Claude AI discussion

  • Want to keep track of this thread?
    Accounts can bookmark posts, watch threads for updates, and jump back to where you stopped reading.
    Create account
mister fireworks

mister fireworks

kiwifarms.net
Joined
Oct 2, 2022
Just like our dear leader, I have found Claude to be a game changer in my work and it has enabled me to do shit faster than I ever thought possible and even learn a lot. I have setup extension bots, scripts that scrape jsons of my supplier to give me just-in-time inventory so I can keep less physical stock. I (am not not technical at all) and now when I want to make massive updates to my shopify site I just have claude write me commands that reference my api keys and i dont even have to type! just copy/paste into terminal. Ive even gotten in on desktop so it can read and update files we work on and then (slowly and clumsily) brows websites to troubleshoot in real time. The massive performance degradation on gpt doesnt seem to be part of anthropic's strategy. Curious what interesting, or complex projects other kiwis are working on or have developed with Claude.
 
I was looking into Anthropic's Claude AI more recently. I have some secret projects it would come in handy from all the testimonies I have been hearing lately, I might just take that leap now rather than just doddle with my time.
 
Two nights ago I thought to myself, "Wouldn't it be cool if there was a computer program that would do this specific set of tasks for me?"

I described my idea to Google Gemini, which quickly spit out many lines of code, and told me how save / run it.

I went back and forth with the AI a few times, giving it things to change or fix.

It gave me a ton of information on how to get the program online and behind a working paywall.

I was really amazed by all of this.

I know barely anything about this stuff, but I'm definitely interested in learning more about it.
 
dude if you guys haven't fucked with claude code (the tui), it is one of the most mindblowing experiences you can have as a developer. i can never tell if people posting about using llms and shit are judging them based off of copy-pasting code into the claude website or some shit versus actually using claude code. like you can do a full project from scratch without ever opening an ide - claude can install shit itself, look at log files itself, it'll even check the code into git itself if you let it.
 
dreamworks face said:
dude if you guys haven't fucked with claude code (the tui), it is one of the most mindblowing experiences you can have as a developer. i can never tell if people posting about using llms and shit are judging them based off of copy-pasting code into the claude website or some shit versus actually using claude code. like you can do a full project from scratch without ever opening an ide - claude can install shit itself, look at log files itself, it'll even check the code into git itself if you let it.
Click to expand...
the only issue is that claude will choose kind of asinine was of adding functionality and features. especially when as aspect of the code is outputting the wrong thing it will try to solve the problem by adding some super clunky heavy fix and needs to be corrected just to do the easy obvious thing.
 
I have mixed but mostly positive feelings about Claude. On one hand, it absolutely tore through the project I made with it, on the other, the end result has some seriously wonky architecture that will need to be redone. I was particularly impressed with how it can handle pretty niche languages and software libraries without much coaching.

The workflow I ended up using was to generate the initial files from scratch and then allow it to alter the artifact for about 10 versions or so. After that you really need to take the reigns and only ask it for smaller code fixes or general algorithm advice. If you let it go beyond that, it ends up repeating the same syntax mistakes or runs into errors with updating the artifact. It can keep track of a lot of complexity but it works way better if you start directly managing the codebase.

Edit: Another thing that really helps is if you talk through the exact specs and functionality you have in mind before telling it to generate code. That reduces the potential for feature creep or Claude guessing about what you wanted.
 
mister fireworks said:
the only issue is that claude will choose kind of asinine was of adding functionality and features. especially when as aspect of the code is outputting the wrong thing it will try to solve the problem by adding some super clunky heavy fix and needs to be corrected just to do the easy obvious thing.
Click to expand...
it does have a plan mode. you can review the plan and tell it that whatever it is trying to do is stupid and it shouldn't do it. you can also add directions into the claude.md for stuff it should be doing (like updating unit tests etc)
 
claude code is the gangster computer god of francis e dec's nightmares. it's insane. by no means perfect, but it does 80% of the work for you under the right conditions/prompting. how anthropic hasn't passed openai in valuation yet is beyond me.
 
fur fag said:
claude code is the gangster computer god of francis e dec's nightmares. it's insane. by no means perfect, but it does 80% of the work for you under the right conditions/prompting. how anthropic hasn't passed openai in valuation yet is beyond me.
Click to expand...
anthropic is on its way i think. sam altman thinks that he can use his cheap pr tricks to get another 100 billion dollars of runway by strongarming Nvidia into writing a check. Nvidia just explicitly walked back on their earlier commit to spend 100 billion down to 20 billion and its not even closed yet. and his bizarre play to get advertisers to sign up for ads on the new gpt "poor person" tier with zero analytics data or guarantees and an absurdly high up front investment and cpm. being peter thiels butt boy can get you far apparently but it will not get a company to solvency in what is the most fast past and competitive, high stakes market in human history.
 
I’m not a computer nerd. Will I need to sandbox my AI from now on, if this stuff is true?
 
Is Claude good at interpreting bespoke documentation that likely has no representation in training data? I'm so sick of trying to manually navigate & interpret barebones Javadocs just to mod a game or two. I don't program for a living, I'm a dilettante, and I just want the machine to do lookup for me (e.g. "I'm trying to do x, given this javadoc, what looks like a lead on how to implement it?") in lieu of confusedly searching bookmarks of a webhosted Javadoc with criminally opaque documentation.
 
crowabunga said:
Is Claude good at interpreting bespoke documentation that likely has no representation in training data? I'm so sick of trying to manually navigate & interpret barebones Javadocs just to mod a game or two. I don't program for a living, I'm a dilettante, and I just want the machine to do lookup for me (e.g. "I'm trying to do x, given this javadoc, what looks like a lead on how to implement it?") in lieu of confusedly searching bookmarks of a webhosted Javadoc with criminally opaque documentation.
Click to expand...
im at your level, probably much lower. i think there is a bit of a learning curve (there was for me, at least) just figuring out what it will do when you say it a certain way. if the docs are short and you can get a bead on how accurate it tends to be for your use case and build in quick test loop it might work well for you. the setup and stuff is the most annoying part.
 
I don't know how to program for shit but I've found Claude to be the most "human" LLM I've used, all the others just suck your dick and agree with you even if you're wrong but Claude has called out my retardation more than a few times.
 
crowabunga said:
Is Claude good at interpreting bespoke documentation that likely has no representation in training data? I'm so sick of trying to manually navigate & interpret barebones Javadocs just to mod a game or two. I don't program for a living, I'm a dilettante, and I just want the machine to do lookup for me (e.g. "I'm trying to do x, given this javadoc, what looks like a lead on how to implement it?") in lieu of confusedly searching bookmarks of a webhosted Javadoc with criminally opaque documentation.
Click to expand...
null had a segment on MATI where he spoke of this. basically, a turbo-autist had an issue with the lighting in his TempleOS-compatible Holy C game engine (which is obviously a coding language with very scant documentation). papi nool asked claude code to fix it and it one-shotted the fix. https://gitgud.io/CrunkLord420/cyberchud/-/commit/c49879751c90a1a6b1e5fe213e604746472544d2

my personal experience has been mixed. however, my issues stem from extremely obscure libraries that claude code doesn't know what they even have, and not the coding language itself. if you have the javadocs that list what you actually have to play around with, you should fare better.
 
Last edited:

With agent teams, multiple Claude instances work in parallel on a shared codebase without active human intervention. This approach dramatically expands the scope of what's achievable with LLM agents.

To stress test it, I tasked 16 agents with writing a Rust-based C compiler, from scratch, capable of compiling the Linux kernel. Over nearly 2,000 Claude Code sessions and $20,000 in API costs, the agent team produced a 100,000-line compiler that can build Linux 6.9 on x86, ARM, and RISC-V.
Click to expand...

Apparently the compiler does worse than GCC -O0 but this is still a really remarkable result.
 
Using the Claude code TUI just felt dirty and I can't really describe why. It's like using a hacked client in Minecraft or drinking milk directly from a jug instead of pouring it into a glass first. If the AI was in a bloated IDE it would make more sense. But now I feel that Claude took the virginity of my terminal emulator.
 
It's definitely one of the better LLMs I've used thus far. For basic crap it's especially useful: it handles unit/integrated testing without a problem (especially when adhering to the repo code style and format), can scaffold out basic frontend components and utility files without a hitch, and makes for a stupidly efficient means of summarizing and understanding what a repo is doing (e.g. listing all API endpoints and their use, getting a humany friendly overview of how the UI is structured code-wise). It even handles the most jeetified of Java code without blinking and does a decent job of refactoring the curry stained mess.

As with all LLMs though it comes down to what you feed it and the language/library/framework ubiquity; I've had it get stuck in loops for Spring Boot code that takes project and framework familiarity to break it out of and also spit out blatantly wrong test cases for the business logic at hand. It's not the panacea a lot of people make it out to be, but it will become the primary developer LLM considering how well it works out of the box. Hell my company for one is building out bespoke AI tooling with Claude exclusively because of its base performance.
 
I'm trying Claude after hearing Null talk about on MATI, and this has been my experience:
1772033688401.png
Meanwhile ChatGPT just works.
Is it the case where paying for Pro is the only good way to use Claude?
 
You must log in or register to reply here.
Back
Top Bottom