Pranjwal Jha

I am 15, I don't play any games. I have maybe played minecraft 3 times in my life at a friends house. My totally accumlative gaming hours are probably less then 5 hours my whole life. My fun is in coding, using ai to code and building stuff

I miss my phone :cry:

@JeroenvanH @ostrisai @krea_ai loss means nothing with a batch size of 1, just run it for 100 steps/image

@ostrisai @krea_ai I'm trying this now, but getting nothing but this, with this loss graph. Obviously my settings are completely wrong here.

I made a training adapter for @krea_ai Krea2 Turbo that will allow you to train a LoRA on the turbo model directly without breaking down the distillation. This works the same way as my Z-Image turbo adapter.

Such a great model and app congratulations

@shreyas_ar54005 @levelsio It's not cheap you dumbass bot

@levelsio Infrastructure got so cheap that even abuse costs less than a dinner. The real question is why scraping defense isn't a default feature on every hosting platform.

The fact that 100,000s requests cost you $20 is more worrrying than Tencent scraping

Great things are happening with Google gemini, at this point delete the gemini app and it's codebase

@JustinLin610 Wishing you luck in whatever you

me stepping down. bye my beloved qwen.

@Alibaba_Qwen 😄 awesome

🚀 Introducing Qwen-Image-2.0 — our next-gen image generation model! 🎨 Your imagination, unleashed. ✨ Type a paragraph → get a pro slides ✨ Describe a scene → get photoreal 2K magic ✨ Add text → it just works (no more glitchy letters!) ✨ Key upgrades: ✅ Professional typography (1K-token prompts for slides, posters & comics) ✅ 2K native resolution with stunning detail ✅ Flawless text rendering + unified generation/editing ✅ Lighter architecture = faster inference Try it now → chat.qwen.ai/?inputFeature=… Full details → qwen.ai/blog?id=qwen-i…

Microsoft is such a cunt ass company with zero morals

Why do I both making something i should just use perplexity to search fuck my chud ass life

wow what goes on web stays on web, single perplexity tab using 50% of cpu and 1 gb of ram 😂 apparently the model hallucinated forever or something

😂😂😂😂

@dopdopinder @TheLBOguy @ThePrimeagen @beaversteever I mean they probably want to test the most basic skills i suppose if a person can remember where to use .front and where .top It's kind of asking kids not to use a calculator in primary school

@TheLBOguy @CatOrange4185 @ThePrimeagen @beaversteever Then why the fixation to point out syntactical errors? "Did you forget to import a library?" Of course I did, all IDEs automatically add them since ages. Thinking can be tested with psuedo-code or even flowcharts, if you are evaluating thought process. But no, write this document

"after 11 technical interviews" "your work has been accepted into our codebase" interviews are getting out of hand

@dopdopinder @ThePrimeagen @beaversteever apparently people who 'code' using AI aren't really the best to hire, who would've thought. And ai first is more of a gimmick than functionality.

@ThePrimeagen @beaversteever Why does an "AI-first" company want people to not use AI or even an IDE in 2025? What sick obsession is that with paper/whiteboard/word processor coding? What exactly does that even test?

@codewithimanshu @iamgrigorev Bot

@iamgrigorev George, that's a brilliant question! Parallelism in PyTorch is crucial for scaling, especially with those frameworks, right?

Curious how frameworks like nanochat actually scale? New blog post: Introduction to Parallelism in PyTorch. Covers async DDP, ZeRO-1/2, FSDP, and TP – with implementations from scratch and practical advice from real runs on different hardware. Even if you are experienced, you’ll likely find something new 👇

All this for a mid ass ui 💔

@VTchuiev @prathamgrv adhd brain can't function without leetcode style problems

@prathamgrv ...why not just open vscode and start writing the python code?

Introducing TensorTonic. A platform to actually learn ML by building it. > Practice ML algorithms in LC style > Replicate cutting-edge ML papers > Prepare with company-wise AI interview questions + blogs. v1 is out. accelerating hard. website link in comments.

Just removed conda, I'm all in on uv now

back to this 🥲 I like zed ui more than nvim sort of

@tobyordoxford Hf solves this by not normalising each advantage term by response length but by total completion tokens I believe

I find it fascinating that when DeepSeek used reinforcement learning to train an LLM to reason better, its chain of thought grew linearly (an equal amount per step of training) — and that amount appears to be exactly 1 token longer per step of RL: