I've finally tried to write some Kotlin with AI ag...

# random

[Robert Jaros]

I've finally tried to write some Kotlin with AI agent. Github gives me Copilot Pro subscription for free for my open source contributions, so I've configured Open Code to use my Copilot access. I've asked the agent to add Quarkus support to Kilua RPC. After less than an hour I've got some not working code and it ate all my monthly AI credits 😉 What am I doing wrong? Was it very naive of me to think I could play around with AI for free? 😉

[Stefan Oltmann]

Copilot sucks, that's the matter.

Try a decent model like GLM 5.3 Flash or DeepSeek V4 Flash - much better results.
They were free on OpenCode in the past.

z.ai gifts GLM 5.3 tokens from time to time. For example at this weekend I got 300M tokens. But claims run out fast.

Unfortunately, currently there is no good model for free on OpenCode.

But there is a z.ai campaign right now that gives you unmetered usage in a certain timeframe. That's what I'm using right now.
docs.z.ai/devpack/notice/event-glm-5.3-flash

[Stefan Oltmann]

If you're looking for 10% off, this would be my affiliate link:
z.ai/subscribe?ic=DBXIS8WQBV

I somehow missed to look for one before subscribing. 😅

[Robert Jaros]

Trying Big Pickle now - it's free on OpenCode and it seems to give me some nice results.

[Stefan Oltmann]

For some reason Big Pickle doesn't list for me.

The best free models I have right now are MiMo V2.5 and Muse Spark 1.2.

Muse Spark 1.2 is a slop generator. Very weak results. 😕
And MiMo V2.5 doesn't convince me, too.

[Stefan Oltmann]

Ah, Big Pickle is based on the older GLM 4.6

[Robert Jaros]

I'm not yet ready to pay for AI 🙂

[Robert Jaros]

It’s like paying for your own destruction 😉

[Stefan Oltmann]

Totally understandable, at least for text generation and the stuff you get for free.

I paid for DeepSeek V4 Flash because it was free for a long time on OpenCode and I liked the results. When it moved to paid, it was still very cheap. It was like working for 3 hours on a puzzle app and getting billed $0.20 😄

GLM 5.3 Flash came disguised as 0x Alpha and was free for 7 days. During that time, it
convinced me so much that I was ready to pay to keep using it.

I did a month of Claude Code once, but I felt it wasn't enough work for the money. The quota is very limited.

z.ai currently lets me use the model for 5 hours for free every day with the Lite subscription. That's a deal.

[Stefan Oltmann]

It’s like paying for your own destruction 😉

True. 😕

I'm really unhappy about the changes in our industry because I like thinking about code and writing it myself. But it looks like the job has changed, so I'm trying to keep up with that and experiment with different models.

That way, if I get a job interview, I can at least say that I'm confident using current AI agents.

I don't know whether that improves my chances of getting out of unemployment, or whether I'm just slowly getting out of touch with writing my own code. 🤷‍♂️

[Stefan Oltmann]

On the other hand I really value AI reviews that find way more issues than any linter.
And letting AI generate boilerplate unit tests that are easily to review and repetive to write is also something I don't want to miss.

[Stefan Oltmann]

I'd prefer a job where I'm not forced to use AI in a specific way (like letting it generate everything and just review), but can AI use to my benefit where I see it fit.
In that case I'd write the core architecture and important parts myself and just let it work on boilerplate.

[Stefan Oltmann]

And doing the reviews, of course.

[Stefan Oltmann]

My 2 cents:

Even with a competent model like GLM 5.3 Flash, I feel that AI-generated code can become complex very quickly. That’s because it sees an edge case and instinctively fixes it in the app. In the real world, for every edge case, I’m doing a cost-benefit analysis in my head. If something is very unlikely to happen but would require complex code to guard against, I often just drop it and live with the risk. That saves my time and keeps the codebase from accumulating unnecessary complexity. AI is diligent in that regard. It sees the edge case and adds the necessary code to handle it. But that diligence can actually become a liability. A model like GLM 5.3 Flash can’t always see why it might be better to knowingly leave a “bug” unfixed.

[Robert Jaros]

It's like working with a child ;) had to google myself to tell the agent how to solve the problem 😉

[Stefan Oltmann]

Free AI isn't worth it most of the time.

[Stefan Oltmann]

I'm still working on getting the slop out MiniMax 2.7 introduced. 😅

[Robert Jaros]

Free AI does have its advantages, though—it kept running for an hour while I was watching a movie 😉

[Stefan Oltmann]

You can have that with paid AI, too. 😄

[Stefan Oltmann]

GLM 5.3 Flash is working for me all day. 😉

I used the prompt I shared in this channel for a review. I did go through the generated TODO.md, removed what I didn't want, added my remarks and let it work on 30 tasks in a row for the last 6 hours or so. 🙂

[Stefan Oltmann]

Due to its sloppy code I won't let Muse Spark touch my code any further, but I also use it to review the code and create issues, because every model has a different view/perspective on the code.

[Stefan Oltmann]

I'm not at the level right now where I create a orchestra of different LLMs reviewing each other, but I think I'll eventually get there one day. 😄

[Robert Jaros]

AI has implemented Quarkus support for Kilua RPC over the weekend. It needed some guidance and help from me, but almost all code was written by AI. Without the support of an agent, I probably never would have done this, given my lack of time and complete unfamiliarity with Quarkus. And I didn't spend a single cent on this 🙂 It was an interesting experience.

[Stefan Oltmann]

Yeah, indeed. Love that feeling, too.

Having DeepSeek V4 Flash and GLM 5.3 Flash in OpenCode (while I lasted) was great.
So much work done without anything spent. 🙂

The best option for free AI is currently hoping to claim a weekend build pass.

x.com/zcode_ai/status/2095847518194307434

I'm currently subscribed more or less because I want AI to finish what it started. 😄

[Stefan Oltmann]

Don't use Muse Spark. I'm finally done with it. Soo many subtle errors. It's a really bad model. Way too stupid for its size.

[Stefan Oltmann]

@Robert Jaros You wanted good & free AI. In ZCode you can grab right now 300M free GLM 5.3 Flash tokens („Weekend Build“).