AI can be based, sort of

Lounge
1

Generally, I'm not a fan. But I'm recently encouraging myself to push boundaries a little further on what I'm comfortable with, and AI is obviously the whole big craze.

I'm also buying a Pixel 10 pro soon which comes with a year of Gemini pro (Buds Pro and Watch too, go hard or go home). But the weird thing with Google is their constant generosity for specific developer cases, and AI is one of them. Recently they've introduced Jules! Which looks like this.

Screenshot_20250825-051205
Screenshot_20250825-0512051080×2424 312 KB

This is the ultimate laziness. The AI runs its own virtual machine, clones the repository, was smart enough to realise the snap build file actually points externally to GitHub, so then follows to the next for repository, reads it's documentation and generated this commit, after giving it a nudge that it hadn't noticed to remove a line that was interfering with other electron work. This results in this commit, that in true fashion, fails spectacularly.

After asking the AI to correct it's mistakes, it begins to panic. Giving me some interesting discussions that result in it being confused on snaps and the exact way I personally do them. That's ultimately a training issue I don't imagine to improve well for me personally.

Screenshot_20250825-052942
Screenshot_20250825-0529421080×2424 322 KB

Screenshot_20250825-053126
Screenshot_20250825-0531261080×2424 325 KB

Very soon, this results in a rapid breakdown in functionality, where ultimately, after having to remind the AI that git clone exists, results in me giving up, as it's getting overloaded.

Screenshot_20250825-053940
Screenshot_20250825-0539401080×2424 312 KB

Ultimately, despite being a massive failure, if say I'm ultimately impressed with it's overall functionality and ease of use.

The free tier is generous, this entire conversation ( which usually wouldn't last half an hour ) only counts of one of fifteen DAILY attempts, meaning, I'd encourage people to give this a go, the setup is trivial, and I'd imagine this would work significantly better on more standard repositories than my own.

But for now, maybe it can at least do the normal upgrades rather than custom patches, I'll try in a few weeks.

(It's one line but I am lazy)

1 Like
2

Since ChatGPT 3.5 was released, I have been using it to help me code for over 2 years. Although I haven't used "fully managed" coding agents like Claude Code, I use it daily, of course. Any regular developer who doesn't use LLMs may be at a competitive disadvantage.

1 Like
3

I am interested in getting GitHub Copilot, personally I don't actually code enough to benefit from it, but I have programmer ADHD. I won't do anything for months and then might do an entire personal project in a weekend on a moments notice. Often none of these are connected whatsoever and I'm jumping between environments systems and entire domains of whatever logic, it's broad, but ultimately, lacking any real guidance or ultimately commitment in anything in particular.

Which makes me think actually despite having limitations, it could be ideal for me, because I've always a gap between knowing what I'm doing but forgetting how to actually practically do it, and wonder if it might help actually get some enthusiasm back. And that in itself bridges a huge competitive gap in myself as a person professionally, tempting for sure.

Is there any distinct reason you pick ChatGPT and Claude? I know Claude is generally rated the best at programming tasks, but I do think Google are in a prime spot given maybe I want to ask my watch for directions every once in a while, despite that, ChatGPT with GitHub integration seems worth paying for too.

I still appreciate when stuffs ridiculously free though, so would also encourage Gemini Code Assist - Visual Studio Marketplace for people who can't afford the first party experiences where it's available. One of my main concerns with AI outside of AI itself is how this will ultimately put an economic disadvantage on education, or alternatively, exposes education to the Microsoft Word effect at a much worse level. But I can't change the world, so I guess we all have to learn to adapt.

4

GPT is the earliest released LLM, setting many standards, and with a decent web client, so I still occasionally use it for things like generating icons. The VSCode Copilot I use also defaults to GPT as its LLM engine, though I still choose Claude 4 in Agent mode, as it still performs noticeably better than GPT-4.1.

In the specific domain of Web, Claude Opus 4.1 has almost no strong competitors, although I also think Claude Code is really expensive. Of course, you can use many tokens to make it seem "worth it." But the main value of LLMs isn't in how many tokens are spent, but in the value of what you create with it. If I could use it to create a project earning $200 MRR, I would definitely choose it, but that's not the case yet, so I stick with VSCode Copilot + Claude Opus 4.1 (API).

Regarding potential educational inequality issues, I'm just an ordinary person who can't change the world, I can only try to adapt to it, that's all. I hope to create things useful to others, and LLMs have helped compensate for some weaknesses in areas I wasn't previously good at - it's a useful tool.

1 Like
5

I've used a combination of claude, gemini and GPT to create a few joplin plugins

I only have rudimentary programming experience (some classes I took years ago), and some "adjacent" experience (my job is technical support/QA testing).

I paid for a month of claude and have been milking the free trials for copilot and gemini for all they are worth :smiley:

I'm a bit mixed on gemini, it's not as good as claude for coding and more expensive than GPT copilot, but the big context window is nice.

Claude has been the most capable for dealing with more complex issues... but it's hard to argue with copilot at $10 a month. GPT 4.1 is good for smaller stuff and you can use monthly allowance of premium requests more sparingly for more complicated stuff. Going forward, I'll probably stick with copilot.

1 Like
6

What kind of domains are you using the AI for?

I've spent a few more hours on a Gemini via Jules for my main work (packaging mostly) and I feel as though even if I swapped in the other models, they'd probably still struggle for where I'd appreciate them.

My main interest in Gemini outside of programming is the multi-modality. E.G., the Gemini Live functionality where you speak to it isn't speech to text. The AI understands the raw sound, it was capable of identifying my accent and getting my home town within 15 seconds of audio, it isn't just doing voice typing and uploading raw text. Or e.g., it's ability to look at a photo and accurately guess my job description ( I look like a nerd mathematically I guess ).

I'm not sure how the others compare in those aspects but certainly for tech coding type stuff so far, I think I'm just unlucky whereas this kind of stuff would be way better for e.g., Joplin itself being on much more popular tech stacks.

I only have rudimentary programming experience (some classes I took years ago), and some "adjacent" experience (my job is technical support/QA testing).

If your job is anything like mine, “it's broken!” And I'm diagnosing anything from web tech to telephony at a moments notice, which would be easier to understand if I worked in IT, but I guess I vibe this idea of having a decent firm understanding in tech and yet in the real world it's a more vague because of how random “it's broken” always ends up being :joy:

1 Like