Big Idea

The argument in one line.

Jev is a new AI model that returns calibrated probabilities instead of written text, running in parallel at sub-200ms latency and near-zero cost, which the video treats as enough to ship a decision-heavy business idea in one sitting.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

A solo builder or indie hacker who wants to add a fast, cheap decision layer (routing, scoring, ranking) to an app without training a custom classifier.
Someone comfortable directing an agentic coding assistant to build and deploy a full-stack app end to end from plain-language instructions.
A builder open to running their own VPS with a self-hosted platform like Coolify instead of renting managed hosting per service.

SKIP IF…

You want a conversational AI feature. Jev explicitly can't chat, write prose, or explain its reasoning.
You're not willing to touch a VPS, GitHub, or environment variables. The walkthrough assumes basic DevOps comfort even with an agent doing most of the work.

TL;DR

The full version, fast.

Jev, a new model from TypeSafe AI, skips token-by-token generation entirely: it returns calibrated probabilities and classifications in roughly 100-500ms at a fraction of a chat model's cost, with a claimed 0% hallucination rate on structured output. The video covers what that unlocks (real-time scoring, self-driving decisions, near-free computer-use agents), then builds a live example: a Typeform-style application form that ranks candidates by qualification tier as they answer. The three-step path demoed is pick a decision-heavy idea worth automating, stand up a single VPS running Coolify instead of renting multiple hosting services, then delegate the build and deployment to an agentic coding assistant.

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 – 01:41

01 · Cold open: Jev goes viral

David Ondrej introduces Jev, a new model from TypeSafe AI that generates probabilities instead of tokens, and previews the video's three-part structure: what Jev is, its wildest use cases, and how to turn it into a business.

01:41 – 03:53

02 · What Jev actually is

Jev is framed as TypeSafe's 'System One' model: it returns decisions as probabilities rather than written answers, built from three blocks (choice, score, probability), and one call can score several criteria at once.

03:53 – 05:25

03 · Old LLM vs Jev, side by side

A side-by-side demo (credited to 'Mattia') compares a normal LLM taking ~8.5 seconds to answer a fraud question against Jev returning the same decision as probabilities in about 100ms.

05:25 – 07:38

04 · Under the hood: parallel decisions and pricing

Jev generates every output in parallel instead of token-by-token, is trained with RLCD to keep its probabilities calibrated, and is priced so output is free and input costs $0.042 per million tokens.

07:38 – 12:25

05 · Real-world use cases and benchmarks

Demos cover predictive spreadsheets, self-driving decisions from sensor inputs, and near-free computer-use navigation, followed by TypeSafe's own charts claiming 0% structured-output and tool-call error rates versus other labs' models.

12:25 – 15:54

06 · Step 1: pick the idea

The framework's first step is picking a decision-heavy idea to automate; the presenter chooses a Typeform competitor that ranks applicants live, alongside a real example of an adversarial QA-testing bot built on Jev.

15:54 – 19:51

07 · Step 2: VPS setup and first build pass

A Hostinger VPS is provisioned with a one-click Coolify install while an agentic coding assistant (referred to as Astra) starts building the ranking-form app from a plain-language brief.

19:51 – 23:28

08 · Watching the agent build and test live

The shipped form is tested live: answers to qualifying questions move the applicant between disqualified, mediocre, qualified, and highly qualified tiers in real time, each round-trip taking roughly 200-500ms.

23:28 – 31:08

09 · Deploy to Coolify and go live

The code is pushed to a private GitHub repo, connected in Coolify, given its environment variables and persistent storage, and deployed; a rejected admin password has to be regenerated before the live app is reachable.

31:08 – 32:12

10 · Recap and close

The presenter restates the three-step framework, plugs a new Discord community and the Hostinger VPS link again, and signs off.

Atomic Insights

Lines worth screenshotting.

Jev generates probabilities in parallel instead of tokens one at a time, so a decision that takes a chat model 8.5 seconds takes Jev about 100 milliseconds.
Jev is priced so output is free and input costs $0.042 per million tokens, meaning 1 million API calls at 1,000 input tokens each costs about $42 total.
On TypeSafe's own benchmark, Jev logs a 0% hallucination rate on structured output versus roughly 1% for the strongest chat models compared.
The same benchmark showed general chat models picking the wrong tool up to 60% of the time on forced fast decisions, versus 0% for decision-only Jev.
Jev is trained with RLCD (reinforcement learning for calibrated decisions), so an 80% probability estimate is supposed to actually happen about 80% of the time, not just sound confident.
One Jev API call can score multiple criteria at once (department, urgency, refund odds) instead of running a separate classifier for each decision type.
A QA testing job that would need a full-time human tester was rebuilt as an adversarial Jev bot that clicks through a site trying to break every release for pennies per run.
The video argues the bottleneck for a new decision-heavy SaaS isn't the AI model anymore, it's a cheap VPS and an idea worth automating.
A single VPS running Coolify hosted the frontend, backend, and could host any agent workloads for the whole demo business, replacing several separate hosting subscriptions.

Takeaway

A probability-only model built for milliseconds, not conversation

NEW AI ARCHITECTURE

Jev trades conversation for near-instant, calibrated probabilities, and pairing that speed with a cheap VPS turns a decision-engine idea into a shippable business in one sitting.

02What Jev actually is

Jev is TypeSafe AI's 'System One' model: it returns decisions as probabilities and classifications, not written sentences, so it literally cannot chat.
Its interface has three building blocks: a choice, a score, and a probability, all from one API call.
Because one call can score multiple criteria at once, you skip training and hosting a separate classifier for every decision type in an app.

03Old LLM vs Jev, side by side

A normal chat model answering 'is this invoice a fraud?' takes about 8.5 seconds of token-by-token generation; Jev answers with a full probability breakdown in about 100 milliseconds.
The tradeoff is real: Jev can classify, score, and rank, but it can't write code, explain itself, or reason step by step.

04Under the hood: parallel decisions and pricing

Jev generates all outputs in parallel instead of one token at a time, sidestepping the weakness where a chat model's answer degrades if its first tokens go in the wrong direction.
RLCD training pushes Jev to report calibrated probabilities: an event flagged at 80% confidence is supposed to actually happen about 80% of the time.
Pricing is asymmetric: output is free and input costs $0.042 per million tokens, so 1 million requests at 1,000 tokens each costs about $42 total.

05Real-world use cases and benchmarks

Demoed use cases include predictive spreadsheets that score hundreds of rows as you type, self-driving decisions from a handful of sensor inputs, and computer-use agents that navigate a UI for a fraction of a cent per action.
On TypeSafe's own benchmark, Jev shows a 0% error rate on structured output and tool-calling, versus roughly 1% and up to 60% respectively for the general chat models compared.
Sub-200ms responses make AI usable inside real-time interactions, like a button click or a driving decision, where a multi-second chat-model reply is a dealbreaker.

06Step 1: pick the idea

The recommended idea filter: find software that's either purely rule-based or uses an LLM in a limited way, then replace that piece with Jev for a faster, cheaper experience.
Two concrete examples: a Typeform-style form that ranks candidates live, and an adversarial QA bot that clicks through a site trying to break every release for pennies.

07Step 2: VPS setup and first build pass

One VPS running Coolify, an open-source self-hosted deployment platform, can host the frontend, backend, and any automation agents for a whole business instead of several separate hosting subscriptions.
Because Jev's own API sits behind a waitlist, the workaround demoed was routing requests through OpenRouter, which already lists Jev as an available model.
The build was fully delegated to an agentic coding assistant: the presenter described the app in plain language, approved or rejected UI decisions, and let the agent write, test, and prep the deployment.

08Watching the agent build and test live

A live test of the shipped form showed the qualification tier changing in real time as answers were adjusted, each round-trip landing around 200-500ms.

09Deploy to Coolify and go live

The deploy path was: push code to a private GitHub repo, connect it in Coolify, add environment variables, attach persistent storage, then deploy.
Small but real friction showed up live: a rejected admin password had to be regenerated and the env vars updated before the redeploy worked, a reminder that agent-built still has manual checkpoints.

Glossary

Terms worth knowing.

System One model: TypeSafe AI's term for Jev's architecture: it outputs probabilities and classifications instead of written text, closer to a fast instinctive judgment than a reasoning chat model.
RLCD (reinforcement learning for calibrated decisions): Jev's training method, which rewards the model for probability estimates that match real-world frequency rather than for sounding confident.
Coolify: An open-source, self-hosted alternative to platforms like Vercel or Heroku that runs on your own VPS and manages Docker-based app deployments through a web dashboard.
OpenRouter: A unified API gateway that lets a developer call many different AI models, including ones behind a waitlist like Jev, through a single key.
Structured output error rate: How often a model fails to return data in the exact schema or format requested, used in the video as one of Jev's headline benchmarks.

Resources

Things they pointed at.

02:07productTypeSafe AI

15:54toolHostinger ↗

16:31toolCoolify

19:45toolOpenRouter

31:39linkDiscord (Humans server) ↗

Quotables

Lines you could clip.

00:18

“It doesn't generate any tokens. It generates probabilities.”

clean one-line thesis for the whole video→ TikTok hook↗ Tweet quote

07:14

“Don't even charge for output. It's free. So the price is only 0.042 per million output tokens.”

concrete, shareable pricing stat→ IG reel cold open↗ Tweet quote

09:50

“100 milliseconds, 150 milliseconds per decision, per action... this is just insane.”

punchy speed claim paired with a live on-screen demo→ TikTok hook↗ Tweet quote

13:50

“This is the life of an agentic engineer. You're just gonna set the direction of the project and you're waiting for the agent to make it into a reality.”

captures the meta-shift in how builders describe their own job now→ 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

So there's a new AI model in town called Jev from types of AI and this is a completely different class of AI models. Look at this. This is a Google Trends chart for the search term Jev.

It's going super viral. In fact, the release video has over 34 million views in like less than two days. This model is very different.

It doesn't generate any tokens. It generates probabilities. It generates everything in parallel.

It's ultra fast. It's ultra cheap. Yeah, there's a lot to unpack.

So the best way to explain this is imagine if you were early to GPT -3, right? Before it was available and understood by the masses.

The average person on the street has not heard of Jeff and he don't understand the implications and capabilities and possibilities that this new type of AI model unlocks. So for context, my name is David Andre and I've been making AI videos for nearly four years now, since 2023. I've interviewed some of the most credible people in the field and I myself have built software with AI that generated multiple six figures of revenue.

So when I say that... Jeff will unlock a new class of software. I know what I'm talking about.

I've spent thousands of hours coding with AI and this model is completely different. Just for example, it is over 200 times faster than the current fastest AI models. It's over 400x cheaper than some cheap models like Haiku or GBD Luna and it has zero hallucinations.

Yeah, this model is... unlike anything we've seen before. So in this video I'm going to explain what Jeff actually is and how it works then we're going to look at some of the most insane and useful use cases that people have managed to figure out with Jeff because it's a completely new type of technology so people are kind of discovering what this can do and how this can be used and number three I'm going to show you how you can build Jeff power software and how you can build this into a business because Think about this.

When AI was released, there was a wave of AI startups. Cursor was one of the first to AI coding and they've sold for 60 billion, right? Now there's going to be another wave of Jeff -powered software.

People who can build this and turn it into a startup, into a real business. So later in the video, I'll actually show you how you can take Jeff, turn it into a new software product and package that as a real business. So first, what even is Jeff?

Well, it turns your text prompt into decisions. And this is what Typesafe, the creators of Jeff, call a system one model. It returns decisions as probabilities rather than written answers.

So you cannot chat with it. It's not a chat model. It will give you like decisions, real probabilities.

Basically, it's something between just a programmatic if statement and a full LLM that's way faster, predictable, doesn't hallucinate. And yeah, you can embed it basically into any type of software. Now, the interface has three building blocks.

So, for example, let's say it's a customer support ticket, right? You get a choice. Which department should handle this ticket?

Then you get a score. How frustrated is this customer? And then you get a number.

What's the probability that they request a refund? And this you get in like 100 milliseconds. This unlocks real -time software because when your users are on a website and they click a button, they cannot wait 10 seconds for Fable to respond or like 5 seconds for Opus to respond.

But if you have Jeff, it's like 100 milliseconds, 150 milliseconds. Another reason this is very powerful is because one model can handle many tasks, right? So you can have a single Jeff with a single API call, and that same model can access the support tickets, search results, agent actions, and product descriptions.

You, as the developer, as the person using Jeff, you simply tell it what criteria you want, and you don't have to train a separate classifier or anything like that. Jeff just understands it right away. Now, I know that it can still be understand what Jeff is and how it works.

So after... seeing a lot of different explanations, this one from Mattia is by far the best, okay? 45 seconds, pay attention.

If you watch this, you will have a good understanding of what Jeff is. So a normal LLM answers like this. You might say, is this invoice a fraud?

I say, based on the line items and the vendor history, this invoice appears to be legitimate, right? Very slow, 8 .5 seconds. Obviously, it's a bit exaggerated, but that's how it works.

Token by token, very slow. Boom, Jeff, completely different. Let me replay that.

Jeff, you get the same prompt, right? Notice that. Same prompt.

Is this invoice a fraud? And then you get probabilities. And it doesn't take seconds, 10 seconds, nothing like that.

It's like 100 milliseconds, you know, 0 .1 seconds. Boom. Gives you the options.

And you define it, what do you want, right? It could be 2, 3, 4, 5, up to you. How many, like, probabilities you want, what categories you want.

So is this invoice a fraud? You would attach the invoice. And Jeff, in 0 .1 seconds, says 88%.

it's clean seven percent it's fraud five percent you need to review it right so this is how it works there's no output tokens it's just one parallel pass and every option you give it gets scored at once and you just get a clean score so for example you can use it in a game right there's literally thousands of different use cases right you have this character and he can shoot forward right left you give the options of what it can do And in a moment, I'll show you how that even applies to something like self -driving.

This character can play the game in real time and basically on the input, right? For example, here we probably like, is the enemy close? Is there a platform above me?

Or whatever. You just give it the relevant inputs and it can make these decisions in a matter of like 50 to 150 milliseconds. Now, obviously, all of the benefits come with a trade -off.

So the positives of Jeff is that it can pick an action, classify, score, rank, root, add like 100, 200x the speed, but it cannot write sentences, it cannot explain itself, write code, reason step by step. It's not going to like replace Astra or Fable. It's a different type of model.

So this is the important thing to understand. Now, before I show you some of the craziest examples I've seen of what people are doing with Jeff. Let me give you a bit more detail, that way you have an understanding, not only just like a basic understanding, but you understand it better than 99 % of people.

So, as I mentioned, Jeff avoids token by token sequential generation. Instead, the architecture is in parallel. This is the biggest weakness of LLMs, right?

Even people like Jan LeCun, they were always criticizing them for being autoregressive. Meaning, if the tokens at the start are shit, the answer is probably going to be shit, right? Because it's going to...

head in the wrong direction. But when you're generating in parallel, not only is this not a problem, it's much faster because it's not like one token at a time. It's like all the probabilities are being generated at the same time.

Same reason why parallel computing existed in the first place. As you saw in that example, one request can answer many different questions, right? The category, the urgency, refund eligibility.

So it's not like you need to like build five Jeffs into your app and constantly pay for API calls. You just have one API call and you list out the criteria you want and you get the probabilities for all these different criterias and unlike LLMs which tell you that every single idea is great and that it finally find the root cause and you know it knows what the issue is now.

Jeff is trained to express uncertainty right? So it was made with a new training method called RLCD reinforcement learning for calibrated decisions. So instead of saying like yes this event will happen it will give it a probability like this event will happen 80 % of the time and if it's not sure it'll give it less.

probability, right? It's expressed to make calibrated, accurate decisions. Now again, I need to really make this clear.

The price and the speed are the striking things here. The latency is between 70 and 500 ms. I only seen like people take like 300 ms when they were processing hundreds of requests, right?

Usually it's around 100, 150 milliseconds. The price is pretty insane because they literally Don't even charge for output.

It's free. So the price is only 0 .042 per million output tokens. So that is $42 per billion input tokens.

So again, output is free. You only pay for input tokens. So for example, if your app made 1 million different API requests with 1 ,000 input tokens each, you'd only pay $42, right?

And you can build a business that makes... tens of thousands of dollars like that. Now, as I promised, let's dive into some of these examples of the most impressive things people have built.

First, predictive spreadsheets from Nader, from Cognition AI. So, this, again, would not be possible with LLMs, or it would be possible, but it would be so slow, and it would be very expensive, right? Here, you can have a spreadsheet, and you can have some data, and you can, for example, type in urgency.

And Jeff will evaluate every single row against the urgency. So literally as you type, it figures out you want each row rated from no follow -up needed to urgent. And again, it takes around 100 milliseconds and it can do that for hundreds of rows.

Look at this. 100, 200, 300 rows. Boom.

Incredible. Literally classifying and labeling large amounts of data has never been faster, easier, and cheaper. Let's look at another example.

As I mentioned, self -driving. This is pretty incredible. Again, you just give it some inputs.

You can see like distance from cars on the right, you know, the speed. The basic inputs. And then it can make decisions like accelerate, slow down, stop.

Okay, red light. It wants to stop and it starts driving. Turn right.

Okay, there's an object. It drives around it. And this is the model.

It doesn't require hundreds of hours of like footage from the road. It just needs these inputs, a few radars or like a camera, and it can self -drive a car. And it can even avoid like unexpected obstacles on the road.

and it can do it in real time, which, you know, if you put in like Opus 5 or GPT -6, it would take a couple seconds to make the smallest decision, which is not possible for self -driving, obviously. If you need 10 seconds to react to a red light, you're probably going to get into a car crash. Another use case that received a step change is computer use.

This is... Something that has been possible, but not as fast and definitely not as cheap. Look at this pricing.

It's basically zero, right? Sam Altman and Dario and everybody, they talk about like intelligence too cheap to meter. You've heard openly I say that infinitely.

But only now with Jeff, this is actually the case, right? So this is getting a flight, booking a flight. Jeff is navigating the interface.

And look how fast he is. Astra and, you know, you cannot even imagine this with LLMs. Like, forget about it.

Even like running in a fast mode, right? It doesn't matter. This is just insane.

100 milliseconds, 150 milliseconds per decision, per action. And it doesn't like misclick. It doesn't, you know, output hallucinations.

Again, it's typed, like type safe AI. It's literally name of the company, type safe AI, right? So not only is it super fast and super cheap, but it also always outputs the structured output you want, the decisions you want.

It never will tell you like, hello world, when you want. probabilities. So this is an incredible improvement to computer use.

All of this took like $0 .004 and took just seven seconds from loading Google Flights to actually having the specific flight. So having the ability to literally embed intelligence into your software at a fraction of the cost and like hundreds of times faster than ever before will obviously unlock a whole new sea of possible softwares.

Like your imagination is the only limit, right? And these two charts should convince you even more. This is the structured output error rate.

And even like the best models, you get like Astra, Sol, Luna, they still hallucinate on structured output, right? So you give it adjacent schema that you want it to always output. And even like Sol and Astra, 1%, 0 .8 % will give you like incorrect fields or some issue.

Jeff is 0 % hallucination rate. Zero hallucinations. Next up, tool call error rate.

So, you know, you say like, use this tool called to browse the web and it forgets to use it. Or like, tool called to open files or scraping or whatever. These models, including Astra, which again, Astra is smarter than Jeff, okay?

These models are not competing. They're completely different category of AI. Astra was still, 60 % of the time, use an incorrect tool or forget to use a tool, stuff like that.

Anthropic models are a bit better in this, but still it's unacceptable to like have 1 .3 % error rate. You cannot build predictable customer experiences or like trading decisions, right? There's many people who are going to turn Jeff into a trading bot.

Obviously, you're going to be insanely advantaged having Jeff trading if you're beating humans, especially on like real -time news, right? Obviously, you're still going to get destroyed by Wall Street, but that's a separate topic. But look, just Jeff has 0 % error rate, right?

So stuff like this combined with the cost, combined with the higher speed is going to make so many new startups and businesses possible. So let me show you how to actually do this, how to use Jeff, how to embed this into software, how to build Jeff powered software. And yeah, how I would think about it if I was starting a fresh business from scratch.

That relies on Jeff because again, this is not gonna be an opportunity in six months. It is a new thing right now It is not mainstream right now. And yeah opportunities like this just don't come around every other day So if you just sit down for the next 15 minutes and pay close attention You will have an incredible advantage over everyone else in the AI field So step one in building Jeff our software or building a business on top of Jeff is the idea, right?

You need to have the right idea. So for example What I will show you in this video is how I would build a Typeform competitor.

Basically an intelligent form that ranks the candidates, the applicants, in real time. Basically solving my own problem when I have a waitlist or I'm hiring for a position. You really cannot have an LM reason over it and take 10, 15, 20 seconds for a candidate until you decide whether to show them a Calendly or not.

But 100 milliseconds, that's nothing, right? Another great idea is from Rafal. He built this test suite, basically, adversarial testing suite that tries to break each release.

And this is just Jeff, like, using our software, right? You can click around, open sites, and see, like, bugs. So this is, like, really incredible.

Something that, yes, you could have built this with an AI model, but nobody would put this in their CICD. It would take literally, like, a couple of extra minutes to do all of this. It would take multiple dollars, probably tens of dollars per deployment.

And, you know, with AI, you're probably shipping to production 20 times per day. At least I am. With Jeff, this is so easy.

This is like so fast, you know, it can run in parallel. He said it costs pennies, right? So it's finally financially feasible and also time.

Like, yes, you might allocate three seconds to this and Jeff just like clicks around and tries to break your site, tries to break your front end, tries to find all of the edge cases. And if you had to pay a human to do this, that's like a full -time job. Like there's a quality assurance people that just click around and try to break releases.

This is literally a full -time human job. Now it costs a couple of pennies with Jeff. So yeah, the first thing you need is an idea.

And again, it doesn't need to be anything like revolutionary. Like I'm not asking you to invent electricity. Just take existing software that is either fully programmatic, right?

Like deterministic, just if -else statement code, or it's powered by LLMs where it doesn't need to be powered by LLMs. A lot of the AI SaaS you see is just cope, right? They're using models in a very limited fashion.

It's not a full agent. It's just like some intelligence, some decision. Jeff can do that.

hundreds of times faster and hundreds of times cheaper. So you can literally disrupt existing working products just by building a better user experience with Jeff. So that's the idea.

You need the right idea. The second thing, you just need a VPS, right? You don't need to spend five subscriptions on front -end hosting, back -end hosting, this and that.

You just need a single VPS and you can host a full -stack web app on a VPS. I know a lot of people are scared of doing this, but it's actually super simple, especially with the latest generation of models. So let me show you how you can set up your own VPS in a couple of minutes and basically use it for automations, powered by Jeff, for building a full stack web app and you can host your full business on top of a single VPS.

You can also use it to host Hermes agent, OpenClaw or any other agent that you're running. But for the sake of this video, we're going to use it to host our full stack web app. without having to have a separate front -end deployment or separate back -end deployment.

None of that. So I'm going to use Hostinger for this. It's also going to be the first link below the video.

All of my VPSs run on Hostinger. Everyone on my team uses Hostinger. They just make it super simple.

It's one of the most affordable options and they have literally templates for everything, right? So you don't have to be a DevOps expert or a Linux professional. You can just...

do a couple of clicks and have your own vps in a matter of two to three minutes so again this is going to be the first link below the video when you get there just click on choose plan go with whatever plan suits your needs i think kvm 2 is a great default for a lot of people and a lot of use cases and it's enough to run a full stack web app for sure.

In terms of period, I would recommend going with 24 months to get the highest savings. Now to get even better deal, go to the right, click on have a coupon code and type in David to get additional 10 % off your hosting plan. Now on the left, you can select a server location, just choose something close to you.

And in terms of the operating system, you can either go with plain OS or select an application. But for this use case of building a full stack web app, I recommend going with a control panel and going Coolify. And this is basically an open source platform that replace a lot of the closed hosting solutions.

So it's very easy to just deploy whatever software you're building on top of a VPS. Like your own server, again, virtual private server, it's your own, and to manage your own deployment. So yeah, Coolify makes that a lot simpler.

As you can see, Hostinger has a one -click Coolify preset that you can just click. And yeah, as easy as that. next let's click on continue and finish the checkout and to finish the checkout just log in with your hosting account or create one if you don't have it and then obviously fill out your credit card and billing details so while the vps is being set up which usually takes like two or three minutes we can literally get started on building this business so i'm going to be using gpt6 astra medium fast and i mean fable is fine as well but astra is the best right now and i'm gonna literally select this empty folder let me prove it to you is this folder empty boom I'll check the folder, including hidden files.

Yes, this folder is empty. Okay. Great, now let's build a full -stack web app powered by the new model from TypeScript AI called Jeff.

Okay, this doesn't spell it correctly, so I need to spell it. Use the API to browse the web to learn more about it, and then help me build basically a competitor to Typeform, where we ask a couple of questions, and the app is kind of split in like a 60 -40 view, where on the left is the 60%. You have the form and we can ask like three or four questions.

And then on the right, we actually see like the admin view of what's happening with this new AI model from TypeSafe AI. Okay, it's also misspelled here. Boom, so obviously I'm doing a dictation to be faster.

We can nicely visually see the probabilities it's assigning for each user, whether he's like highly qualified, qualified, or like mediocre or disqualified, right? So these will be the four different things we want in the structured output. And yeah, so go ahead and just build this app full stack right now.

We're going to deploy it on Coolify on our Hostinger VPS. No overthinking, get to work and build this app. We're just going to watch it with Astra guys.

I'm not even joking. Let's check back to Hostinger. All right, so the VPS is ready.

As you can see, we're running Coolify on Ubuntu 24 .04. So again, the same VPS, you can reuse it for many different things, right? You can host your web app.

You can host your agents. Like you got a lot of CPU memory usage. utilize it.

Run it for experiments, run it for automations, anything that needs to be available 24 -7, just host it on your VPS, right? Literally every single month that goes by, I have more VPS servers than I had before. If you boil it down, it makes sense because as AI gets better, the two biggest markets in the software world will be hosting and inference, right?

And by far the best hosting option for your own VPS is Hostinger. So again, This is going to be the first link below the video and thank you Holstinger for sponsoring this video.

Now let me jump back into BB to see what's happening. So it needs a types of API key. So I'm going to cancel that and I'm going to say, we're actually going to use open router.

So find an open router API key somewhere on my MacBook and use that copied over into this project into a environment file. So actually what Astra tried doing there is try to ask for types of API key, which is obviously the type safe. is the company that created Jeff.

So they have their own API. The problem is that it's behind a waitlist, right? And not everybody got accepted.

So obviously there's workarounds. You can either use OpenRouter and just type in Jeff, boom, there it is. Or Vercel AI Gateway if you're already using Vercel.

It doesn't matter. It's any AI provider, inference provider that has types of API integrated. You can just use it to implement Jeff into your own software.

So let's see how this is going. I already have OpenRouter API key, so I... Just told it to find one on my MacBook.

If you don't have one, you can just create an account. Takes 20 seconds and give it to Astra to just cook. Just make sure to give it some limits so that it doesn't burn it.

But honestly, with Jeff, you don't even have to worry about the limits. You can literally charge up $5 and that will last you thousands of users in your app. Like really, the pricing here is incredible.

And yeah, so now we have our own VPS. We have... Astra building this thing.

So this is the life of an agentic engineer You're just gonna set the direction of the project and you were waiting for the agent to make it into a reality I mean with Astra, let's be honest with Astra anything is possible, right? So if you just have the right idea and a bit of work ethic You can build anything Okay, so Astra is finished there I had to tell it to stop overthinking because it was running a bit slower than my liking even though i had it on medium but anyways let's see what it built okay it called it signal so we have the classic ai thing just like lots of bullshit small text but that's whatever let's focus on the main thing right which is the core feature the form so what would you like to okay actually i'm gonna tell it to in the meantime remove all of the small text and details like everywhere you just added a lot of unnecessary UI elements like these small text descriptions.

None of that I approved. Just remove them to make the frontend a lot simpler, cleaner, more minimal. I'm not really sure what position this is hiring for.

Maybe I should have told it that, but let's see.

I want to delegate recruiting and hiring. Okay, continue. Boom.

Alright, so here we have some signal journey. Alright, criteria. Okay, here it is actually.

evaluating fit for a business automation service you can inspect the rules behind every decision highly qualified okay so nice this is exactly what i was looking for so this is how jeff in real time evaluates me as a potential lead so obviously i didn't know it was for a business automation service so let's uh let's do a founder executive continue let's see how that changes things okay slightly more qualified let's do 20 000 plus Okay, qualified now.

A lot more qualified.

Would you like to make it happen as soon as possible? So now we should be even... Yeah, so we went from qualified to highly qualified.

Nice. So obviously my first response wasn't the best because I didn't give Astra the criteria. It just kind of figured that out.

But yeah, this is how Jeff works. You can see that like basically it's real time, like 100 milliseconds. 150 milliseconds whatever actually it's here 488 milliseconds round trip okay so it's a bit higher than average but still pretty clean pretty good so see like this is insane so i'm just gonna turn on fast mode here and go low well stop doing the refactor and just help me deploy this on uh my host singer vps this is the beauty of the ai agents you can literally have them help you deploy it help you set it up they are very very familiar with uh Ubuntu Linux machine, which is what a VPS is, right?

Even when you're using Coolify, remember that this is still running Ubuntu Linux and the agents love Linux. It's the most documented operating system, the most open source operating system. So yeah, the agents, if they could definitely choose, they would want to be ran on Linux.

So that's why you should run a Linux VPS. Most of VPSs are Linux. There's some that are Mac based like Apple Silicon for different purposes, but just go for Linux one.

And with OMRG, a lot of people are switching their personal computer to Linux as well. So it's an easy prediction. The rise of Linux is literally happening in front of our eyes.

And, you know, in one or two years, more and more people will be running Linux purely because agents are becoming the way we interact with computers. All right, so Astra is asking, what's my qualified dashboard URL?

So here I can just click on manage panel. And I'm just going to show you guys how to do it. for the first time.

Literally, you don't have to know anything. Say like, I'm here right now. Just have the agents help you set it up.

It's as simple as that. I know it can be intimidating, like setting up a VPS for the first time or using Coolify for the first time, but you can literally just tell what you want done. Speak to Astra, speak to Fable, and these agents will guide you through all the decisions.

Okay, so first we need to create an admin account. So let me do that real quick. Boom, create account.

okay so next i'm going to take a screenshot of this part and share that with table done i'm here now okay continue boom so what did it ask for dashboard url what now providing screenshots to these am models are super op and to make it even easier obviously you could just use codecs computers to click through it but some of these setups actually do recommend you do yourself because you learn a lot by going through the setup so let's click on this machine and create my first project boom so let's click on deploy your first resource and then i don't know if we're going to go through github or just direct we'll see what's easier yeah so we're going to go with github here so astra is asking if we can create a private github repo Yes, create a private GitHub repo named JeffDemo and push the entire project in there.

Just make sure to git ignore everything like the environment file. It was already suggesting it, but it's a good practice to just double remind it. Let's click deploy and we need to connect this.

So since I'm already authenticated with GitHub CLI on my computer, Astra can just easily use it to create a new GitHub repo. and then it's going to give me the link to that. Alright, so here's our GitHub repo right here, gfdemo with all the code and correctly getIgnored environment variable, fully JavaScript, this is the code behind the Signal app right here, and with here we have the info, webhook endpoint, instance endpoint, and should have enough.

Alright, so here let's click on register with GitHub and obviously make sure to If you don't have a GitHub account, make sure to create one. It's one of the most essential accounts to have in the AI era.

I have two FA, so I need to use GitHub mobile here. Do not share one -time code, guys. I'm just joking.

It was a viral tweet of somebody sharing their one -time six -digit code. So this jumps directly into creating the app. So let's just click create.

And this will probably redirect us back. Boom, insert Paul Stories. Select the Paul Stories.

Let's select this Jeff demo. We're good like this. Install.

okay add a resource to the git repository okay let's select this load repository continue so on the left let's click on build pipeline let's go with a docker file here and again i'm just gonna have astra help me set this up oh yes the main thing is the environment variable here things are correct environment variables all right here on the left boom add add a open our api key boom value right here obviously do not share this with anyone keep your api keys private i'm gonna revoke this before uploading this video add variable okay all right so we should see four different variables two in production two in preview Now let's go to the left to persistent storage.

Right, slash app, slash data. Boom, add volume. Then domains.

We don't have any domain for this. Okay, we should change this to HTTPS. Here icon.

Boom, save. All right, so let's go top actions and deploy.

The build is in progress. Okay, so it's finished. So let's go to the top right.

Go to the link. here we need to enter the username and password so put an admin and the same password you put in the variable app password so it's just rejecting my password so i'm just gonna reset it i had to ask to generate a new one it's a random character so let's go back to the environment variables and change the app password right here boom make sure this is good okay okay update i'm gonna update this one as well boom boom boom looks the same update all right so again we need to go action and redeploy this is going to create a new deployment and hopefully the password works this time okay it's done 10 seconds faster than last time let's click on links top right boom gonna paste the password first admin and we're in so this is the new simpler version of the app like we requested so again This is for building automations, right?

So you can put this for anything like hiring developers, hiring lawyers, whatever. So let's see how this works. This time I'm going to be a qualified candidate.

I really have a lot of like repeatable processes in my business that I'm outsourcing to cheap overseas labor. And I feel like there's opportunities for automation. Yeah, let's see if I go into highly qualified.

Mediocre? Damn, why? Okay, I'm just going to go try to...

Try to complete the troll it so I'm gonna do just exploring So it's mediocre as well. I'm gonna say under thousand dollars that should move me more disqualified. No, okay Just researching Okay, so I'm still mediocre So obviously we could play with the criteria a bit more for Jeff But you can see that last time when I filled it out successfully I went from like mediocre to qualified to highly qualified now.

I was doing the worst possible submission and i'm staying in mediocre right so again we fully delegated to astra i haven't even configured the problem for jeff or anything like that but it already works we already have something that is extra intelligence compared to google forms or typeform and you can see that this really unlocks a sea of possible applications that were not possible to build before so if you want to build a business there's never been a better time to do this before and if you're not sure what ideas to go with think about what is now possible with jeff what existing softwares will be disrupted by either having a lot faster and a lot cheaper AI if it already has some AI or if it doesn't have AI implementing type safe predictable quick reliable intelligence into that software to make it better.

So really these are the three steps that it takes to build Jeff powered software and by the way I just launched a new discord server for humans it's literally called humans so if you want to be a part of that it's going to be the second link below the video make sure to join it's completely free and that's how you know about things like Jeff faster than everybody else so make sure to join the discord server and once again if you don't have your own vps what are you doing in 2026 go ahead click the first link below the video and set up your own hosting vps Again, one of the most affordable and easy options out there.

Super easy to set up all of my own VPSs run on Hostinger. So with that being said, hopefully you found this video valuable. And yeah, Jeff is incredible technology.

I cannot wait to see what this type of AI models becomes over the next year or two. With that being said, thank you guys for watching and have a wonderful productive week. See ya.

The Hook

The bait, then the rug-pull.

TypeSafe AI's Jev doesn't write a single word: it takes a prompt and returns calibrated probabilities in about 100 milliseconds, for a fraction of a cent. This breakdown covers what that unlocks, then follows the reveal all the way through building and deploying a working Jev-powered app on a VPS.

Frameworks

Named ideas worth stealing.

02:07concept

System One model

TypeSafe AI's term for Jev's architecture: instead of writing text, it returns decisions as probabilities and classifications from three building blocks (a choice, a score, a probability) in one API call.

Steal forany real-time decision UI: support ticket routing, fraud checks, lead scoring, ranking form submissions

06:36acronym

RLCD (reinforcement learning for calibrated decisions)

Jev's training method, which rewards probability estimates that match real-world frequency (an 80% call should hit 80% of the time) instead of rewarding confident-sounding answers.

Steal forevaluating whether any 'confidence score' feature is actually calibrated or just cosmetic

12:40list

Three-step Jev business framework

Idea
VPS
Build

The presenter's recommended path to turning Jev into a product: find an existing rule-based or LLM-lite piece of software worth disrupting, provision a single self-hosted VPS instead of multiple hosting subscriptions, then delegate the build and deploy to an agentic coding assistant.

Steal forscoping any small AI-powered SaaS idea before writing code

CTA Breakdown

How they asked for the click.

VERBAL ASK

15:54product

“Go ahead, click the first link below the video and set up your own hosting VPS.”

The Hostinger sponsor read is embedded as the literal second step of the tutorial's own framework (Idea, VPS, Build) rather than a separate ad break, and the same link is repeated as the video's closing line.

MENTIONED ON CAMERA

15:54toolHostinger ↗

31:39linkDiscord (Humans server) ↗

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.