¶1Okay, so what you see here is now Jev actually controlling this first-person shooter game. You can see we are actually using some inputs here to decide what it's going to do. So, here it's going to take flank. You can see the probabilities of the next move here. And we also have signals.
¶2These are all the signals that is coming in. So, you can see we have health, ammunition left, position, and all of these information are kind of fed [music] into Jev here to actually help it make decisions. So, you can see we won the first round here using Jev. So, it doesn't make all the decisions, but it decides to break contact, take flank, push, hold, find cover, and reload. So, if you just watch it now, you can see it goes around.
¶3Now, it's doing take flank. Now, it's breaking contact. You can also watch it from here. That's pretty cool. So, you can see Jev here running around in this small game here.
¶4So, the reason I wanted to test this model was to see if we can use it in some kind of maybe some sports book on Kalshi, Polymarket, or maybe in other places too soon. So, this is the new model from Type Safe AI. They call it Jev. So, it's a bit of a different model. I think it's pretty interesting and it's very cheap to use.
¶5So, on their homepage, they call this system one models and Jev. So, they have a few example of what is kind of differences. So, this is kind of reinforcement learning of calibrated decisions. It's a bit strange. >> [music] >> Input is unstructured data with an emphasis on structured program state.
¶6Type safe structured values. That is pretty interesting. And sampling is not sequential. It's parallel. That is why it's so fast, right?
¶7So, that is why it's good at doing stuff like you can see here end-to-end response time is 70 milliseconds to 500. This can range from 40 to 200x faster for same levels. So, that is why it's so good at uh yeah, the game we saw, right? And we have confidence, use cases, AR powered workflows, smart if statements, structured outputs slot into ordinary software as faceted decision rules, right? So, you can see map producing our big data.
¶8So, I definitely think there is something interesting about this model. Uh I just want to explore a bit today how we can actually use this uh on sports book [music] and on other news and stuff like that. But I have no like great use case yet, but I just wanted to create this video to explore this model a bit because it's been really interesting testing it out on stuff I like to do. So, if you just take a quick look here, so what is Jev? So, it's an AI decision model that assigns probabilities to your options.
¶9So, for example here, yeah, this is just a couple of examples I have. So, we can input like a some context, a question, and some options. So, in our game, right? Enemy health is low, what should I do? So, we can ask Jev that with the input and context it has, it can evaluate, but it's going to give its response in actions and probability.
¶10So, attack 25%, retreat 15, take cover 60. So, when uh if we instruct it to always take the highest probability decision, it's going to take cover, right? So, yeah, your code can choose the highest probability action here, take cover. Also, we can kind of do how Jev makes game decisions. So, we feed it game data as I saw.
¶11So, we feed in data like health is low, a player is nearby, cover is available, right? And we can ask uh Jev again, baby, what should we do? Decision, take cover, confidence. So, it's always can also give like a confidence score. Um yeah.
¶12So, basically, we can give it some data, some options, and it can give us a response like a decision with a confidence score. This doesn't mean that it's always correct, right? It's just a confidence. If it gives like 50% confidence, then it's just a coin flip, right? So, Jeff chooses from the actions you allow, the game engine moves the characters and apply the rules.
¶13That's was from our game. But, I also wanted to do like a more relevant example that is I was looking at today in my yeah, testing of this model. So, today there is a game in our sports book, Chelsea versus Brentford. I guess it's Brentford versus Chelsea. And we can see the prices here on Chelsea on Calci, 39, 36, 27.
¶14Okay. So, there is a the striker on Chelsea is injury concern. So, he isn't exactly ruled out yet by Chelsea. They just said that he's uncertain if he's going to play or not. So, with this information, we can actually ask Jeff, "Let's do the relevance question.
¶15Is this injury news relevant to Chelsea winning?" So, the relevancy is yes, because he is their starter striker. And with him out, this team news could reduce their chances of winning, right? That is pretty interesting. And will this affect Chelsea price? This could be our next question.
¶16And our probability of yes is like 80%. This is just an example, but you get the point. So, if we can use this information is how I would use this information is to look at the price. So, we can try to gamble a bit, let's say before the lineups. If he is out, maybe the price of Brentford will go up.
¶17So, I had some plans to actually go check out the price on Brentford uh or Chelsea when the news came out about uh if how Pedro is going to start, if he's in the lineup or not. But, I see the time is already too late. So, I didn't pay attention. But, this should be priced in. So, this were like heavily rumored, and everyone should have priced this in into these prices.
¶18I think it would be different if he he does start now. I'm going to check it afterwards. I just see the timer is too late. But, we're going to go to Calci now and check the price. And if he does start, then I think we will see the price of Chelsea going up.
¶19So, let's take a quick look at the lineups and Calci. Okay, so as I thought he he does not start. So, if you kind of scroll down here, he's out, right? So, if you look at the prices, basically nothing changed. Just because uh we can go back like but you can see here around the lineups, nothing changed because uh yeah, that was basically priced in, I would say.
¶20I think maybe if he did start, we could have seen like a small spike here up. Uh but uh yeah. But, that was just an example of how you can use that. It's not like very That That is pretty easy to understand, but it's just like a simple example. So, let me show you a couple of the demos I built today to try to understand how we can actually use Dev in more of these type of trading, prediction markets, sports books, uh yeah, ways to use this.
¶21Okay, so here was kind of the example I did for the Chelsea game. So, is the news relevant? Should Chelsea move up or down? So, you can see here uh let's say we are on uncertain. So, this was kind of before the game.
¶22You can see no one is ruled out. We'll see if he will play. That was kind of from the head coach. Final decision will be pending before the game, right? And here you can see uh Jeb judges the meaning relevant to Chelsea price, yes.
¶23First order direction, uh yeah, 100, right? So, you can see uh here we kind of cold test the market. You can see now the price is 38 five, right? With without While we are keeping this uncertain. So, what could happen if he's ruled out?
¶24Uh, is this relevant to his price? Yes. João Pedro is unavailable, has not been uh yeah, something like that. Then we could see something like Chelsea pressure down. We didn't see that now because it's already priced in, but that was kind of the theory or the hypothesis.
¶25So, it went down to 35, right? Based on this news. Uh, so let's say he started, then it could be the opposite, right? We got the information semantic news that he is starts, then we can run this through Jev with some other data, and we can try to We could also try it to maybe to predict uh with some previous historical data what the price will be if he starts. Uh, but yeah, you kind of get the point here.
¶26So, this is one use case. I thought about maybe diving a bit more into Jev. Uh, also I built uh this. So, this is more like um Here we can look at more Yeah, it could be like an order book, let's say for 15 minute Bitcoin up and down. By the way, I have a really cool experiment running here.
¶27You can see I've been running this now for a while, and we are $12 up so far. So, that's going to be a a different video, probably early next week or something. But let's look at this example here. So, is pressure structurally supported for the order book? And we can feed it some data here, top imbalance, full depth imbalance, offset, and yeah.
¶28And then we can kind of let's say Here we kind of simulate. We run Jev on this question. Is top of book directional pressure supported? And the answer we get is like 72 here with the other probabilities. Uh, we can do uh liquidity.
¶29How fragile is the displayed book? We can ask all of this, right? Run the analysis. Bilateral thinness is our most is our answer to the question, what displayed liquidity topology will govern transmission of the next small truck? That was pretty advanced, but yeah, you kind of get the point.
¶30Or we can just do something like, is the market under reacting? That is a kind of more relevant. You can see we have the BTC spot that is up plus 84 basis points. Broad positive, and we can ask the question, ignoring direction, how does prediction markets repricing compare with the supplied external BTC impulse? And we can run this analysis, and it's under expressed.
¶31So, what I found out is this model is not good at actually trying to predict, give like a prediction. Here we have like a semantic compiler, right? I think that's the way to use it. So, we can kind of do ask some questions that is really hard to code. So, this could be like fuzzy questions like these are not so easy to code, right?
¶32So, they are really easy to ask in semantics, but they are not really easy to hard code. At least we need a lot of if statements and stuff like that, a lot of rules to ask those fuzzy questions. I think this is what I'm going to try to use the model for, right? So, like a semantic compiler by asking data some fuzzy semantic questions about, yeah, decisions. I also did try to build like a streaming where I connected Calchas API football from Pinnacle and Jev.
¶33Uh, it was working sort of, but I don't think it's the right uh exact right setup. So, it was working like it had some interpretation of the fuzzy regime layer here, but uh yeah, I couldn't really find any value in it doing this like this way. So, yeah, it's definitely interesting. I would say that. And another interesting part of this is the pricing.
¶34It's basically 0.042 in and zero out per million tokens. So, it's really cheap. It's almost free to use, to be honest. So, you could have a lot of fun with this. You could do things like I did in my game just because it's so fast and you can parallelize so many questions.
¶35We could do actually things like this with it. So, that's pretty interesting. So, today uh I just wanted to talk a bit about it because I found really interesting. And if you haven't seen people actually look at this model, I think it's definitely something you should maybe check out. Uh like I said, I'm going to try to apply this to prediction markets, sports books, maybe on hyper liquid, on some stuff like that.
¶36I wouldn't go too hard. So, I'm going to try to do something in shadow and just keep testing in paper mode and see if I find something that is interesting. Uh I guess I'll show you this, but I have been playing around maybe trying it in my 15-minute Bitcoin up and down setup. This is not it. This is a different system.
¶37I'm going to talk about it in a few uh in a few days, I guess, in the next upcoming video. So, definitely look out for that. But like I said, I think this is interesting and I'm definitely going to test it out and see if I find something. I'm also going to keep scouring and see how people are using it. So, maybe in a few videos I can come back and have a more interesting use case.
¶38But this video was just to talk a bit about uh how I have tested it so far. And yeah, hope you actually go check it out. Like I said, I tested it on Open Router and the prices are so fair that you don't really lose a lot of money even though you have to use the API. So, pretty cool to just see something new. Even though it doesn't work out for anything we can do in like real uh real setups, I think still it's pretty interesting.
¶39So, yeah. Thank you for tuning in. Not the longest video today, but uh yeah. See you soon.