Open Source, Faster Jev is HERE transcript

I have this AI playing the snake game right on a machine locally. Everything is running locally using the new classification model named Leia. As you can clearly see in here, it makes almost like 30 decisions per second. And for Emperor, it takes sub 30 milliseconds to complete. And after just leaving it for a little bit, now the length is more than 140. And it's getting so long, but it still never fails. It finds a way, finds the right direction to actually eat the points.

And this new model is called Leia and it's basically the same mod system models as of Jeff the model that took the entire web and the entire machine learning or AI industry by a storm and it falls under the system one models which is a special type of models that only gives you type decisions like structured decisions and it's meant for classification. It does not output text as usual LLM for like you're asking it something it gives you an answer but this one is more for classification.

So you give it the issue with the different classes and it gives you a decision in sub milliseconds. And by sub milliseconds, this one takes less than 33 milliseconds to make an entire decision. And since type safe AI announced the new model, Jeff, people have been doing all sorts of things with it in really amazing stuff. So for example, people were making it play Super Mario Bros. And this was all being played in real time.

You got how fast it is because Jeff is really, really fast. and it just makes a decision super quickly and clearly tell like oh right jump um left jump move forward or pause or something like that until it reaches the finish while some of the for using it to have or find high intent leads and personalized outreach messages. So they give them like for example 700 leads and Jeff is going to classify which one is going to respond, what is the perfect message for this model or for this particular person and how to reach out to the person itself and some other folks are using it to book flights like literally control the browser using browser use.

So the guy behind browser use in here use browser use plus jev to have this super ultra fast crazy cheap. So this was entire like entire booking took 7 seconds and it only cost like 0.003 0039 to book a flight. And by book a flight, it literally opens Google flights and he searches for a flight for you. He selects a dates and he finds the cheapest flight. So yeah, the introduction of Jeff is literally changing everything.

And we're going to see a lot more models or companies actually adhering to the system one design and creating more models like Leia and like for example Kev, which is a tiny jeike family of decision models built on top of Queen 3.5 and it works pretty well as well. like there's 9 billion, 4 billion and 0.8 billion models and it's open source and uses open weights and you can just download it, run it inside of your computer, same as Leia and it's so good.

So you give it a decision in here questions depending on classification problem and it's give you like a structured output JSON with different probabilities and a score of which one to choose or what is the right decision to take. So I took the application or the demo that was built by Gregor from browser base and I wanted to it's open source by the way so you can just go and check the source code. It's called Jeff ultra fast and of course he uses Jeff like the API J because Jeff is is a closed source model and it's hosted in the API so you need the API but it's super cheap so you're not going to notice anything.

It's pretty cool. It's pretty fast. Well, what's even cooler is to be able to run your own open- source open waste model layer right inside of your computer and it would be much faster than Jev, maybe a little less accurate because it's completely just very first version. And it was released literally 2 days after Jeff's announcement or like 4 days or something. And it just works on an M1 Max, M3, M2, M4, any model, silicon, any MacBook Pro or any Mac.

Literally, it works super fast and super well. So I said why not just take this projects import it to use layer and I asked cloud code to do that for me and it did really well that's I was like shocked because cloud literally finished it in one session and I've got a demo working using layer completely undependent from Jeff's or using Jeff's API or paying anything completely in just free and this is the prompt I gave cloud code in here to take the source code and integrate Leia into it uh like swap between Jeff and Leia and it did a pretty well job like I didn't have to go like a back and forth between a lot of stuff.

It just like almost one-shotted everything and it was amazing to say the least. This is what the new demo looks like and this is the prompt we're giving away. So, find oneway flights from Zurich to London on October 20, 2026 for one adult in economy. Stop when matching flight options are visible. Do not select or book a flight. And it's selecting the Google flights demo in here. So, as soon as I'm going to go ahead and click like start demo, it's going to open up a new tab.

You can clearly see on my Google Chrome browser in the top right over here. It's going to wait for it for a little bit and it's going to start controlling it uh using layer and browser use in here to find exactly things. So right now it's like in a manual mode. If you want to just like go ahead and manually click on like choose next or you can click run automatically. This exactly what I'm going to do. So watch this.

It's going to be so fast. Like you're not going to be able to keep up. It's choosing Zurich is choosing London right now. It's going to choose probably the uh dates. So it's opening the dates picker. Waits for it. Chooses October 20th. It's choosing the one way in here from round trip and it just clicks submit in here just to find or see all the flights. And there you go. Once it finds all the flights in here for us, it just go ahead and like says Jeff reports uh complete in here.

You can inspect the page and this is exactly what we end up with. Completely controlled browser session. And this is the session that was controlled with a tab that was controlled by Leia and the the model that we have and as well as like browser use. There's another fixture demo in here using a fixture reading room and the prompt for it is like open the article about using finite choices to control browser agents. If we start the demo, you can also notice uh format which is like a fixture website for reading books or something.

It's going to open up and this is actually the decisions are actually now um starting to make. So if I click run automatically now you see it's going to be super fast. So he just clicks for us and he opens the page and he kind of like knows and he infers that this is the page about like choosing finite uh choices to control browser ages. Exactly what our prompt is trying to allude to. So it's it's really good at like you know uh following instructions and classifying things really well.

I think for me right now currently I would say Leia is a little less precise and accurate compared to Jeff cuz Jeff is much much more accurate. Well, as a start, I think this is a really great start for something big and and huge cuz cuz Jeb was released like a few days ago, right? And now we've got a open- source open ways competitor called Leia just few days after that. And people like at Jev they took well I mean their announcement they say I think they took 3 years in stealth mode developing and creating this engine and the decision model before they release it now and people just like cloned it literally in like five or 5 days or like a week.

So I think the future is bright for us especially for open source and open waste lovers like me and you. We love that and things are actually going to change. So it's not going to be closed source anymore. It's going to be literally so fast and so free and open source and you can run it right inside of your computer. So I would assume stuff like controlling the or computer use stuff where the model is going to use your computer and classify what to click and stuff using voice models or voice mode where you can talk to it and it just does instructions for you super quickly because we're talking like sub milliseconds like 33 milliseconds to make a decision and click something.

So that would be crazy and it's going to be life-changing. I'm pretty sure a lot of like use cases and new cool demos and applications are going to kind of like be born from Leia or the older models are going to come exactly the same as Leia cuz it's going to run inside of your machine. So, it's going to completely change the perspective. And for those of you who are thinking what the heck is system one models, what is system one?

Long story short in here, there's this book called Thinking Fast and Slow was published in 2011. And he explained the difference between system one models, which is more of like a fast, instinctive and emotional type models, and system 2 models, which are slower, more deliberative, and more logical models, which is what we have now as LLMs, which they tend to reason about things and find solutions for us, and we can talk like chat back and forth.

But system one models are more of like fast instinctive and emotional models. And that's exactly what Jeff or Leia or other type of models that fall under system one which are classification models come from. So an example of like system one uh for instance in here you can determine that an object is greater distance than another. That's system one. Localize the source of a specific sound. Complete a common phrase like war and peace or something like that.

Solve basic arithmetic like 2 plus two. read text on a billboard, drive a car on empty road, think of good chess move if one is a chess master and is to understand simple sentences where system two are more like slow, effortful, infrequent, logical, calculating, conscious type of decisions where like prepare for the start of the sprint, direct attention towards certain people in crowded environments, look for a person with a particular feature, try to recognize a sound, sustain a faster than normal walking rate, determine the appropriateness of a particular action in social setting.

Count the numbers of A's or other letters in a given text which basically now LMS even LMS are failing when you say count the number of Rs in the word strawberry. It struggles to find the number of Rs in that word and basically in a nutshell that's what and the very cool thing that tape safe AI in here they announced like yesterday evening that Jev is now available for everyone. There's no weight list and you can start using it.

And of course, I went ahead and started using it. And I would say their dashboard and the playground they created in here is phenomenal. I'm going to use this one to basically give you an idea of what the API looks like, of what interacting with Jeff looks like. So, in a nutshell in here, Jeff takes two inputs, the states and the questions that you have to ask it that you want answers for. So, we're going to start with the questions in here.

The questions are formatted in JSON format where you say, for example, is sandwich. And later on it's going to respond with like what is this sandwich? And depending on the type in here it's going to be responded whether with a score whether with a selection like a choice or a yes or no which we they call it null. So you select the type in here like null. You give it instructions. So is food a sandwich? And as you can clearly see in here we're putting food between quotes because food is a state variable.

So we're using that variable from the states to put it right inside the question in here. And you can clearly say here we're saying food is hamburger. We can give it a definition. And you can give it much of like other characteristics and feels as well to describe exactly what the state looks like. And for criterion here is actually what the model is going to use to define whether what we're asking it is true or false.

So we're saying true if a sandwich is a food dish where fillings such as meat, cheese, vegetables or spread is placed between structural stretch or false the food has no bread enclosing or filing or uses only a single slice of bread yada yada yada. And you can just like have your own definition. So the model can use it to answer that question for you. And because we're using food as hamburger in here, I'm going to go ahead and do run request.

It's going to give us exactly what we need. So is sandwich 94% true, 6% false. So it's going to respond with true. If for instance, we're going to replace the food in here with like for example avocado toast. And I'm just going to go ahead and like do send request. You can clearly tell in here now it's telling us 6% true, 94% false because avocado toast is not considered a sandwich. I mean per our definition in here of course and this what in raw code the response would look like with like the model that was selecting it the answer the users like number of tokens and stuff request ID so you could just have you know follow-up request and evaluation time so it took 93 milliseconds for the entire decision and just to give you an idea of why I'm super excited about the layer model the new open source Jeff alternative type of model you can clearly tell here they're almost the same example they have a state in here of like an email then they have like a little more complicated and advanced like questions.

So like saying department which email or which department does this email belong to and they giving it criteria like billing technical sales and the urgency of the email in here with the criteria the churn risk the refund request whether it's a refund request yes or no clearly these are the results in here and it's telling you it's billing confidence 94% the language is English everything is good and the amazing part about it it did all of that in 39 milliseconds I mean if you compare it to Jeff's actually timing or time.

It It's like three times faster. So yeah, I'm I'm really looking forward to the future of system one models and classification models like this. I'm going to be using him quite a lot inside of my projects, the demos I'm going to creating. I'm pretty sure a lot of people would be as well because this is just phenomenal. So anyway, thanks for watching. Um love to see you more on our videos next time and uh yeah, see you in the next one.

Why?