Claude Opus 5.5 + Jev = 10x Power (I'm Never Going Back) transcript

I'm starting to think that Jev was the most important AI tool release of the year. You can see right now Claude is working alongside Jev to do work on my behalf. And it's not because Jev is the greatest AI model. In fact, it isn't even an LLM. It doesn't even interface with text. It's what it can do that is really impressing me. the speed in which it can process data and the capabilities and different use cases it unlocks are going to enable us to get so much more out of our LLMs like Claude and Chat GBT specifically.

I've been using it alongside Claude in order to automate a lot of work by speeding up decision processes, speeding up what would used to take far too long in terms of inference. And there's been a variety of use cases I've already been testing this on. One application is in my trading as you can see in front of you. I built an entire newsbot to send me news that may impact prices of certain assets. And because Jev can make decisions really quickly, it's able to categorize headlines as bullish, neutral, or bearish.

And in just a few milliseconds, automate a trade. I've also been doing this across highfrequency training using Claude as the brain for inference and Jev as the probabilistic determinator that is able to trigger a trade in around 400 milliseconds. And I've also been using it across business development because it's able to process data much faster, which is not only increased the amount of leads that I am able to outreach to, but it's also reduced cost versus purely using a model like Opus 5.5.

And in today's video, I'm going to actually show you some cost comparisons as to how Jev not only increased output, but actually reduced cost across the board. So what is Jev? Well, a traditional LLM speaks in words, whereas Jev speaks in probabilities. So, if you speak to Opus 5.5, which is an LLM, you'll ask it a question like, is this invoice fraud? And the LLM will use a bunch of words in order to determine an outcome.

So, it'll say based on the items in the history, this invoice appears to be legitimate. Whereas, when you ask Jev the same question, is this invoice fraud? It looks at the data and then it uses that to determine a probabilistic outcome. So, it assigns a probability to each potential outcome and just gives you a number. So it doesn't waste time needing to write a bunch of words. It just looks at the data, gives a probability, and gives you an outcome.

Agents within an application like Claude run in a loop. So an LLM decides what to do. A tool executes on the agents behalf a model. In this case, let's say Opus 5.5 evaluates the results and then continues in that loop until the task is complete. However, the problem with an agent loop is that it's extremely slow and it's extremely costly because every single decision requires a model call. And as we discussed before, every time you are relying on a model call, the LLMs, because they are large language models, that's what LLM stands for, they are conversing.

They are determining a factor and each word is counted as a token. But Langchain, who developed Jev, decided to take an entirely different approach. They didn't build an LLM. They didn't build a large language model because Jev doesn't operate with text at all. They describe this as a system one model which is a model that's just built to make fast structured decisions that software can easily use because it's just evaluating a state returning typed answers and probabilities and those probabilities can then be used to trigger the next decision in a chain.

But the power in my opinion isn't replacing Claude with Jev because that doesn't make sense. You can't talk to Jev. It's not a chatbot. We still need LLMs. The real power is using both for what they're good for. Using Claude or GPT as the brain. That's what they're smart at. They're great at planning. They're great at researching. They're great at writing. They're great at coding. But then using Jev as a decision maker, whenever you have something that can be scored, quantified or determined, which is the case for a lot of data applications, whether it be highfrequency trading, whether it be SEO, whether it be business development or data scraping, that is where Jev comes in.

So, let's look at some use cases where you could stack Jev on top of Claude and it could be a lot cheaper and faster. I think SEO is a good example. Borgger posted this example and then I'll show you some things that I'm doing internally where Jeff scraped 584 pages. You could see it alongside Opus 5 and to scan all of these pages for SEO it cost around 21. Now Opus doing the exact same task would have cost around $43.

Massive, massive difference in cost. And that's because Jev doesn't have to call upon the model every single time it looks at a web article. It doesn't have to say, oh, this web article has the word SEO in it. It doesn't have to ask the model, does this web page have the word dolphin in it? And then the model says, hm, I'm checking. Okay, it does. Or it doesn't have the word dolphin. And then it relays that data back.

Jev just looks at it and assigns a probability, in this case 100%, that that specific word is on the page. And then it's a very binary yes or no. It's a fundamental difference to how the tool works. This has great applications for something like business development, which obviously this isn't using my real data, but this is a real system that I'm using. And we were across the business development team reaching out to over 50 individuals per day across multiple platforms for a variety of reasons.

It's usually outbound in order to generate leads for our business which is either our consulting business or our agency business. And what we are now able to do is categorize as you can see here with a percentage the likelihood that this person will respond based on our criteria. Are they a warm fit? Are they in the correct market? Does it have the keyword triggers that we are looking for in their profile in order to reach out to them?

So, we can scan a bunch of data across the internet and we can verify our lead list and then our lead list is much more targeted. And if it's above a 50% probability, then Jev automates the outreach process. With Claude, it would cost a lot more and take a lot more time because every single time it would look at a potential candidate for outreach, it would have to do all of this inference. Whereas if you just give something a probability and then based on a confidence score, as you can see here, you have a percentage where it will trigger a specific outcome, then you're saving a lot of money and you're also saving a lot of time.

And I've ran some comparisons [clears throat] on this cuz I wanted to find out, all right, if I use Jev for a task and I just use Claude for a task, how much cheaper actually is it? And for emails specifically, per 10,000 emails, Jev cost around 14, whereas Claude Haiku, which is a much cheaper model, if I used Opus or Fable would be way more, cost around $2.20. So this is an example where the same action on the Jev side was so much cheaper.

And this is the real power in my opinion. It's being able to use Claude at what it's really good at, which is planning, which is the inference, like anything that requires thinking, strategic thinking, then delegate the probabilistic determination to Jev in order to make decisions. relay that data back to Claude. Claude gives feedback and then it runs that feedback and relays it back into Jev to tweak things even further.

For something like AI trading, which is something I'm really interested in. I'm running tons of trading bots and that's something I speak about on my other channel. This is great because it reduces the lag time on trade so you can get better execution, which is especially important when it comes to news trading. And you're not causing this extreme cost, especially with high frequency trading when Claude needs to constantly do inference on like should I enter the position, why am I entering the position, is it in line with the guardrails and it goes through that process every time because of how LLM's work.

Jev's able to just assign a probability. This is a setup we want to take. It fits the parameters. Bang. So I think if you're interested in trading as a specific use case, this is going to save you a lot of money. So how do you set up Jeb and how do you start using it? Well, on the Typesafe AI website, there's a quick start guide for you to try it in playground. You can also use it in Versal or you can install it through the Python SDK.

But what I recommend is that you just install it with Claude code. It's the easiest, most direct path. And what I'm actually going to do down below in my free school community is I'm going to leave the full setup guide. So you can drag the prompt into Claude Code. It'll not only install Jev, but it will also give you the exact architecture that I'm using across my workflows. For example, across the trading workflows, the news workflows, the BD workflows.

Anything that I do with Jev, I'm also going to give you for free. But the architecture and the theory behind how Claude communicates with Jev is applicable to a variety of use cases. So you can just set it up with the prompt and you're good to go. By the way, if you want to find the assets for any of these videos, you just go into classroom in school. You go into free assets library, scroll down to the video that you watch, for example, the Jev video, and then the file to download is going to be available there.

This is opening up a lot of use cases for people. There's a great website. It's called shipwithjv.com. And you could actually run this through Claude and give it your context to find out which business model could be good for you because there are tons of new use cases now which are available. Gaming is a big one because of how it's able to more quickly arrive at outcomes. We've actually seen Jev be able to make decisions a lot faster which actually help it compete in games like Clash Royale which before you know things were just far too slow.

That's obviously a little bit gimmicky but I mean if you're creating your own game that's going to be a massive benefit in terms of data applications. I have some friends that run analytics companies. They are already using it to determine the outcome on a particular piece of information. For example, if it's a company that has X data and looks at what posts outperform, it can scan for three set variables and give a probability that a post will go viral.

So, you could start to see the applications there as well. There's also a bunch of like AI powered agents. I really recommend just checking this out cuz people are doing really cool tools. You can go trading and markets. We're already seeing tools using Jev for onchain trading, for per trading. If you want browser use cases, there are already a bunch here. So, this is the website that I really like, shipwithv.com. There also some GitHubs as well, but a lot of it is here.

I recommend that you check it out if you do want to brainstorm some use cases because once you install the prompt and you want to start trying stuff out, you can start just downloading some of this stuff and testing it out yourself because ultimately, I can't tell you what to use this for. I've told you how it works, but you need to use it to speed up things that you do or try and pursue opportunities that previously were too expensive or were too slow and now you can do them faster or more feasibly.

That is really where this is going to come in. I'm going to have some more specific guides soon. So, make sure you subscribe to the channel on specific builds that I'm using with Jev. But I wanted to bring out this video today to give you the information. And if you do want more information on the trading stuff cuz that is one use case which is really applicable that I've already started using Jev for alongside business development and some other small things within the business and the way we review data and we look at content.

You can check that video out on my finance channel Mars Deutsche Finance and you can get that full video too. I'll see you in the next one. Have a lovely rest of your day. Peace out.