AI is moving so fast. Does it ever let anyone rest?
Originally written in Chinese. Translated with AI, reviewed by Gary. Read the Chinese original →
After a whole day of different meetings, I got home and was about to wind down before bed when a partner suddenly sent me an AI update. Something called Jev. I only meant to take a quick look, but the more I read, the more awake I got, and my head started exploding again. Fine, no sleep then.
What exactly is Jev?
Let me try to explain it with the business logic I'm more familiar with. When we use ChatGPT and ask it a question, it composes an answer and shows it word by word. Behind that is a unit for measuring usage called a token, which you can roughly think of as small pieces of text. When companies plug AI into their own systems, they usually pay by this usage. How much data it reads and how much it generates can both affect the cost.
Jev does something a bit different. It specialises in judgement and choice. Think of it as a "plug-in" in your workflow: you hand it the data, tell it what the options are, and it decides which way things should go. Technically it connects through an API, of course, and someone still has to set up the workflow.
Imagine you have a warehouse with thousands of parcels coming down the line every day. In the middle stands a sorter who glances at each one: this small one goes to a rider; that big one needs a lorry; this one's details are unclear, so put it aside for a colleague to check. Jev is a bit like the person standing in that spot, dedicated to routing quickly. Once sorted, booking vehicles, sending notifications and arranging shipments are handed on to the systems behind it.
These judgements could already be given to a general AI model. It reads the data, generates an answer, and tells the system what to do next. The answer can be short; it doesn't have to be an essay every time. Jev, though, is designed specifically for this kind of work and returns a choice and a probability directly, with no need to generate an explanation. Making ten decisions a day, you might not feel the difference. At a hundred thousand or several million decisions a day, the difference starts to get interesting.
The price is the real shock
What surprised me most was the price. I did a simple estimate using public API prices. Assume one million decisions, each reading about 1,000 tokens, with a general model outputting only 20 tokens. Using the cheaper GPT-5.6 Luna, it comes to about US$224; with GPT-5.6 Terra, about US$2,240; with Jev, about US$42. Under the same usage assumptions, that's roughly 81% to 98% cheaper. Of course, this compares only the cost of the judgement step. It doesn't mean the whole system saves that much, and it doesn't mean the three models are equally accurate.
But I still found it stunning. AI's division of labour can already be this fine. Content models write content, analysis models handle complex analysis, and small, repetitive decisions get a dedicated model of their own. If the cost of judgement keeps falling, could the customer messages, order exceptions and after-sales issues we used to only spot-check all be checked and routed, every single one?
Exciting and exhausting
This is the most tiring part.
You've only just understood one thing and started thinking how to fit it into your business, and another new thing arrives. Meetings and running the company by day, trying to rest at night, and your head is still going: hey, where could this be used? Can the system we're building be changed again?
Should you chase AI or not? Don't chase it, and you're afraid of missing something that could genuinely change your business. Chase it, and it really feels like there's never a moment to stop. Maybe the next thing to learn is how to judge what's worth chasing and what can wait. But that's easy to say. When I see something like this, I still can't help reading on.
Honestly: exciting and exhausting at the same time.