An irrelevant but interesting take is that this applies as an analogue to a lot of stuff in electronics related space.
It is harder to receive data than to transmit it, because you need to do things like:
match your receiver’s frequency with that of the transmission (which might be minutely different from the agreed upon frequency), to understand it
know how long the data will be, before feeding into digital variables, or you might combine multiple messages or leave out some stuff without realising
this gets even harder when it is wireless, because now, you have noise, which is often, valid communication among other devices
Getting back to code, you now need to get in the same “wavelength” as the one who wrote the code, at the time they wrote the code.
I weirdly love reading code and figuring out what it’s doing. Debugging is cathartic.
It might take a while and I might be cussing up a storm saying, wtf is this shit? Why the fuck would you do it this way? Why the fuck did you make this convoluted for no reason?
Right now it’s unfucking some vibe coded bs where instead of just fixing an API to get the info we needed accurately, it’s trying to infer it from other data. Like, there is a super direct and simple route, but instead there are hundreds of lines to work around hitting the wrong endpoint and getting data missing the details we need.
Plus letting the vibe add so much that is literally never used, was never needed, and on top of that returns incorrect information.
So? Some of the people pushing out ai slop would be perfectly capable of writing their own llm out of widely available free tools. Contrary to popular belief, they are not complex pieces of software, just extremely data hungry. Does not mean they magically understand the code output by the llm when it spits out something.
given that expert systems are pretty much just a big ball of if-then statements, then he might be considered to have written the app. Just with way more extra steps.
Yeah but the people who made it like that probably understand whether to trust it to write code or not. The AI Tony wrote, he knows what it does best and he trusts it to write his code. Just because it’s AI doesn’t mean it’s LLM. Like I trust the errors compilers give me even if I didn’t write them because it’s good. And I trust my scripts to do things that I wrote them for, specifically since I tested them. Same with the AI you yourself made, you’d test it, and you’d know the design principles.
an ai is not a script. You can know what a script does. neural networks don’t work that way. You train them, and hope you picked the right dataset for it to hopefully learn what you want it to learn. You can’t test it. You can know that it works sometimes but you also know that it will also not work sometimes and there’sjacksjit you can do about it. A couple of gigabytes of floating point numbers is not decipherable to anyone.
The point is that no vibe coder could design an LLM without an LLM already existing. The math and tech behind machine learning is incredible, whatever you may think. Just because we can spin up new ones at will doesn’t mean we ever could have skipped ahead and built Jarvis in 2008, even if all of society was trying to do so - because they were trying.
In the fictional universe where a human could singlehandedly invent one from scratch in 2008 with 3D image generation and voice functionality that still exceeds modern tech… yeah, that person and their fictional AI wouldn’t necessarily be producing slop.
the “target” is to get useful software out. The ai is the tool. In this example, the ai is the gun. It is the tool used to achieve the goal.
Anyone can make an improvised hammer. Stick a rock or a piece of metal on a stick. But that doesn’t make them carpenters, even though they made their own tools.
if they build software using mainly ai generated code, then they are a bad coder
I’ve tried this on personal projects, but not work projects.
My verdict:
To be a good vibe coder, one must first be a good coder.
Vibe coding is faster to draft up and POC, longer to debug and polish. Not as much time savings as one might think.
exactly, you can only really verify the code if you were capable of writing it in the first place.
And it’s an old well known fact that reading code is much harder than writing it.
An irrelevant but interesting take is that this applies as an analogue to a lot of stuff in electronics related space.
Getting back to code, you now need to get in the same “wavelength” as the one who wrote the code, at the time they wrote the code.
i like the analogy
I weirdly love reading code and figuring out what it’s doing. Debugging is cathartic.
It might take a while and I might be cussing up a storm saying, wtf is this shit? Why the fuck would you do it this way? Why the fuck did you make this convoluted for no reason?
Right now it’s unfucking some vibe coded bs where instead of just fixing an API to get the info we needed accurately, it’s trying to infer it from other data. Like, there is a super direct and simple route, but instead there are hundreds of lines to work around hitting the wrong endpoint and getting data missing the details we need.
Plus letting the vibe add so much that is literally never used, was never needed, and on top of that returns incorrect information.
enjoying it is a different issue. You probably enjoy it because it’s more difficult, which is perfectly valid reasoning
Exactly how I feel about it as well.
Even if you’re the one that built, programmed, and trained the AI when nothing else like it existed?
So? Some of the people pushing out ai slop would be perfectly capable of writing their own llm out of widely available free tools. Contrary to popular belief, they are not complex pieces of software, just extremely data hungry. Does not mean they magically understand the code output by the llm when it spits out something.
Stark would have developed their own way of training their AI. It wouldn’t be an LLM in the first place.
and he stil wouldn’t understand its output. Because as we clearly see, he doesn’t even try to look at it.
What if it were just a really big Expert system?
That’s usually the thing that you call AI players or COM players in computer games.
given that expert systems are pretty much just a big ball of if-then statements, then he might be considered to have written the app. Just with way more extra steps.
I think you missed the part where “nothing like it even existed”
so? someone invented current llms too. Nothing like them existed before either. If they vibe coded with them they’d still be producing slop.
Coding an llm is very very easy. What’s not easy is having all the data, hardware and cash to train it.
Yeah but the people who made it like that probably understand whether to trust it to write code or not. The AI Tony wrote, he knows what it does best and he trusts it to write his code. Just because it’s AI doesn’t mean it’s LLM. Like I trust the errors compilers give me even if I didn’t write them because it’s good. And I trust my scripts to do things that I wrote them for, specifically since I tested them. Same with the AI you yourself made, you’d test it, and you’d know the design principles.
an ai is not a script. You can know what a script does. neural networks don’t work that way. You train them, and hope you picked the right dataset for it to hopefully learn what you want it to learn. You can’t test it. You can know that it works sometimes but you also know that it will also not work sometimes and there’sjacksjit you can do about it. A couple of gigabytes of floating point numbers is not decipherable to anyone.
The point is that no vibe coder could design an LLM without an LLM already existing. The math and tech behind machine learning is incredible, whatever you may think. Just because we can spin up new ones at will doesn’t mean we ever could have skipped ahead and built Jarvis in 2008, even if all of society was trying to do so - because they were trying.
In the fictional universe where a human could singlehandedly invent one from scratch in 2008 with 3D image generation and voice functionality that still exceeds modern tech… yeah, that person and their fictional AI wouldn’t necessarily be producing slop.
Even if they build the AI doing it from scratch, all by themselves?
yes. Because that would still mean they didn’t code the app.
“killing is bad!” “but what if the murderer 3d printed his own gun?”
More like: “killing is bad” “but what if the ‘murderer’ designed, build and produced their own target?”
You can’t kill a robot, so it isn’t killing.
the “target” is to get useful software out. The ai is the tool. In this example, the ai is the gun. It is the tool used to achieve the goal.
Anyone can make an improvised hammer. Stick a rock or a piece of metal on a stick. But that doesn’t make them carpenters, even though they made their own tools.