, me personally, I use Grokbot for research purposes.

And on X, there are a bunch of really cool projects that people are working on. And this is like another way to kind of discover what things you can... use to test Jev out based on existing projects on X.

And so it basically gave me this list of different projects. And this was one of the tweets that I got from Anshu. So big shout outs to Anshu on Twitter.

But basically, as you can see, you can write down a bunch of different prompts on what you want the... canvas to have as a painting and it will automatically create those paintings using pure javascript no ai image generation jeff predicts every pixel's color in parallel and paints them and what's great is that he also included a github repo so we can click on this github repo and then we can copy that and then we can just say like try try to replicate this and then basically this is what we get right so a painting of grass and water let's click enter And it automatically paints that.

You can see the green and blue. We can say yellow, orange, and red carpet. Click on enter.

New canvas automatically creates that. We can say half black, half yellow. And we get like a kind of half black, half yellow thing.

So this is just like a quick example of what Jeff can do, right? So think of these like. different pixels.

If we zoom in, we can kind of see the little pixels over here. And since we give Jev a definitive answer or command, unlike an LLM that generates tokens, Jev can output thousands of structured decisions at a time. And thus that saves you a bunch of time.

As you can see, it generates it in seconds and cost. And something else that I found on Twitter was this Chrome extension by Peter Wang, where you can label all X posts on your timeline. And it can tell you if each post is clean, engagement bait, promo.

And it costs only three cents for a thousand posts. And he open sourced this project, which by the way, big shout outs to Peter Wang. And so not only can we build the same thing, but we can kind of tweak it to something else, like maybe YouTube, for example.

So we can copy this GitHub repo. And so for this, now we can say, build this for me using Jev and my API key that I gave you in that other project, which is also in the same workspace. I paste in that GitHub repo.

And then I can say, but let's make it for YouTube instead of X. And then generates that in eight minutes using Astra. And then under the Chrome extensions, you want to load that unpacked folder.

In this case, it's called YouTube scanner. Now I can go on YouTube and for example, I can hover over this video and I can view the results of this. So I can see that this is a tutorial.

I can hover over this one and I can see that this is more specific and more vague. And then I can scroll down and all of these videos kind of load in real time. As you can see, Jeff gives you these results in real time as well.

And then we can even remix this a little bit more. Let's say that me, for example, I work a lot with web designers and web developers. And what would be like a useful tool for these types of people?

We can make a new Chrome extension that... you know, we can visit any website, hover over any section and it automatically writes up a prompt for us to use to give to any agent so that that agent can then build that section that we got from that website one -to -one in Codex or Cloud Code. So over here, I'm in the TypeSafe website and I can click on this new Chrome extension that I created.

And now I can literally, as you can see, hover over anything. I can hover over, for example, this card or this section of cards click on this we see a snapshot of this and we also get the prompt that we need to use in order to build this and we can copy the prompt and also save that reference png and then just for an example i can paste in the prompt attach the image as a reference click on send and so as you can see we get the two cards one to one in this just random like little site that codex gave me so yeah you can really be creative and look for all of these open source projects and remix them to your own use cases like this one over here.

And now you don't always have to do this. You can always start from scratch as well. Now that you gave Codex your API key and you gave it that initial skill to start off with, we can say something like, let's use Jeff to basically provide a link to a product launch video and it can scan the components of that video.

And then it can use the Fusion Studio, which is an awesome motion design tool that you can use with AI as an MCP. And I have a great video on that that I just released.

Feel free to check that out. I'll link it down below. So using the Fusion Studio to generate those components one -to -one to then animate them.

And so it gives us this, and this is just like a random sample video. There's nothing special. But we can go on X and look for a really cool product launch video like this one.

And you want to just download this MP4. And then you can go over here, click on add video. You can either paste the video link or upload a video.

I'm just going to upload my video. And we can click on open. video is ready it's reading the frames and now all of the frames are red we have the video over here and we can kind of scroll down and look at all of the different components so as you can see like it's highlighting these different components and so we have like all of these different components like the text we have like the the little uh prompt input over here we have these like drawings we have these texts and i can click on select all and then hit build components.

And that automatically opens up this file inside of Diffusion Studio that I can eventually use and remix with other copy, with other colors, and eventually create a nice product launch video. And now apart from this, just going to the final chapter of this video. something a little bit more fun.

Using Jev with gaming is really interesting. This person literally vibe -coded Subway Surfer and it got Jev to basically play this game in super fast speed, controlling all the movements with a super low cost. But I thought to myself like, hey, maybe it would be cool if we also vibe -coded Subway Surfers, but instead of having Jev controlling the game, like the movements, we can maybe make like some like little microphone.

functionality where I can change the environment as I speak, as I manually play. I can change the environment of the game. I can change the difficulty of the game.

And this is what we get, right? We have Subway Surfer and we just have this little mic button over here. All I had to do was like prompt it and maybe adjust a few things here and there in terms of like the environment.

But basically I can turn on this microphone and I can adjust the environment. So let's turn it on. Miami.

New York City. Rio de Janeiro. The moon.

Mars. The sun. Tokyo.

The ocean. The Sahara Desert. Make it harder.

Make it harder. Make it harder. Make it harder.

Make it easier. Make it easier. So yeah, guys, really cool little example here.

But that's it for today's video. If you have any comments or questions, please let me know down below. But yeah, thank you all so much for watching.

Hope to see you next time. Goodbye.

The Hook

The bait, then the rug-pull.

TypeSafe just released Jev, a model built to decide instead of write, and builders on X are already using it to score tweets, research competitors' ad accounts, and label content in real time. This video works through five of those community projects, remixing each into something new: a YouTube video classifier, a website-section-to-rebuild-prompt tool, and a voice-controlled game.

Frameworks

Named ideas worth stealing.

03:26list

Starter Jev use cases (from Codex)

Smart inbox (sort incoming requests)
Creative brief matcher
Transcript clip finder
Natural-language controls
AI fact checker
Lead prioritizer

After installing TypeSafe's agent skill, asking Codex 'what can we build with this' inside an existing project surfaced these six concrete, classification-shaped ideas.

Steal forany workspace that needs cheap routing/scoring/classification instead of a chatbot feature

CTA Breakdown

How they asked for the click.

VERBAL ASK

01:44subscribe

“more than 90% of you that are watching this video are not subscribed to my channel... if you like these types of workflows, consider subscribing”

soft mid-video ask tied directly to the content ('if you like these workflows'), not a hard sponsor-style pitch.

MENTIONED ON CAMERA

00:00toolTypeSafe / Jev ↗

00:43channelOri Silver / Maxfusion breakdown video ↗

07:33channelGPT-6 & Diffusion Studio video ↗

00:00linkRob Hallam's SuperX tweet scorer (X post) ↗

00:00toolSuperX tweet tester ↗

00:43linkOri Silver's Maxfusion ad-research post (X) ↗

10:49linkWeekly10x newsletter ↗

Storyboard

Visual structure at a glance.

hookcold open00:00

built to decide

valuebuilt to decide01:15

valuetypesafe.ai02:05

valuejevinci repo04:28

YouTube extension

valueYouTube extension06:09

valuesection lens06:41

diffusion studio

valuediffusion studio08:16

valuesubway surfers10:28

ctaoutro10:41

Frame Gallery

Visual moments.

Frame at 00:12 from Testing TypeSafe's Jev: A Pixel Painter, a YouTube Classifier Extension, and Voice-Controlled Subway Surfers

Frame at 00:22 from Testing TypeSafe's Jev: A Pixel Painter, a YouTube Classifier Extension, and Voice-Controlled Subway Surfers

Frame at 00:28 from Testing TypeSafe's Jev: A Pixel Painter, a YouTube Classifier Extension, and Voice-Controlled Subway Surfers

Frame at 00:38 from Testing TypeSafe's Jev: A Pixel Painter, a YouTube Classifier Extension, and Voice-Controlled Subway Surfers

Frame at 00:44 from Testing TypeSafe's Jev: A Pixel Painter, a YouTube Classifier Extension, and Voice-Controlled Subway Surfers

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.

13:53

Lukas Margerie · Tutorial

Building a Claude Code Agent That Rebuilds Generic AI-Made Websites

A step-by-step build of a Claude Code agent that takes any URL, redesigns it with premium components, generates matching imagery, and ships two finished variants — then gets deployed as a chat-triggered agent.

August 2nd

08:54

Lukas Margerie · Tutorial

Claude Code's New Browser Mode Is a Game Changer for Designers

A Claude Code tutorial: circle inspiration on any live site and the built-in browser rebuilds your app to match it.

July 14th

10:36

Rob Shocks · Review

Jev: TypeSpace's Decision-Only AI Model Explained

TypeSpace's new model can't write you a sentence, but it can make a decision 200 times cheaper than the LLM you're already paying for.

September 17th

12:51

Mark Kashef · Tutorial

Jev Explained: A Classifier for Fact-Checks, LLM Judging, and Routing Decisions

TypeSafe's Jev swaps free-form generation for a constrained classifier: give it facts, a question, and the only answers it's allowed to return, and it answers with a confidence score instead of a guess.

September 18th

04:07

vogel · Demo

Jev Classifies 1,500 Real Inbox Emails in Real Time

A live test of TypeSafe's new classifier model Jev, run against the creator's own 1,500-email inbox export to see if it can actually replace manual triage.

September 16th

12:00

Nate Herk | AI Automation · Tutorial

Higgsfield Dropped a Pay-Per-Use API, Here's How to Wire It Into Codex

A walkthrough of connecting Higgsfield's new usage-based video/image API to Codex with a free skill, then the math on when that actually beats a $60/month subscription.

September 18th