Claude Sonnet 4.5 can already replace most indians
Indian must be really bad then, because AI, even "muh claude" produces garbage.
>>106738592The plan-build-test-reflect loop is equally important when using an LLM to generate code, as anyone who's seriously used the tech knows: if you yolo your way through a build without thought, it will collapse in on itself quickly. But if you DO apply that loop, you get to spend much more time on the part I personally enjoy, architecting the build and testing the resultant experience.
>>106738664I have hard time believing, in a true professionnal context (your fizzbuzz isn't), you use it for more than boilerplaty stuffs (templating, unit tests generation or generic helper methods), because we are getting shoved this shit at my company and it quickly becomes a mess if you try more than boilerplat stuffs. I have hard time believing some companies release products that were mostly generated by AI.
>>106738744>because we are getting shoved this shit at my company and it quickly becomes a mess if you try more than boilerplat stuffs.which language?
>>106738787Mainly C#/Java/SQL/Typescript
>>106738592I'm at a FAANG with private access to Claude (it literally runs on our hardware) and for anything moderately complex that isn't clear cut or has exceptions it craps down badly
>>106738797>Typescriptwell gemini 2.5 pro made this in svelte i was just the guiding hand.https://prisonrpg.pages.dev/prisonRPG/
>>106738827i think AI is mostly about being able to steal code that has been written a thousand times, if you copy it from Github you have to care about the license, if Claude "generates" it, it's license free
fuck it all kinda suckswish I was born in 70s and became a professional programmer in the 90s. everything was simpler and cooler back then. and I'd be long senior by now
>>106738836Seems quite basic, to be frank. No backend (it looks clientside only), just some reactivity and states. I tried to make a simple forum with sveltekit/drizzle/postgre and it crapped out very fast. Also the code is shit, and it tends to use legacy syntax (easily fixed with an instruction file and context7 mcp).The code it produced was also shit (lack of modularity and all the basic stuffs), it creates a shitload of useless stuffs (redondant scripts) and overcomplexified things (I just wanted a little sql script to seed some data for testing an it ended up making a typescript script and some bash files to launch it, it is incapable to clean its shit. I've got to tell it was trying to make the simple forum only using AI prompts. I'd probably have done it twice as fast myself.
These models are still terrible for everything besides CS101 shit scraped from stack exchange.Because of how context works, they tend to generate God Objects and completely miss really obvious opportunities to separate concerns, use design patterns, etc. Not only is it unmaintainable by humans, but the complexity of objects piles on until even the AI (with context) can't edit the file without breaking something. I occasionally use Claude to write the implementation of a function, but I don't let it design anything. It's really bad.Since most "good" code/architecture lives on private/enterprise repos and isn't public, I think that LLMs are going to struggle to produce complex novel software for a long time. The training data just isn't there.
Coding with Claude/LLMs is like mentoring a junior engineer with a learning disability, forever.
>>106738987>Since most "good" code/architecture lives on private/enterprise repos and isn't public, I think that LLMs are going to struggle to produce complex novel software for a long time. The training data just isn't there.all private repos in github or gitlab have been scrapped
>>106738592Tried it on my large production codebase, ca. 800k lines of code, including tests etc.It's good. Like, really good.
>>106739415proof for>all
Claude still can't even get out of Celadon.
>>106738664bro what's your stack/setup/philosophy/wtv. i tried this and it just fucking sucks cock, it struggles to actually solve the problem no matter how hard i try with details and different ideas.for sake of clarity, and this may identify the problem immediately, i've mainly used Codex via it's online interface & github to write stuff and help architect when it can, but i've also tried Cursor w/ gpt-5 with maybe a bit better success.
>>106738592>shareholders find more ways to raise profits
>>106738613well... yeah
>>106738971>I tried to make a simple forum with sveltekit/drizzle/postgre and it crapped out very fast. Also the code is shit,meh i dont believe you more likely youre shit>I'd probably have done it twice as fast myself.lol
>>106738917> everything was simplerThese days, almost any software (and hardware) info you might want is at your fingertips. Same goes for non-computer info you might need.In the 1990s, you'd typically have to acquire a physical book, or else talk to someone who knew the info you were after.> and coolerHaving to find something else to occupy yourself with while the machine was compiling certainly had its advantages, but could also get tedious sometimes.
>>106741193>good morning, aydsharrrrr
>>106740867
You are going back to kolkota, ranjeet
has anyone thrown these bots at reverse engineering tasks? For example I want to decompile some psp game would they be able to do it?
>>106738592>less than 2 points improvement from sonnet 4 to opus 4.1 and less than 3 from opus 4.1 to sonnet 4.5these are dogshit results. you will never notice the difference. was it worth it to spend millions in such a small improvement? AI has really reached a wall, huh?
>>106745926it's also always against the same problems, so it's overfitted...
>>106738744That sounds like a skill issue honestly. AI is a tool after all and not being able to use it correctly just means you can't adapt and that you'll eventually be replaced.>I have hard time believing some companies release products that were mostly generated by AI.We do at least. I'm a senior software engineer and honestly of the code I've pushed these past few months, about 90% was AI generated. Learn to prompt.
>>106738613You have never worked with indian software devs straight from school.
>>106738971Skill issue
>>106743849In my experience they’re great for webdev, reverse engineering? You could try it, but doubtful.
>>106743849Someone just did to port apple system7 to x86Look it up on hackaday
>>106741193The funniest part about this study is that it's probably about how only 5% of them are top 5% in some coding competition
>>106738592Claude literally keeps giving me unusable slop and when I complain it insists it should work and then sends the slop again lmaoI'm getting 'works on my machine''d by a fucking aic
>>106738613>>106738592Qwen3-coder-30b can replace just about most indians. Not really an achievement.
>>106745926You clearly don’t know how these numbers work, the closer you are to 100 the exponentially more accurate you are, there is a huge qualitative difference between 39>49 vs 89>99 for example
