Due to not being able to secure a job I had to leave Dublin and went back to Occupied Area early this month. Shortly after I arrived I got picked up by an old friend to work for his new startup, which I'm more than willing to do. (Always glad to help an old friend out… although technically it's me that's being helped more in this case.) Via him I obtained two accounts that allow me to use Claude Code - you're expected to use it in your work. I used it for a few days and noticed a few things:

• It's good, but it's not going to replace me if status quo remains the same, although it indeed would make convincing people who don't code that you're still needed ten times harder because on surface it looks really impressive.
• You absolutely have to monitor the agent's work all the time - you have to know what your agent is doing, because:
• Other than context, the limit of the state of things is also the limit of any coding agents. They are not capable of going beyond this limit, which requires human input. Sometimes you'll need to untangle things yourself so that your agent of choice could continue, and because you need to untangle things from time to time you're still required to have traditional software engineering skills.
  • This is the case currently. Would it continue to be the case? I'm not sure.
• Some people might say having a coding agent enhanced their abilities to explore things they previously don't have the time and/or the ability to explore. From my experience you really don't get this enhancement for free: the more unfamiliar you are with the new things you have "become able" to do with AI, the weaker your ability at untangling for those things, and the more painful it would be when things go awry. With things as they are now, proper learning is still necessary if you want to be consistently able.
• If you expect to be able to create things via vibe-coding without paying effort to read and understand, you're simply wishing for free stuff and that'll never work in your favor.
  • Unless you are filthy rich. But then again if you're filthy rich you'd either know this already or don't need to know this at all.
• If your assessment of coding agents' abilities is only based on the Copilot that comes with VSCode, your opinion is outdated because Copilot sucks big ass.
  • Claude Code is at the very least two generations better.
  • I tried to have Copilot make a simple Golang backend for me in the weekend. It generated code that I have to throw away because it's more pain to make it into something maintainable than to write the thing myself. It also stopped editing the files directly for whatever reasons. Not good. If that's what people been looking at of course they're gonna have opinions.

I believe this era we're living in can be aptly called the "post-coding era" - an era where people truly have to accept the new reality of code becoming yet another automatically manufactured commodity and have to deal with whatever consequences that follow. It sickens me that my old software engineering principle can be interpreted as in support of vibe-coding (I suppose God do love playing pranks on His creations) - the fact that this is somehow possible does not in the slightest imply that I was correct, because when I wrote that thing it's still one year before Karpathy had come up with the phrase "vibe-coding" and what I had in mind as the opposition was the over-complicated method of Robert Martin's "Clean Code"; and the fact that vibe-coding might have been more user-empowering than whatever people have done in the last 20 years is so ironic it bothers me to no end. Alas, the pandora's box has been opened and the state of things is never going back.

One must understand that predictions, even when they are proven to be correct eventually, mean nothing if they cannot lead to actions. That being said, if you don't mind reading useless predictions, here are some of mine:

• Both AGI and AI replacing engineers are likely not going to happen in 2026.
  • But 2027… I'm not so sure.
• As much as I hate to say this - local open-source models are probably going to lose eventually. I doubt another Linux vs. Windows vs. macOS situation is going to happen for AI.
• I believe in Yann LeCun's judgement - there's a limit to the LLM approach and his "world model" is the next step. I also believe that LM-based agents (with storage and retrieval) can be seen (and treated) as a "prototype of prototype" of world models at an immense cost if and only if they can be continuously trained.
• Instances of models/agents should be able to be continuously trained by the environment they're in. Things should be in such a way that if, for example, I deployed an agent six months ago, that same agent should be "smarter" six months later and if I deploy another one of its kind now the newly-deployed one should be less efficient than the old one. If this is possible and is achieved, language model based agents can (at least partially) overcome its inherent limit. It may not be AGI, but it will be half a step closer.
  • Someone - either other people or myself - is likely going to do this in 2026.
  • I have my doubts here - not every continuously training algorithm is ever capable of becoming AGI, e.g. network switches continuously learn about the latest MAC addresses of the other side of its ports, but we're not going to see them starting wars against human anytime soon.
• Old-school coding would at the very least stay around as a recreational activity. One day people would post YouTube Shorts about manual coding like how they post YouTube Shorts about woodworking today. To be fair I would genuinely hate this, but what can you do at this point?

---

Back

Last update: 2026.1.13