Chasing The Dragon:
The "Development" Of Kapteeni v1
It's not as cool-sounding as you might think.
"Chasing the dragon", I believe, is an old Chinese euphemism for smoking heroine (or cocaine?). Apparently, when you heat up the drug, it will give out a curving thick line of white smoke, not unlike the Chinese Dragon, which is a long thing. You'd then have a straw of some sort to suck those smoke into your body, and this action was called "chasing the dragon". Of course that was just some random tidbit I've read on the Internet a long time ago; I never did hard drugs and never plan to (unless I got into some kind of serious incidents in which the doctors absolutely have to use opiates, but that's a very different story). But the whole strings of events I'm about to tell felt just as sinister as doing hard drugs; the first thought that came to my mind after it all (kind of) ended was quite literally, "ugh, I need a shower".
You might have already heard of the recent announcement of Jev. If you don't know what Jev is, it's a small LLM that's made to generate values within a pre-determined set; using LLM Agent jargons, it can be (very) roughly described as a zero-shot classifier, used for taking natural language inputs and routing them to different parts of a system. The traditional way of calling an LLM to generate the value is quite slow, and Jev did a bit of work and seems to have made it very fast. Almost immediately, it's known that similar things have always been readily able to achieve; and because Jev isn't open-weight, a bazillion open-weight replacements of various performance happened almost overnight; and eventually, someone made a benchmark suite for it.
And after a few days Kapteeni was born.
Before this I was in the middle of training an LLM; very small, just 0.5B parameters, mainly just for fun, definitely not suitable for any serious work (maybe I could somehow LoRA and make it into some kind of very tiny tool-calling model, but that would be another story). The machine I use for training is so slow it's not going to complete for another 3 weeks. You can't have things stopping you for 3 weeks, that's like a whole year in AI time.
So I made an executive decision: pause whatever I was doing, make (& publish) an open-weight Jev-like model as quick as possible, and try to get an entry listed in the benchmark list. Even if I have no use of it, getting something ranked publicly can still be a good line in my CV.
The Chase
Fickle promises of treaty, fatal harbingers of war
Futile orisons
Swirl as one in this flight
This mad chase
This surge across the marshy mud landscape
Until the meaning is forgotten
-— Van Der Graaf Generator, Arrow
Apparently 80+ people have had the same idea (and as I'm writing this part of the blog post there's apparently 10+ more), which means I'm probably already very late to the party, so I immediately got the coding agent started. At first I asked GLM-5.3-Flash to generate me a plan and execute it; that one didn't work out and wasted me an entire day. I switched to GLM-5.3 and this time it seems to have progressed successfully; within about half a day, I have a trained model and a whole set of code that wraps it into an HTTP API sitting on my hard drive. This, I thought, was clearly not enough: it's something that any rat bastard could've made with an LLM subscription within half a day, how am I going to compete and win over other rat bastards who at the very least is going to do as much as I did?… In hindsight, this may be the correct judgement, but it probably wouldn't have mattered too much, since there's only so much room you can have before you hit the capability ceiling of the base model, and with the help of coding agents everyone (including me) is going to exhaust that room and hit that ceiling very quickly(of course we are only going to fine-tune existing smaller open-weight LLM models and not going to train from scratch, which would take much longer, is borderline meaningless at this stage, and can easily costs hundred or even thousands of dollars on GPU rentals). I'll have to choose between getting into the benchmark list early (and risk being ranked low) and trying to achieve a better score (and risk not being listed in the benchmark list early enough and/or not being to get the said "better score" anyway). I chose the latter, because with so many people already lining up in front of me I might as well assume I have a day or two to spare, but if I'm being listed at a very low rank it's basically as good as not being on the list at all.
I asked the agent if there's any other ways that can give me a better benchmark score other than switching to a more capable base model. It suggested a few ideas, we spent about a day and a half testing them all out, and eventually all of them failed and we ended up having a version of the model that scored better in certain metrics than the old one but didn't manage to produce a better benchmark score. In a stroke of some kind of genius that I suppose I still have inside me I said to the coding agent, "if they're good at different things while having about the same score, we'll release them as different variants of the same version. We'll call them different names, like Intuit or Meticulous.", and that's what we did. I had the coding agent post the GitHub issue for me because I was too ashamed of doing it myself (for one reason or another), and that was that.
The Dragon
How long the time seems, how dark the shadow…
Oh, how straight the eagle flies, how straight towards his arrow!
How long the night is? Why is this passage so narrow?
How strange my body feels impaled upon the arrow!
-— Van Der Graaf Generator, Arrow
I think you've heard about the phrase "AI psychosis" already. I suppose most people, upon hearing this phrase, would think of those people who treated a chatbot as their quote-unquote "real" boyfriend/girlfriend and demanded Sam Altman to revert OpenAI's decisions when they decide to take down GPT-4o; of course there's more than just this one example, but overall, it's a phrase commonly used to describe people who ended up believing things that are ultimately untrue (both about themselves and the world they're living in) under the influence of consistent use of AI.
This may just be a part of the kind of AI psychosis that I myself am in, but I was slightly surprised that despite using it almost every day ever since I returned to China from Ireland I feel like I'm still at a considerable distance away from a full-on manic episode while so many supposed-to-be "elites of the industry" have fully accepted its speed and pretend (or maybe not pretend after all) that yes, having something that writes the software for them without any kind of deep thinking at all is the best thing ever since sliced bread and exactly what they've always wanted, examples of such kind of people including ESR, Robert Martin and DHH. Unfortunately for me everything about this still feels utterly sinister. Is this actually supposed to give you euphoria from supposed-to-be sense of accomplishment from vibe-completing yet another vibe-coded project? It's seriously been nine months, I felt absolutely nothing besides bafflement and utter confusion from this everything, from the hype around Jev, the dismissal of it essentially being "a fast text classification thing", and last but not least, the chase of that tiny bit of reach, of fame/noteriety, and of perceived legitimacy of your CV that you can gain by putting yourself on a leaderboard that some people probably would look at.
The chase… especially the chase. As I'm writing this the issue tracker of that JevBench repository has added yet another 30+ more issues requesting for "their" fine-tuned models to be evaluated. It's hard to say that I developped Kapteeni - or even any "development" had happened - especially when all I did was asking and waiting. At least with the tiny LLM I mentioned above I did learn (even if I do forget) how a basic autoregressive LLM works; this time I learned a near-zero amount of things that I do want to learn, and a lot of things that I probably don't want to learn (for mental health reasons). The drug is toasted, the smoke dragon is inhaled, and just like some ADHD people who apparently felt no high and only better focus after doing cocaine, I felt no sense of joy from all of this at all.
After The Chase
"Compassion", you plead as though they kept it in a box, hah!
Well, that's long since been empty
I'd like to help you somehow, but I'm in the self-same spot
Oh, my condition exempts me
-— Van Der Graaf Generator, Arrow
Fuckwad!! Of course someone directly posted their Jevs on Hacker News and of course the post on Hacker News got 500+ updoots and the repository got quite literally a thousand GitHub stars, because of course that's a much better way to farm whatever people are expecting to get than to wait in a stupid queue of someone's stupid benchmark. Of course. Man, what's this sense of disgust and complete defeat I just felt, and was it just me being too stubborn and clinging to the outdated morals?
I do plan to make a new version of Kapteeni though. People don't care about any language other than English, but I do. Maybe I can make it specialize in Chinese contexts (and maybe Japanese, and Korean) and gain an edge that way. This kind of things is what I have to consider now, when most of what I wanted is just writing my code in peace. Utterly insane…
Last update: 2026.9.30