Vuk Rosić

Research on training an LLM with DeepSeek & Kimi architecture

Project to research, optimize and train our LLM - Blueberry

Full course on becoming AI researcher from scratch

No description.

No description.

LLM that can be trained on 1 or more GPUs for research.

Leo optimizer, variation of Muon that runs faster

Become GPU kernel engineer step by step.

Use Muon optimizer instead of AdamW.

NextJS-14 ChatGPT clone.

Ultimate AI research and engineering course

No description.

Push open-source AI research to the frontier by the end of 2026 - open frontier science for everybody.

No description.

NextJS, React, Prisma, Clerk WhatsApp Clone for YouTube Tutorial

Auto AI research tools

No description.

Setting up a NextJS + Convex + ShadCN + Stripe + Tailwind CSS + Typescript + React project

Development of New Neural Network Optimizers

One-command installer for Deli Chen's AutoResearch skill — turn a coding agent into a long-horizon autonomous researcher.

No description.

Paper and experiments for reasoning by repeatedly refining a latent state.

Build LLM & write paper from scratch

Updated edition of full course on becoming AI researcher

No description.

Mixture of Experts LLM implementation

python-to-triton-llm

Analyzing what patterns text forms when converted in raw bytes - using persistent homology

No description.

Analysis of Muon optimizer in modern Large Language Models.

Learn tiny reccursive models

No description.

Analysis and Design of Novel Optimizers for Neural Networks

hybrid-llm-mamba-transformer-research

No description.

No description.

Code & Train Qwen3 From Scratch

Researching Muon optimizer in LLM pretraining

No description.

LLMs and VLMs that take in bits or bytes directly without tokenization.

No description.

Build decoder-only transformer (LLM) from scratch

Code and training for pretraing of an agentic LLM

Course on building Claude Code from scratch

No description.

Code Kimi K2 From Scratch

No description.

No description.

No description.

Pytorch from scratch course

Research on muon optimizer in LLM pretraining.

Code & Train GPT-5 small version - for learning.

No description.

hybrid-llm-mamba-transformer-research-2

No description.

Tools to build and release your own LLM

Small byte level AI model takes and outputs text and images, research

No description.

No description.

Telegram Bot connected to OpenAI API, GPT-4-Vision, Dall-E-3

Send multiple images to telegram bot that will process it with GPT-4 vision.

Materials I teach in my course.

LLM research, v2

Attention mechanism research

Learn to do small LLM research

muon-newton-shultz-research

Reproduction and research code for QK Norm rank collapse analysis in LLMs

No description.

No description.

Clone of WhatsApp with Next14, Prisma, Clerk

No description.

Blueberry LLM version 2

Byte-level LLM research

Research on FP4 LLM pretraining

Research on text converted to raw bytes, data distribution

LLM research experiments 3

No description.

Train your own reasoning LLM in 20 minutes — GRPO RL post-training on a $0.10/hr GPU (Unsloth + TRL + vLLM). Raw base model 44.5% → 73.5% on GSM8K for ~$0.05.

From-scratch reproduction of SuperBPE (Liu et al., COLM 2025) on a laptop, $0 — two-phase superword BPE + encoding-efficiency sweep.

No description.

No description.

No description.

Nvidia's paper Pretraining Large Language Models with NVFP4

No description.

hybrid-llm-mamba-transformer-research-3

Ecommerce stores and product fontend designs in NextJS / React.

Fullstack app from only nextjs + stripe

No description.

This is built only with nextjs as frontend and stripe as backend.

No description.

Blueberry LLM

Code & Train Grok 4 From Scratch

Code & Train Anthropic Claude style LLM from scratch.

Train best LLM possible on 1 x NVIDIA RTX 4090 GPU

Linear, Mamba and Hybrid attention analysis of MoE LLMs

Official PyTorch Implementation of "Diffusion Transformers with Representation Autoencoders"

No description.

Generating images with diffusion

Research on RoPE on a small LLM

No description.

LLM RoPE improveemnts

Speeding up the reinforcement-learning rollout loop with speculative decoding.

Build lossless speculative decoding + DeepSeek's DSpark drafter from scratch (torch, CPU). Companion code for the from-scratch video.

Turns natural date phrases (next friday, third thursday of march) into ISO-8601 dates. A ~1M-param byte-level transformer trained 100% on code-generated data, with an honest map of where it breaks.

No description.

No description.

Cuda kernels and other optimizations for OmniDrones.

Research tools

No description.

No description.

No description.

No description.

Representation autoencoders research

Tutorial notebooks from my channel - https://www.youtube.com/channel/UC7XJj9pv_11a11FUxCMz15g

No description.

Cinnamon - Open Source Alternative To OpenAI's Reasoning Models

Place text you send to this API on a black square image. You can also enable Dall-E-3 as well. Deploy this to your Vercel.

No description.

No description.

No description.

3D product card for NextJS 14.

No description.

Select multiple images and prompt and get OpenAI API GPT Vision response.

Notion.so clone - Next.js, Tailwind, Convex, React

No description.

CodingTeacherGPT

AI Research blog and posts

DeepSeek OCR

No description.

No description.

🚀 Efficient implementations of state-of-the-art linear attention models

No description.

No description.

Automated AI researcher

Mini experiment on LLM scaling laws for learning

Model that takes and outputs raw bytes. Image and text.

Fast diffusion image generation

Future of AI is raw bytes as input and output. We research it here.

Transformer architecture research

Neural network data precision comparison

LLM attention mechanism

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

No description.

Research on Blueberry LLM with PaTH attention.

Random AI research I do

This is AI generated AI research

Random AI research tutorials

Study of activations functions.

Comparing different activation functions in attention and feed forward.

NanoGPT (124M) in 3 minutes

Open benchmarks for maximizing open-LLM inference throughput on real GPUs — vLLM tuning (quant, CUDA graphs, speculative decoding) per hardware tier, with reproducible serve commands.

Load a pretrained LLM base and RL-train it to reason with GRPO (Unsloth+TRL+vLLM), rewarding correct step-by-step answers on GSM8K.

Predicting missing parts of the input in representation space instead of predicting exact tokens.

voidlab research thread: long-context + agentic test-time compute — fork of llm-research-kit

voidlab research thread: continual learning — fork of llm-research-kit

voidlab research thread: weight-tied recurrent depth (latent reasoning) — fork of llm-research-kit

No description.

Fixes writing errors a spell-checker can't see — their/there/they're, its/it's, then/than. A ~1M-param byte-level transformer. Ships a model-vs-script benchmark + technical report.

Pronounces words English spelling can't (though->DH OW, tough->T AH F) — a ~1M-param byte-level transformer that learns grapheme-to-phoneme and generalises to unseen words: 72.1% exact vs 8.4% for a per-letter script.

Corrects a misspelled word to its intended word (recieve -> receive) — a ~1M-param byte-level transformer that beats nearest-dictionary AND a frequency dictionary; trained 100% on code-generated data.

Restores the vowels to a vowel-less word (wrld → world, brd → bread) — the inverse of a lossy map that no rule can compute. 84.7% vs a naive script's 68.7%. A ~1M-param byte-level transformer trained 100% on code-generated data.

Converts messy separator-less identifiers (sdkmodel, HTTPREQUESTHANDLER) into any case style — the slice a regex provably can't split (99.7% vs 8.2%). A ~1M-param byte-level transformer trained 100% on code-generated data.

Make GitHub repositories safe and legible for coding agents.

No description.

Your own personal AI assistant. Any OS. Any Platform. The lobster way. 🦞

No description.

No description.

Course on plastics from scratch

This code implements the algorithm of FIPO, a value-free RL recipe for eliciting deeper reasoning from a clean base model.

PyTorch code and models for VJEPA2 self-supervised learning from video.

text and image to video generation: CogVideoX (2024) and CogVideo (ICLR 2023)

Copy paste prompt to auto optimizer your AI model training (auto research)

Researching weight decay + Muon optimizer in LLM pretraining

An open source library designed to provide community examples of Joint Embedding Predictive Architectures (JEPAs). It contains code and examples for learning representations from images, video, and action-conditioned video, as well as planning using JEPA-based models.

No description.

No description.

No description.

No description.

Learn to implement and train JiT https://arxiv.org/abs/2511.13720

QeRL enables RL for 32B LLMs on a single H100 GPU.

No description.

No description.

Bento grid product layout from Aceternity UI in NextJS 14.

Affiliate marketing platform

No description.

No description.

Automated set of tools to help you upload video to YouTube faster.

I Vibe Coded Some Muon Optimizer Research: A Comprehensive Analysis of Newton-Schulz Orthogonalization in Modern Optimization

Deep dive into LLM reasoning and reinforcement learning

Small research on muon optimizer.

AI Research Paper Generator for Ideation and Prototyping

Inspect Muon optimizer

Build a Jekyll blog in minutes, without touching the command line.