pixlmint

65 karmajoined 3 anni fa

comments

4 giorni fadiscuss

I have OpenWebui hosted on my homelab, but you can also just have it live on your machine in a docker container. I honestly just embrace the iffy feeling. Openrouter has very good telemetry (which is partly why I went with them) and it'd be pretty easy to notice when someone other than myself uses my keys. For the little agentic coding I do I like to use OpenCode, and if I need to ask a question in my editor I use CodeCompanion (neovim AI chat plugin). I quickly went to check what OpenCode does with the API key, and it doesn't seem to store it in the user config, so that's at least something. But yeah, really recommend OpenWebui as a ChatGPT replacement (though there are a lot more alternatives out there, I just already knew owui from when I was playing around with local models)

4 giorni fadiscuss

Last month, I cancelled my Claude Pro subscription and instead used those 20$ to purchase Openrouter Credits. Most of my knowledge-seeking questions can be answered by Gemma4, for basic code editing, Qwen3.6 27b is enough, and for really difficult tasks, GLM5.2 doesn't leave me hanging. I'm by no means a heavy AI user, so I'm even saving money going the API Credit route and relying on the smallest possible model depending on the task complexity.

8 giorni fadiscuss

I feel it's a little more extreme for swiss companies though, especially outside the country. My aunt who works in Rome in the tourism industry told us of a local company that had 'swiss' in the name, simply because of the positive connotation, even when they don't seem to have any relation to the country.

It's why I feel wary of making business with any company with the word 'swiss' in it's name

9 giorni fadiscuss

It'd honestly be pretty alarming. They just took the entirety of the internet to become rich, and now that they have something to take, they intentionally worsen the whole thing (by investing additional training data) to ensure resellers/ distillers get worse responses.

I'm obviously not taking the side of the resellers here - if their T&C don't allow reselling, then by all means restrict their access. But intentionally training the model to give worse responses when it sees a specific date format, would be pretty messed up.

9 giorni fadiscuss

ZHAW, a swiss institute. The locked-down browser we use is developed by ETH, a pretty highly regarded swiss university.

10 giorni fadiscuss

What do we think are the chances they trained their models to behave worse or even malicious if those special apostrophes are present in the system prompt?

12 giorni fadiscuss

I don't really see the value in testing something like this in the exam format. We have a lot of programming projects we do, where software engineering is the primary focus, and there I think is space to give an introduction on how to use AI for software engineering, what to avoid, current best practices.

But the class I was talking about in my earlier comment was specifically on the fundamentals of programming (how to write a for loop, how to write and call functions... all that sorta stuff). If you need AI to complete this, why even get the degree. Maybe in 10-20 years we'll think differently about this, but as it stands right now, I think someone holding a degree in computer science should absolutely know the fundamentals of computer programming.

12 giorni fadiscuss

all our exams are in-person.

12 giorni fadiscuss

I'm currently getting my Bachelor's in Computer Science, while in-person is absolutely necessary, it doesn't necessarily need to be hand-written. We now have the exam questions printed out, and we respond on the online exam platform, just like we would before. It is explicitly stated that no aids, including AI, are permitted. There's usually the prof and at least one more TA walking around the room, if anyone is seen with something other than the exam platform open they fail the exam.

Some profs also started requiring special software for doing the exams, which works fine but is pretty annoying to use since it requires Windows, but that application basically hijacks the OS, making it impossible to navigate to any different webpage/ application.

In 2024, we had some coding exams, and AI was explicitly NOT forbidden, which lead to them making the exam more difficult (because with AI, you can solve it so much more quickly, so gotta make it harder), which was really annoying, because it made it practically impossible to solve it yourself because of the time crunch. I hope they realized how wrong that was.

16 giorni fadiscuss

I can only speak from my perspective, as someone who's lightly neurospicy with a good serving of crippling social anxiety on top, but having to jump on a quick discord call with the maintainer of a project I was excited about wouldn't be a deterrent to me.

Yes it sucks, but it's better than not regulating whatsoever, and at least this way I could be more certain my contributions didn't get drowned out.

16 giorni fadiscuss

I'm pretty sure decimalenough was talking about having the project structured in an org and only allow org members to contribute, not to automatically allow contributions from people that are members of a completely unrelated, corporate-managed org like google.

17 giorni fadiscuss

JetBrains Mono does make it very hard to justify spending anything more than nothing on a good font though.

17 giorni fadiscuss

> Once most of the bugs are fixed

my brother in christ, I hope you're actually trying to be funny here.

18 giorni fadiscuss

This is cool, I vibecoded something similar for a school project, but this looks much more mature, thanks for sharing!

18 giorni fadiscuss

Anthropic has some very interesting views on Intellectual Property xD

19 giorni fadiscuss

Did you have your agent talk you into making this something separate over building on top of git?

19 giorni fadiscuss

Do they do the same when using the model through API in something like Opencode?

19 giorni fadiscuss

I cancelled my claude pro a couple weeks back. Had some doubts over it, but not anymore

26 giorni fadiscuss

Like I said, I never really used it for agentic work. I had previously evaluated locally runnable models with opencode (such as qwen3-coder), but found that it wasn't really feasible.

Since then I've adopted a different philosophy, and I actually prefer it this way.

I still very much enjoy doing most coding myself, but when I tried using tools like Claude Code, it felt very difficult to return to the codebase after letting Claude make some changes. Maybe that's just because of poor AI-use discipline, I don't know. But with smaller models, that's not even an issue. I can't just let it do all the coding and thinking for me, however if I can describe a function I want to great detail in plain english, then Gemma can write it for me, and it will most likely work. It's perfect for boilerplate.

I also recently worked with a web framework I'd never worked before, though I'm deeply familiar with other ones. So I asked it "I know how to do this in Y framework, what's the best-practice approach to doing it in Z framework?" and it was incredibly helpful, even pushing back on some of my 'bad' attempts at solving a problem.

I think GPT5.4 mini might fall into a similar category, in that it probably performs best when not overwhelmed with too many tools/ skills/ mcps, instead being given clearly defined tasks by an orchestrator model. I call those my token burners, as they're super cheap to run and have high tokens/second.

27 giorni fadiscuss

I’m working on my developer portfolio that deeply incorporates the forgejo API, which is where I host my code. It basically gives a personalized dashboard for my personal projects.