delete and do over. all good. jr devs same, less hurt feelings when code is delete. 😁
i dont know yet how people start from scratch, they are prompt wizards. i think its a lot of describing the architecture in an outline, feeding it lots of context and possibly creating that initial app skeleton, pick the libraries etc.
If you start from scratch, then you have to do all of the work you normally have to do when writing manual code, but you still end up with crappier code. I don't get the point, at all.
I think the opposite: the best thing it's good for is setting up the skeleton, arranging the boxes on the page, importing all of the libaries, etc. Basically, like a framework.
It's the programming/logic part, that it's really bad at.
yup, it's a glorified scaffolding tool, really inefficient at that
i can build scaffolding code generators in a few hours, with no bugs
I'm going to start over, but refer to this one, and reimplement everything, step by step. Faster than googling StackOverflow, at any rate. 😅
this is increasingly making me realize, this is the moment if i dont learn to prompt like the current CS students do.. out to pasture w me 😂
Yeah, gotta move with the times. 🤷♀️
This gets you a good prototype to show customers and can help you out, in a pinch, and find bugs.
This actually went quickly because I told to use the Alexandria repo as a code base. Everytime it got stuck, I told it to go look at Alexandria, again, and that got it pretty far. Would probably have to feed it more Kind 01 examples, to get it moving from here, but I need to clean that up, first.
i'm not wasting my time learning it until at least the code generators don't mangle the types in my Go code, every. single. time. doesn't even pass the lexical analysis step.
probably the hype about AI and code is based on the fact it can cope with the retardedly simple type semantics of javascript, which are basically nonexistent
ironically, it is easier for humans to reason about and more efficient for compilers to use strict typing, there is no compiler that outperforms Go on the basis of lines of code
I don't know if it has enough Go examples, yet. Typescript and PHP are more ubiquitous, so those are easier to AI.
I wouldn't AI code in C, either. Would be hilarious, tho. 😂
i think that they are optimizing for data size and compute time and more depth of causal chains means exponentially bigger models are required
probably you can make it work but you have to allow for 10-100x as much data in the model and about the same difference in computation, and my guess is that renders the efficiency gain null
and i don't think it's gonna be that much different in C, Rust or C++, they are all very strongly typed and have deep semantics with side effects that you don't see with the little scripts of javascript and python code
The bloat and redundancy is incredible. And just the nonsense and the ded code...
ded internet is gonna drive the peoples to the nostr
ded code is the open secret about AI code generation that the gigacorps like M$ and NVDA are not gonna include their marketing and all the suckers will buy it and all the managers will push to use it and eventually the actual decrease in productivity will be revealed and suddenly there will be a market crash because the NASDAQ right now is holding up only on that hot air
Yeah, it makes the whole thing run noticeably slower. Click on a button in Alexandria and it immediately does something. Click on a similar button in this thing and it's like, "Yo, hold up. I need to go get myself a coffee before I show you that note. Like chill. Take a seat."
Spinner spinning... 😏
Explains why relays are so fast, but so many clients are really slow.
They kept saying it's the relays, but nostr:npub1wqfzz2p880wq0tumuae9lfwyhs8uz35xd0kr34zrvrwyh3kvrzuskcqsyn and I were like
anyone who claims that AI helps them write software, except for summarising documention, is lying, either for pay or because tehy don't want to seem like a grumpy kitty
It's also true that connecting to one relay is lightning fast, but waiting for data streams to resolve across several relays can quickly slow things down.
and it is kinda worse when each realy is basically sending the same data too
Yeah a good SDK ought to be able to optimize those issues away.
I've used Claude a bit to help me explore and learn C++ language features, but I don't trust it to write C++ code as much as I trust it to write TS or Python.
yes having it go read docs, and other repos, even local checkouts.. thats kinda what ive been doing. it really likes the markdown in the nips for example..
Great for writing in-line documentation, too.
This is the point, where the vibe coders usually give up and phone a dev friend, but I'm the friend, so I'll just take it from here.
If you came in with an architecture/component diagram and a clear set of requirements, you could probably get pretty far from scratch just by prompting. In that case, you've already frontloaded the design and code organization work, and now you're just handing that design to a junior dev (Claude) and saying "go do this."
But then you're just doing all of the hard work and the AI gets to do the fun ShadowySupercoding stuff
😭
Yeah pretty much
for me the time savings is if it can boiler plate what ive done, but with small tweaks. generally im using too new of stuff for its training so i have to keep a tight leash but for some things when i know its trained or has enough context i can open up the throttle 😁 anyway, its fun and exciting
Yeah the LLMs don't know a ton about Nostr yet. It's good to be on the bleeding edge.
Can't wait for the #GitRepublic LLM, based upon Aedile. Wish I could use it, today. 😭
Someone who can code and test, who uses this will be a Supercoder. Someone who can't code or test, who uses this will be a Vibecoder.
Dynamic doesn't change, everyone just ups one level.
yes, in my experiences, it has a lot better chance of working when you have a solid, hand-written base that it can infer from.. i like to give it feature work sometimes and see what it can do. its hit or miss, sometimes its really impressive and saves me lots of time, other times it majorly gets in the way and is waste of time.
eventually you get a feel for what it has a chance of doing vs. not. and how to prompt it so it has better chances.
