---
title: 'Jev SystemOne Model Explained!'
source: 'https://youtube.com/watch?v=6piQcPa6V-U'
video_id: '6piQcPa6V-U'
date: 2026-10-02
duration_sec: 1071
channel: 'Piyush Garg'
---

# Jev SystemOne Model Explained!

> Source: [Jev SystemOne Model Explained!](https://youtube.com/watch?v=6piQcPa6V-U)

## Summary

This video explains what Jev is, a System 1 model from Type Safe AI, and why it's a game-changer compared to traditional generative LLMs. It covers the problem of using heavy generative models for simple decision tasks and demonstrates how Jev provides fast, structured answers.

### Key Points

- **Natural Language as the New Interface** [03:04] — Traditional systems used structured data like JSON, but with LLMs, natural language has become the primary interface.
- **The Overkill of Generative Models** [09:57] — Generative LLMs are trained to produce text, which makes them slow and prone to verbose or unexpected outputs for simple decision tasks.
- **System 1 Models Defined** [11:16] — System 1 models are built for fast, calibrated decisions, returning only the requested answer (e.g., yes/no, label, score).
- **Jev: The First System 1 Model** [11:31] — Jev is Type Safe AI's flagship System 1 model, trained with reinforcement learning for calibrated decisions.
- **Jev's Performance Advantage** [15:47] — Jev is two orders of magnitude faster and more efficient than existing LLMs on System 1 tasks.
- **Use Cases for Jev** [17:08] — Jev can be used for LLM routing, categorizing tickets, scoring issues, and building decision trees.

### Conclusion

Jev is a System 1 model designed for fast, calibrated decisions, not text generation. It's ideal for tasks like classification, routing, and scoring, offering significant speed and cost advantages over traditional LLMs.

## Transcript

Alright, so hey everyone, welcome back, welcome back to another exciting video and in this video, we are going to talk about GIF. So, in this video, we are going to understand what is GIF, what is the problem statement it is solving and it is not just another LLM model.
You know, there are a lot of LLM models in the market. For example, you have GPT-5.5, you have GPT-6 Astra, you have some models from the cloud like OPPO 5, OPPO 5.5, you have some cloud sonic models.
but this is actually something different you might have heard these terms like type safe ai system one models or jiv this video cannot demystify to understand that what exactly is this jiv and why this particular thing is a game changer this is a fundamental change this is
something built from the ground up and this is a very very very optimized and a very very very innovative model so with that let's start with the video so you might have seen this particular article which is by type safe ai okay and this is launched uh this is released on september 15th
1996 so this is a very recent one that is introducing system one models and jib okay before diving into system one models let's first understand a problem statement okay so what is our problem statement let's say you have
a system okay forget about ai forget about elements just a very very simple system design so in a system design you have a user right you have a user and this particular user is talking to some kind of a reverse proxy let's let's just take in the next for an example so let's say you
have a reverse proxy here and then you have a bunch of servers so let's say this is a server right so awsx server so you have let's say multiple servers horizontally scaled and this is a part of
let's say a auto auto scaling group ticket so this is a reverse proxy and then after party out you do entire services okay they are processing something or the another thing so it is accepting the request processing something giving back the response and giving the response back to the user
now typically systems
they are actually doing some kind of an event processing to the Kafka then on the Kafka you might have some consumer groups right so you have let's say this is a consumer and this is also a consumer so just for the simplicity let's say you have just two consumers
and then these consumers are taking some kind of rapid decisions that what should I do and what should I not do based on that you are doing some kind of a DB operation so let's take some MongoDB let's take MongoDB here
so let's say this is a MongoDB you have done some diet crud operations some crud operations let's say you have done here you have a Postgres DB as well so let's take a Postgres So, here you are doing some reads and writes in some courses, this one is also doing some reads and writes and this one is also doing some reads and writes in some instances.
So, typically a high level design, something looks like this of a system design. Now, with the AI, the natural language has become the first thing, okay, so natural language.
So, as soon as AI and LLM models came, everything has changed to natural language. Because when AI models were not there, what happened here, everything was very static. You used to work on JSON or XML. So we used to work on a kind of structured data.
The request of the user will come in JSON format. Then everyone will process the JSON format. On the basis of this JSON format, you will fire some events. Then JSON is processing and then we store it in either our MongoDB or database.
So you have some kind of a hard-coded JSON kind of a data. This is not a natural English language. But as the LLMs came, people are moving away from JSON. and what they are saying is that, hey, what if I can control whole my system using natural language?
What if everything is automated? For example, user can just say, what is my account balance? What is my account, account, if I can spell that correctly, balance? So, if we get this input, this is a natural plain English language, right?
Like user can go this much plain, my account balance is this much. Now, the thing is that you have to fetch the account balance from the database and this is a pure pure natural language so what you basically do is you give it to the
back end now in the back end instead of having these things instead of doing some kind of a db operation you have llm models all over the place so let say for this example my open ai let say so you will have an open ai model server let say this is a server for your AI You are going to send this particular query to the LLM backend And what will happen to it It will have some tools For example switch balance can be a tool Switch balance is a tool Your LLM is going to call this particular tool then do some kind of a looping mechanism So here we have a loop It is going to do some kind of a loop get back the response as a JSON and then send it back to the LLM
the user. So this is now how the modern systems have evolved. Har jagay, har ek jagay aapne kya kya LLM ko inject kar diya. Even here, yaha par bhi ek LLM injection ho sakta hai. Ki ho sakta hai mujhe
kuch tags add karna hai, mujhe usko kuch categorize karna hai, whatever is the user query, the human in the loop feedback. So bahut saari cheez hai. So this is how the modern systems have evolved. Yeh ek choti se layer hai, which is not that choti.
Yeh ek choti se layer hai, jo humare modern systems ke andar ho choti hai add. now let's take one more example and let's identify the bottleneck that as soon as I try to scale this particular architecture or as soon as I put it in such a collection
so what can be the problem, what can be the fundamental problem so let's say user said that hey I need a refund simple, very simple thing
and what our system will do if user said I need a refund for my order So let's say for my order number, let's say 37. So what should happen is that in our database,
we just need to add some kind of an urgency to it. Here this particular query of the user, how urgent is it, I just have to save it in the database. So I just have to put a tag on it that how much is its urgency.
Okay, what is the urgency? I need its order ID. So what is the order ID? And what needs to be done next? That is process something like this. Yes or no. Okay, I just need this much. Now what you can do possibly is that this type of query, if this query goes to your backend,
then you will give it to your LLM. Now this LLM can, what we will prompt the system is that for the given user query, these are few questions which I had answered. What will LLM do? It will process them,
will do some tool calls and based on that it can give you such a JSON or not. talking about the structure output you can get defined of a JSON but what this dev says
what this particular architecture says that this is actually an overkill this is where the problem starts if I want a text instead of predefined predefined
questions I want answers so for that why are you using a traditional element A traditional LLM.
Because the traditional LLMs are trained to generate some text, right? The traditional LLMs are trained to generate some text. If you would have said generative. Okay, these are generative in nature.
The whole form of GPT is generative. That means if your query says something like write a point. Here you explicitly say that make me a point. You know, means out of nowhere generate a point.
and maybe write a YouTube script. Write a YouTube script. What is this particular thing? This is basically a generative thing. Generate for me. For example, I can just say write a code to do something like this.
These are generative things. What you can do with this? You can use your OpenAI model. You can use your cloud. You can use your grog models. Perfectly, perfectly fine. Because these models are trained. These models are very heavy.
to actually do these kind of tasks. But, if my question is that, Okay. But, if my question is that,
for this query, let's say user has said, I am saying user here, and user is saying, you know, let's say he said something, where is my order?
And what is the requirement for me? The requirement is, how urgent this message is. I just need urgency on the scale of 0 to 5. Am I generating something
new The answer is no I just want to categorize this particular thing I want to label it Just for example the question can be that is my order shipped Okay And what is my requirement Answer me in the case of true or false by looking up from the database
Am I generating something? The answer is no, right? The user said, is my order shipped? So, my LLM model gives me just true or false. Based on that, let's say, yes, your order has been shipped. You are getting my point, right? I can ask
something like, if someone has written an article, user has written an article, okay, article on, let's say, LLM models and what is the requirement of my system that do one thing,
just categorize, just put a label on it. Put labels on it. Labels can be AI, ML, or it can be something like backend. Just give me labels on it, that the user has written an article
or something, I just need these labels on it. Is it something like generating? No. So, for these types of tasks, if you are using a generative model, so what you are basically doing is you are using more time because those models are very heavy.
You are using more tokens. This is the absolute problem. And because the default nature of these models is generative in nature, so it is possible that you will not get expected outputs here.
For example, if you tell it that I want true or false, then it is possible that your traditional element will say yes to you it will say yes, it will say ok, it will say true with a capital with a small case t, it will say true or it will say
alright, yes, this is this order is shipped, ok so yes, this order is shipped
something like this, because their first instinct, their trading data that is trained to be generative and nature. So, for these types of tasks, where you have to take a decision in some way,
you want to do some kind of a decision making, you have some predefined outputs and you want just an NLP or something like an intelligence on that, intelligence,
so there basically you can use a Jiv model. And I will not say Jiv, I will say a System 1 model. System 1. And in System 1, the model which was released, that is a dev model. So type safe AI is actually the company name which is working on a system
one model so these type of problems are solved and the name of their very first model is dev latest. So dev does this. So if you go to dev documentation this is where the things actually become clear. Dev is a type safe flagship model for the fourth system one model.
Okay send state and type the questions get structured answers you can code using directly. Okay you can use your code. So if we go here I want to show you an example. Let's see an example in Quick Start.
Yes. So if the user says, Hi, I have been trying to connect my Stripe account for 3 days and integration keeps failing. I am losing sales. Please help me as soon as possible. So here you can ask a question to the jev model.
The question is urgency. N-O-U-L means to give an answer in Yes or No. Does this message express urgency? Your jev model will give you an answer in Yes or No. It is going to be a Boolean. That is it.
It can't frame the answer on it. Because if you use open or cloud here, it can happen because they have the context about the site, they start giving you the answer. That have you tried this? Have you tried that? No.
Here, Jav doesn't do this. Jav will only give you yes or no answer on the basis of this instruction. Because these models are trained from ground up, if you go to a table here,
because these models are trained on reinforcement learning for calibrated decisions, then first of all, it will be fast because it doesn't have any context about strike. It doesn't have any context about strike. It just analyzes what the user has said and whether there is urgency on its basis or not.
This is so fast, right? This is the example, okay? After that, if we move ahead, let's move on to the example use cases. So, for example, if we look here, let me just go into the example.
Okay, choice. So, here you can do what? Choice. What programming language is this code written in? you have given the option Python go and rush and then after that what can happen you give it code.
J model will just detect you and tell you that this code is written in Python. To analyze the code? No Do I have to do bug fixing in the code No If I have to do that then I can use openAI or cloud for that But it this this is the code and tell me one of these options
Similarly, what type of meeting is this based on the title and description? So, I have given options and I will get an answer from it. So, this is basically how Jez model works. If you see here, my running shoes arrived late in a wrong size.
Can I swap them for size 10? Which team should handle this? Should the exchange team handle this? Delivery? Or charge this? Then what will happen? Your job model will just return one of these.
It doesn't matter. It will not frame the answers. It will not generate on it. It will not write sorry messages. It will only answer as much as it has been asked. You can actually ask multiple types of questions.
Similarly, you can also ask scoring questions from here. So if we go down here. For example, the export button caches the settings page in Safari. it works in Chrome but a few of our customers use Safari
so what I want I want a score how swear this reported issue is ok and this is the criteria on that basis what I will get that number will be that's it it will not try to solve it it will not give any special tip
this is the right word it doesn't give you special tip so this is as much as you ask it will give you the answer then you can use a null null is basically yes or no thing right I asked three times
now can I please just talk to the real person. Now, is this customer asking for a human agent? Yes or no. Is repeated contact? Has the customer contacted support before? Yes or no answer.
And this is the reason that it is very very very fast. So if you see here the benchmarks, then there is a video here. This video is actually very good. So if I play this video,
So this is typeface AI and this is your normal tradition LLM. You can do things like LLM. But let's see how fast it is. Enter and you can see the answer is right here. So whatever questions you ask, you get the answers.
But if you look here, this is a generative model. So it has to generate all this. Okay. And here the probability is in LLM that it will hello-cinate and start giving some text output.
So this is basically your system 1 dev model. Okay. So if you see right here, we have built our first public model, Jiv is available today. Jiv achieves similar level of intelligence on system 1 tasks compared to existing LLMs
while being two orders of magnitude faster and more efficient. So this particular model, Jiv, it is actually more faster because I have told you, it is not trained to generate something. Ask a question, it will just give pinpointed answers and it is much much more cheaper.
So this is basically the thing that is here. okay and you can see you can read this article just in case you want like what is basically there and the documentation is actually very good okay so how to build a typeface so you can read all
these bathrooms and everything but if you want a dedicated coding video do let me know in the comment section i am on the wait list of the gist once i get the access i am going to create a coding video as well but i hope you got the gist that the gist can do this model but it is actually
an overkill to use a generative model just for getting a predefined answer. And this is where the Jev is basically pre-trained from the ground up. It is made from the ground up for these kind of tasks.
So, I hope it makes sense that if there are such decisions in your system that you categorize it, give me an answer in yes or no, what should be done next, I just want an answer in set of questions. I just want to do some kind of a processing
based on the text. I don't want to generate something. That is where the Jev comes. one very good use case can be an LLM router for example user asks a question you can basically ask
which model is capable of doing this if it is a coding question so plot is good, if it is a generic question so GPT is good and if it is some kind of a rage bait kind of a thing then Glock is good. So this kind of LLM
routing is also where the GIF can sit in an open router kind of architecture. So if you are making an LLM router or you want to do some kind of But decision tree, that is where the GIF can sit in between and give you the answers.
So, that was all about this particular video. Think of GIF like an if-else of the LLM world. Okay. It has faster and it just does the work. So, that was all about this particular video. I hope it gave you a gist on what is the GIF model.
So, meet you in the next video. Until then, bye-bye and take care.