Jev in GTM: what it is, what’s overhyped & what’s great

So if you haven't heard of Jev, it's been exploding across X, with tons of different GTM use cases. There have been some very big claims that this changes everything for GTM. "This is insane." But some of these claims I think are very questionable, and there's been a lot of accusations around these claims being fake.

I've personally been working very closely on a lot of decision-making GTM systems quite recently, so I'm very close to this problem.

So in this post I wanted to go over what Jev is, what it's good for, what's overhyped, and then actually how I'm going to use it within my GTM engine.

(For anyone that's new: I'm trying to push GTM engineering as far as it can go. Each week I bring on a different expert or try out a different tool to basically add towards my GTM engine.)

and if you prefer the video version it’s here:

Very overly simplified, it is a classification model. It's just super fast and kind of generalized, and it's got a really good API. So that's what people are using it for.

But the selling point of Jev is you don't have to train it. It works fully out of the box. You just basically define some options and the state, and it picks.

So you're usually saying: okay, what are the things that you want it to classify? For example, what team should handle this billing message? If it is a payment, a payout or an invoice, it should be returned as billing. Then there's technical, etc. So that is one example where it's a choice.

Another example, which is a little bit closer to what we do, is lead qualification. So you have the state here, like: hey, who's ready to buy, level 0 to 3? And then you can have some sort of agent guardrail, so it's true or false.

And I like the analogy someone on X said: using Jev is kind of like throwing a ball to someone and saying "think fast". They just catch it, and whatever they catch, they get. If you throw it to an LLM, it takes it, it goes and researches, and it kind of does something else. With Jev, no matter what information it gets when it receives it, it just forces that into the structure that you've defined.

So one of the questions is, okay, what's Jev good for? Jev is really, really fast, and it's really, really, really cheap for classification.

Riley's was a good use case of Jev. An example here was email classification: okay, which emails belong in which category? If you're going to fill this out with an LLM, it'd be one by one, and a lot more expensive and a lot slower. Watch Riley's walkthrough here.

There's a few other examples, like video analysis. You saw from Movez where they're basically taking in all those posts and classifying them. See the post.

One of the cooler use cases, which is a little bit more on the agent side, is something like what LangChain did, with Jev as a judge. As someone that's done a lot of benchmarks before, I do think Jev as a judge could be pretty cool as an eval for agents. Because it is really fast and because it can classify really fast, you can do some pretty cool evals. I recommend checking out the LangChain blog. See the post.

Another example is Gojiberry, where they're basically classifying loads of outbound messages and defining them on a criteria. I don't like the word "predict" in this context, and I'll go over that later with some of the claims about predictions. But the use case and the demo they showed is pretty cool, and it obviously did very well. See the post.

If you're going to strip it back, it's not a scientific breakthrough in a lot of ways. BERT and other types of systems within this lane were just a little bit more complex to use: you'd have to fine-tune them and do other things to get them set up. I think it's kind of like OpenClaw in a lot of senses, where the hype around it is big because it's very, very simple to use. See the post.

And in all fairness to them, from the benchmarks I've seen, it seems to be the fastest. See the benchmark.

I also want to mention context.dev. They had another classification model which I saw on X was doing pretty well. I didn't have time to try it, but I thought I would mention them because it looked pretty good.

So what are people complaining about? Steve basically put out a video of all these fake demos of people using it for things that it probably won't work very well for. See the post.

I think a lot of the questionable demos were coming along when people were treating it as a decision maker or a predictor. An example here is the Tetris demo, where it's basically just not making very good decisions.

I think this video from Dread Jordan was pretty good to sum it up. There is a lot of what I'd say is misleading marketing. They claim it doesn't have hallucinations, but he was very right to bring this out, and I've seen other developers do this too. See the post and the video.

There's this idea on the landing page that says: hey, it doesn't hallucinate, so zero hallucinations. Where that is true, it's because you are configuring the things it can choose from. So it's saying: hey, based on the criteria you've given me, I'm going to choose option A, B or C. I'm not going to hallucinate.

But it doesn't understand any of your training data, and it doesn't have context of your brand. To build on this a little bit more: it might have three decisions, and a criteria for those three decisions, and it's not going to hallucinate at all, because it's always going to pick one of those three decisions with a confidence level. But it could be 100% or 90% confident something is right, and out of those three decisions just choose the completely wrong one.

Some examples of this are people using it in Tetris and trading, and it's making very questionable decisions. That's because it is trained to be generally good at a lot of things, and that is why it's so easy to use out of the box for a lot of use cases. But it doesn't have any context of your data, so there's clearly going to be a lot of things it's going to be confidently wrong about.

So it does have zero hallucinations in some sense, but that doesn't mean it's always right.

Another known limitation is the context window. I spoke to a developer about this.

For a lot of the use cases I've mentioned, you need to be able to give it data within its context window, to have enough context around your brand to make a decision. If you're doing lead scoring, you have to give it context around: okay, how do we score leads, or this type of lead tended to do well, that kind of stuff. But they've only got a context window of 32K, so you're not going to be able to give it enough data to make that really good decision.

So it's going to be good for the things it's been trained on and can easily understand. But for a lot of the decision-making process, the good decisions are going to be based on your data, and you're not going to be able to feed it good enough data.

so this is when we get into some of the claims. Not to shit any creators so i could be wrong, but I don't necessarily agree with this stance. I don't think it's going to be the decision-making layer as it is right now, and I don't think it has the context window for that. I don't think it's the right architecture to be a central decision-making layer for a GTM system.

I think if they make a much bigger model that you can fine-tune, and it has a larger context window, then maybe it could be the start of changing it forever. But going back to Dread Jordan's video, which I fully agree with: once they do make it bigger, is it going to be as fast? Are the benefits of what it has now still going to be the same? I don't know. So at least in its current form, and with what they've done so far, I don't agree with it.

I think they are pushing towards this decision-making framing, and I think it's good. I think they'll do very well. They're talking about this idea of smart software, and making it so software can make decisions faster. Clearly this is just the start, and I'm really excited to see what they do next. Watch their explanation.

when I see things like "Jev is changing ad testing", . this is an example of something I don’t think it can do very well. This isn't doing what it says it's doing, and there are a ton of different GTM examples like this, because people assume it can do predictions. But hey, if people want to do fun demos, that's fine as well. Just be aware: I doubt it's going to work very well. See the post.

Before I go into what I'm using it for, I want to explain a concept I've been thinking about a lot recently, and I think chess is actually the best analogy I can think of.

I want you to think of your GTM like a chess board. When you're making a decision about your GTM, like should I do ads, should I do this, should I make a post, it's made of many different moves. There are many different players, across many different places on the board, and many possible directions you can go. You can make a post about this. You could send a message to that person.

If any of you play chess on chess.com, you'll see that when you play, there is a best move, or a kind of ranking system for the different moves you can make. I think about GTM the same way. There are a lot more things at play, and much more than just one person making one move. But at a very high level, at one point in time, there are many different moves across marketing, sales, advertising, outbound, etc.

For example, you could say: okay, one of the channels I'm focusing on is content marketing. So me as the founder, I could have many possible moves or posts I could do today to act on my strategy. One of the moves could be: hey, I want to make a post about Jev, because it's trending. Another one could be: I want to do a demo of my thing.

But then you go into your employees that can make moves. You go into your buyers that can make moves: they could see a competitor. You go into agents that can update your CRM.

There are really only three types of players: you and your team can make a move; your buyer can make a move (can they reply to something, can they ask for pricing); and agents can make moves as well.

So really, your GTM is thinking about every possible move, and setting up systems so that people make the correct moves within their different lanes of ownership. If you're a solo founder, it's just: hey, today, what moves can I make that best benefit me?

I saw a lot of slop GTM demos for Jev, so I wanted to make something a little bit more grounded and useful.

For GTM, one of the things people never do, and I try to enforce as much as possible, is update the CRM. Even if we can update the CRM automatically when doing campaigns in HeyReach, oftentimes when it gets past a certain stage you might be messaging someone on WhatsApp, or you might have met them in person. So I want to give my GTM system the context of all the other states of the actors.

The use case I have currently: I have a channel in Slack for my GTM engine called buyer move update. Any time I want to update my larger system with something a buyer has done, I take a screenshot, add some context and post it in Slack.

In the demo I added some fake outbound messages from HeyReach, just to protect my data, then posted an update like "just followed up on LinkedIn, they said they want to meet in person". It replies and confirms the move has been recorded.

I'll be honest: I did try another use case where it pulled in loads of data from my CRM, my messages, etc., and routed them all to loads of different people. It basically told me different moves I should play. It worked okay, but I don't think it really worked. I feel like I could have got it to work much better through other methods. I'll keep updating on that.

So basically, in conclusion, I think it's a really cool technology, and I'm going to keep playing around with it. There clearly are a lot of cool use cases and things you can do really fast in GTM: the demo above took me like 30 minutes, if that. And I'm sure there are a lot of things I'm missing with cost cutting use cases!

but i just think if you think it’s going to predict the future based on data it’s never seen you’re ngmi

if you have got this far you should DEFINITELY subscribe to the newsletter here, since the whole thing is me building this GTM engine in public, trying new tools each week

thank you.

roman

(PS: If you do want to build systems like this, I'm the founder of a company called Scale Intelligence. We build GTM engines on a frictional gtm engineering basis or you can use our signal infra to build your own. Apply to work with us here.)