Plugin pricing in the AI era

[quikquak]

AI - Know your code. It's about comprehension, not the speed of writing.

I found this guy quite interesting, "The tax we pay when building with AI" :

[ghettosynth]

AI - Know your code. It's about comprehension, not the speed of writing.

I found this guy quite interesting, "The tax we pay when building with AI" :

I'm not even going to watch the video. Comprehension debt isn't a new phrase, now, that is, we've been discussing this for a bit now, it is a real thing and anyone who builds with AI must be aware of it. However, building with AI does not imply acquisition of meaningful comprehension debt. The debt that you encounter is a function of your chosen SDLC.

What matters is the opportunity cost of skipped manual comprehension. It's not all equal, and that's what admits AI development as useful. Moreover, the need to know is a function of the exposure of your software, what is the blast radius.

[quikquak]

Watch it all, no skipping later parts, no TL;DW. There might be something you haven't thought of.
Or didn't want to hear.

Yeah, if you treat it as a programmer, not a tool, then you're stuffed!
I would never ship unknown code, it's bad enough using OS(whatever) library slop but that's supposed to work, right? I have used AI for small C++ classes, very useful - it made a small doppler effect for me, but it was incorrect, had bugs all over it, and it was not very well optimised, I ended up refactoring the whole thing - I should have built it myself from scratch - I was feeling lazy that day, but it ended up taking longer then expected.

But it did create a KD-tree algorithm for me, and nailed it. Here's a fun demo on the P5 .js editor, just press the play button and move the mouse about on the screen:
https://editor.p5js.org/QQquikquaki/sketches/QQrFFqunV
It also converted it to a C++ version for me, which was nearly perfect! Nearly.
I have also found it very useful for translating code, especially SIMD to NEON.

I'm not convinced yet though, and thankfully for us humans, it can't design anything new! LOL AI.

[ghettosynth]

I'm not convinced yet though, and thankfully for us humans, it can't design anything new! LOL AI.

Define new? It absolutely can create things that are not in the training data. That's a good enough definition of new. It has designed and built software tools for me that don't exist in that specific form, granted, with much supervision. It wasn't being a genius, it was amplifying my time in a way that tasks that would be extremely time consuming and that would have a low return on my intellectual investment became possible because my engagement with those tasks were reduced to testing the resulting output to spec and managing the overall design of the tool. What I didn't have to do was read the hundreds of thousands of lines of code of the interfaced target to understand how to work with it as a library of tools.

I watch very few videos today. This is a key area where AI is helpful, avoiding unnecessary linear consumption. What I find is that it gives me far more time for valuable linear consumption. Now that I've automated so many menial tasks, I have more time for both casual reading as well as deeper study.

[pekbro]

Definitely dependent upon your definition of new. They can not think, one thing to remember about that. Previously untried combinations of stuff sure…

[ghettosynth]

Definitely dependent upon your definition of new. They can not think, one thing to remember about that. Previously untried combinations of stuff sure…

Why do you want to "remember that"? It's a rejoinder to a position that no serious user has or cares about. A calculator can't think either, neither can an alpha-beta pruning tree, but both provide "intelligence", noun, that is perceived as reasoning by some humans in some way. When the video game beats you at chess, it's not thinking, but we still say that it is "thinking k moves ahead" depending on how deep we allow the tree to reduce available moves to a score. In a sense, to coin a popular LLM term, it's a category error. So, misapplying the terms reasoning and/or thinking because researchers might use them is missing the point that they are convenient labels that are understood in the same way. To even argue that LLMs can't think is a failure of language because we don't actually "know" how humans think. We have models, and we can try to replicate those models. What we can argue is that LLMs fail certain aspects of our models of human thinking in ways that demonstrate that they are not engaging information in the same way, or to the same scale, that humans do, and moreover, that the end result is a failure on specific tasks that we associate with "thinking."

It is unlikely that either scale or LLM architecture will change this. However, that is not the same thing as saying that we know what the view is from the top.

[pekbro]

Oh gawd, forget it dude sry I said anything. What they are doing is clearly defined if you are not an idiot.

[ghettosynth]

Oh gawd, forget it dude sry I said anything. What they are doing is clearly defined if you are not an idiot.