LLMs Are the Future of Classification

I was one of the many lucky individuals who got access to Jev on its release day before they restricted signups due to demand, and of course, I was thrilled, seeing the implications it could have thanks to its sheer speed and cost-effectiveness.

I mean, it's nowhere near as smart as even the open-weight models we have now (well, to be fair, these have gotten really good), but the fact that it's so cheap and just smart enough for what it will be used for is what makes it so impactful. There's a reason why OpenAI cloned it with their Decisions API; it's that important.

Using Jev for Moderation

For the past few weeks, I have been running Jev as a reviewer, something akin to Discord moderators or Reddit moderators, if you will, to moderate content on some of my platforms.

As an example, above is the usage graph of Jev on my Discord server for the past week. I have it look at users' messages, edits, then historical messages when there's nothing to do, and flag any content that violates the rules.

Around 25,000 requests with each input of roughly 400 tokens, it has barely cost me $3. The best part? Everything is practically instant. What a time to be alive, eh?

If I were to use Luna 5.6 (which is what I used much more sparingly before Jev), it would've cost me roughly $25:

Say we disregard the price, there's still Jev's output return speed, which again, pretty much instant for Jev, is simply too valuable. Even the super fast models that the free web versions of various LLMs provider cannot come close, not even Celeris 1's incredible 158 ms response.

Anyway, Jev so far has been running very smoothly. I have it set to automatically dish out timeouts and deletion on any flagged content with a confidence above 0.95; a bit anecdotal, but from what I've seen, that's where Jev is almost always right with its flagging. For everything below that, a human review is needed, and I'm loving it so far, as it makes my moderation team's job easier.

So Cheap It May as Well Be Free

Seeing how cheap it can be on a big server, I have deployed Jev on the official Tower Defense Simulator Wiki for a few days to let it moderate wiki edits, user comments, etc. The wiki generally has less activity (200 active users at the time of writing), but the things that do happen over there are much bigger content-wise; hence, the token input would be much higher than usual.

Yet, all that cost me was just three cents. Yes, three cents for a completely automated worker that can moderate content 24/7, keeping your site safe and free from spam and abuse; it's incredible, is it not?

Not to mention Jev's scaling potential; it stays extremely cheap when you start small, and barely scales up in cost as your product grows. I'd say the cost is simply negligible compared to the value it brings; you might as well just forget that there is any cost to begin with.

I don't really want to sound like I was paid by TypeSafe AI, but really, it's incredible. You should be looking at Jev as if it's a function that a library provided for you to use, not a product you need to buy. It's extremely fast, extremely cheap, as if it's completely local! Heck, it's so cheap that if you just wanted to try it for a few days, you absolutely can! I genuinely doubt that some quick A/B testing would net you even a cent, compared to a single prompt from something like Gemini, haha.

Expect More Like It

With how great Jev already is, you can expect more like it in the future. I mean, as I've mentioned earlier, even OpenAI announced their equivalent of Jev already, and I expect so much more from other competing AI providers.

I'm not an AI expert, but I feel like these models will have a much easier entry point for new labs to develop their own 'System One' model, compared to the frontier where you'd need billions of dollars to even get the infrastructure ready.

So, try it out for yourself and see what you can build with Jev! TypeSafe has removed the waitlist for registration on September 27th, so you can get started right away (that, or use through a third-party gateway).