I'm really wondering: Guys, what's going on? This blanket aversion to anything with AI on it is almost religious and reminds me of the "mobile phone rays" opponents. Man, I'm really not a fan of this hype either, but calm down.
And blocking people who use AI-generated images across the board, well... I don't get it.
@phranck oh god this is the future isn't it. People just linking to unvetted LLM output as search results and I have to check every one to see if it's BS.
So... theft has been done before, therefore it's ok? What?
Leaving that aside, there's a difference between stealing the idea for a specific product and stealing everything, everywhere all at once IMO.
But the boiling the oceans part is the biggest for me. Why can't AI boosters ever engage with that one? Like seriously what is the plan?
@phranck I think it’s less a developer thing than a “Mastodon anti-capitalist” thing. This place attracts a certain kind of crowd.
It’s just the flip side of (and, to a degree, reaction to) the hype-infected hardcore rah-rah AI people.
Normal people are somewhere in the middle, seeing room for nuanced criticism without ignoring the benefits of the technology. These aren’t the people who get all the attention, though.
@phranck It’s about anshittification. AI rarely ever adds useful features that could not be added better without AI. But it’s a) a hype to add AI b) faster. It’s the tale of “good enough”.
If anyone without technical inside asks me what I think about AI, what I tell them is: “It’s great if it does not matter whether the result is correct. Want a artsy image? Great use case! Want to know the size of a screw? Nope, use calipers. The output of an AI should ALWAYS be assumed to be wrong.”
@phranck Yeah, that’s what our professors told us.
I studied computer sciences in the previous century.
Technology evolved faster than you think,
but way slower than you expect.
@phranck The problem is that AI is added as if it provided true value. Which it can (“artsy image fast”!) but if you work in an area where *CORRECT* is the most important property, or is one of the important properties, or matters at all, AI does not help. Proving the output of AI right is way to hard and way to simple to skip.
It’s a risk calculation: <Gain by AI>/<Risk of AI>
The gain seems high but is low (how often does being wrong not matter?).
The risk … is huge.
@teilweise It all depends on the models and how they were be trained. If the training base is "The Internet", well...
But if you feed it with curated and verified data, then it will be WAY better! I think for now we need more specialized models.
@phranck I’m still searching for a good tutorial on how to create a ML model from scratch.
Like: “Here’s a problem we want to solve. Here’s how to choose the number of layers and the types of layers. Here’s how to design the input layer and the output layer.”
The teaching (as you said: Use lots of verified data) and using part are well covered.
Also missing: What makes the test data “good” data?
~/phranck 
Uckermark MacGyver 
aburka 🫣
alpi
Jeff C. 🇺🇦🇬🇱
Patrick Gerken
Hannah
Voline
Der Teilweise
Phil Dennis-Jordan