• 0 Posts
  • 17 Comments
Joined 11 months ago
cake
Cake day: August 7th, 2023

help-circle
  • I kinda get where he is coming for though. AI is being crammed into everything, and especially in things where they are not currently suited to be.

    After learning about Machine learning, you kind realize that unlike “regular programs” that ML gives you “roughly what you want” answers. Approximations really. This is all fine and good for generating images for example, because minor details being off of what you wanted probably isn’t too bad. A chat bot itself isn’t wrong here, because there are many ways to say the same thing. The important thing is that there is a definite step after that where you evaluate the result. In simpler ML you can even figure out the specifics of the process, but for the most part we evaluate what the LLM said or if the image is accurate to our expectations. But we can’t control or constrain the output to exactly our needs, because our restrictions largely are just input in a almost finished approximation engine.

    The problem is, that companies take these approximation engines, put them in their product and consider their output fact. Like Ai chatbots doing customer support, and make up facts like the user that was told about rules that didn’t exist for an airline, or the search engines that parrot jokes or harmful advice. Sure you and I might realize that these things come from a machine that doesn’t actually think about it’s answers, but others don’t. And throwing a “*this might be wrong because its AI” on it is not an acceptable waiver of accountability.

    Despite this, I use chatgpt and gemini a lot to help me program, they get a lot of things wrong but also do great. It’s a great tool, exactly because I step in after the approximation step, review and decide. I’m aware of the limits. But putting these things in front of “users” without a review step means you are advertising that you are either unaware of this flaw, or just see the cost-benefit analysis and see that if noting else it’ll generate interest during the hype.

    There is a huge potential, but throwing AI into a situation where facts are needed when it’s only making rough guesses, is the wrong way about it.





  • It’s worth adding I greatly prefer MS Auth style authentication, since I don’t have to find the right entry to read the Auth code and then write it on the other computer. Instead MS pops a notification and you either type or select the right number, verify with fingerprint and done. Much more convenient.

    It often tells you what you login into and where you are attempt to log in from, so it’s a few extra layers of security for those that have that awareness to check those details.










  • All of them are OK, except mkv is less a file type and more a container. What should be specified is the code for video, which for most things I’d say AV1, but high res movies might not be the most suitable. Throw in opus for the audio track, and you can use mkv, but might as well use webm anyways since it’s more clear what’s behind it. (though can still be other things)

    I’d also add that jxl should be the standard for lossy images. Better than jpg. And you want something other than png for massive images because that quickly gets costly in terms of size due to png being lossless.



  • I’ve been generally happy with the documentation, but sometimes you need to see how something is supposed to be used.

    It’s not the python documentation, but the Qt documentation (which is used to understand PyQt) where I wanted to play some video. Well, you need a QMediaPlayer to control the video. But the video apparently needs to be a QGraphicsVideoItem for the actual video, but wait that also has to be in something, Turns out you need to put it into a QGraphicsScene, which ALSO needs to be in a QGraphicsView. And it wasn’t easy to find out, because some of these needed to take another item as init argument, while others needed to call a method to add them. So you are left to figure out what all these things are and what purpose each of them do differently, which order all should be created in, and try too keep track without messing up all the similar type names. One place where loosely typed python gets real weak. It’s seemingly obvious from the way I wrote it though, but having just the video playback as my starting point, finding out what classes to use to begin with was hard.

    Another case isn’t Python, but C#. Where some framework docs give class definitions but don’t tell you that those are used to be “subclassed” to override a specific method and then the class is set as a property on a object you want behavior to change for, using a class name that really wasn’t letting on what it was doing at all. This happened after ChatGPT came into existence, and it throw me a complete example. Other times tho, it makes up a similar thing… which sucks