Strange prediction, you say, didn't you fail to see LLM being useful and have to accept your defeat once? Yes I did, but this time, it's different.

As I have mentioned before with an insane amount of luck I got a job from an old friend after I returned to China. Unfortunately he's been a big fan of LLMs ever since GPT-3 became a thing (and I do sincerely feel guilty about being such a luddite and not going for the first-comer's advantages for him), but I was in a genuinely bad state so I was in no position to reject things. And as we met IRL and he told me all the things he's able to do w/ LLMs in the last 2 years I was like "yes, yes, very wonderful, very exciting" because what other responses could I really give that wouldn't make my predicament worse, really? Maybe my two years in Dublin did change my views on things, but after hearing he said that he could work "even when I'm, let's say, on the toilet!" my very first thought was exactly these words in English: "oh no… this is gonna blur the line between work and life isn't it?" and I immediately become slightly worried. And as the day goes by everyone was sharing things like "oh this thing can improve your productivity so much" and "oh that thing can improve your productivity so much" and I was sitting in my chair having this insistent worriness of "this can't be right" stuck to the inner back of my skull for whatever reasons like how those dried chewing gum residue stuck to the floor and being an absolute ass to clean up.

YMMV, but working on your smartphone is not a good feeling. The form factors of phones always make it feels like they're connection & communication tools instead of working tools for me, like how you'd have a palm-size notebook for taking notes but you'll never print reports in anything smaller than an A4 (or Letter if your country loves ridiculous sizes), but I digress.

Almost 4 months later, none of this acceleration ever slowed down (let alone stopping), and there's no sign that it ever would. I'm not necessarily happy about this, and that's why I'm going to make this whole blog post happen.

So right after I joined the company I started thinking about what will come next, some of which I've said in a different blog post. Basically my prediction was that self-improving agents are going to happen within 2026 and they will be able to self-improve via modifying themselves. This of course requires the agent to have coding abilities, so my idea was that before doing that I should make a coding agent myself just to see how it goes, which later became Tarha.

Before I started working on the idea I was thinking about actually going through the Agentic Design Patterns book, but as I was working through the book w/ my Ollama subscription and manual coding, OpenClaw happened, and it somehow became a very big thing in China. I'm not kidding you - within weeks after the initial rush of buying Mac Minis, they started to have meetings and conferences and lobster merchandises for this (mainly because it's nicknamed as 龙虾 and later shortened to 虾) and some started calling it "Web 4.0" (!) or some shit, and a good majority of them were hold by crypto bros because of course if anyone's going to hold meetings and conferences for this thing this early it's gotta be the crypto bros.

I suppose you could dismiss it all as yet another stupid craze, after all among those who bought it who really needed one? Definitely not every single one of those who lined up in front of Tencent's building for it. But I do invite you to stand in my shoe here - as someone who just ruined his family and himself financially and still got a good amount of credit card debt, you probably don't want to put yourself in further peril by not acting because your ego insisted on some random principles. So I was like, "there might not be enough time for me to do it the traditional way", and decided to stop reading that book and just vibe it.

Tarha is not the first project that I've vibe-coded for myself (I did a Claude Code session visualizing frontend thing before this because I didn't know you could just /resume) but that's not important. What is important is that it only took an afternoon for the project to go from zero to actually being able to code and Git Commit and all that. Of course it only worked some of the time and broke majestically during the rest, but as I sat there watching it crunching tokens I had this God-awful complicated fear that from now on only people with money and people who can shit out the fattest lies will survive since everyone who've seen this would naturally decide that people with knowledge of execution is no longer necessary after LLMs.

I kept working on it after that afternoon (I really doubt "working" is the correct word here, I just asked for things and things got done), adding all sorts of things a coding agent would have and also the "self-healing" abilities with which the agent should be able to collect its crash information (if it crashes, which happened a lot) and fix its own code. It didn't always work out - most of the time it didn't, and I am not really familiar with Erlang so it's a big pile of shit now. Good thing we have LLMs to deal with this, because I surely doubt if I should.

Yes I did another one; I decided to start anew because Tarha is a mess. For some reasons the self-improvement part never quite worked out despite starting anew. This one has a web frontend but it kept hallucinating non-existent WebSocket message formats nearly every time even if all I asked it was just to change its own color pattern. Sometimes it does succeed but it's not frequent enough. Now that Hermes came out I've basically giving it up and giving up competing in this front.

There were rumours that Hermes copied the self-improving idea off some Chinese team. I don't really care about this, but I would absolutely copy other people if I had the chance, especially when it comes to something not made to solve actual problems but to spend more tokens. "Oh no, they are thieves! They stole my idea!" - as if it's actually something that's so important it must be kept as a secret at all time. What makes these people think their idea worth as much as they thought they would in the first place, and why do people like this always think they're the only one who can come to the idea on their own? Not to mention that ideas has always been cheap and the fact that the execution part is cheap now doesn't mean the idea has risen in value.

you know things are getting dire when i start to buy brain supplements despite them being expensive as fuck

– yours truly, 2026.4.16

Three months into this whole vibe coding thing I had the most painful burnout I've ever had. You'd thought that the thing that made me want to quit computer science must already be bad enough, and yet somehow it managed to become even worse. Whatever has been made to require people in this industry by the adoption of LLM is so fucking next level. I genuinely struggle to understand how this can be felt as exciting and addicting. Seriously, don't you ever feel tired to have every single judgement being juiced out of your brain the entire time?

And even being put on a mental treadmill isn't enough, because you're expected to handle many agents at the same time as if context switching has always been a non-issue all of a sudden despite years after years of people memeing about how programmers hate it when managers disturb them with unsolicited meetings. If I remembered correctly someone from OpenAI said the upper limit of agents that one person can practice effective human-in-the-loop with is about 3 to 5. That sounds about right, I myself start to loose track of things when I get more than 3. A lot of people claim they have been managing 10+ agents on parallel; I'd bet none of those people have ever written a single line of code in their lives, and I'd bet all of the "judgements" and "decisions" are simply just them saying "yes" without a single fucking thought. Some people would also argue that C-suites have been always doing that, but of course they'd conveniently ignore the fact that that C-suites rarely get into the loop themselves, and a lot of times a lot of are not doing a good job managing people anyway.

Making judgements is not the same as thinking; ask any kid who had used a makeshift die fashioned out of their pencils or erasers for their school exams and they'll tell you that lesson. There's no so-called "training" of the so-called "taste" happening either: what kind of "training" is this where all you do is just pressing the Enter key, really? And because you have less time for thinking and less time for reaching those judgements since your bosses won't leave you the fuck alone and demand you keep up with the "improved efficiency", your thinking ability degrades, and in the end you're just making more and more flimsy and terrible decisions while burning out at a pace quicker than ever. I also find it absolutely ridiculous that someone somehow even thought the fact that people have always been writing shitty code could justify all of this; it's almost like saying that because you have already cut someone's fingers off accidentally so cutting off the rest of the hand is totally okay.

I don't know (other than trying to bet on world model so that I can at least have some chance of winning of course). I don't have a good conclusion to both this problem and this blog post. The only thing I know is how tiring this whole LLM thing is. Always the next newest thing. The next model, the next tool, the next agent, the next "claw", the next Indian (or not) tech influencer who'd make a fortune just by fellating himself for 3 weeks with a 20 USD Claude Pro subscription, the next whatever. I don't want this. I genuinely don't want this. But what choices do I really have?

I wish all this could just stop or at the very least pause for a while. Alas I'm not seeing the world willing to give me what I want any time soon.

---

Last update: 2026.5.2

Back