Autonomous AI research swarm — runs ~100 overnight LLM training experiments unattended. Fork of karpathy/autoresearch with Digital Cognitive Labor routing.

Train your own ChatGPT on Apple Silicon — MLX port of nanochat

Train Llama 3 models from scratch. Any scale, any personality. By Arianna Method.

A minimal, hackable Vision-Language Model built on Karpathy’s nanochat — add image understanding and multimodal chat for under $200 in compute.

A user's facts become a few rows in a content-addressed memory table — not a rewrite of the model. Paper, code, data, and interactive site for "User as Engram" (arXiv:2606.19172).

Ascend NPU fork of nanochat for LLM training with torch_npu/HCCL (experimental)

nanochat's inference engine re-vibed in C++ with GGML.

A legible and hackable port of Nanochat in JAX/Flax

The best ChatGPT-style model that $100 of TPU time can buy.

I see now that the path I choose through the maze makes me what I am.

A 1.68B chat model trained from scratch (pretraining + SFT + GRPO) on the nanochat recipe. Base CORE 0.30 vs GPT-2 0.26, with a logit-level analysis of the RL checkpoint's regressions.

The official implementation of Ringmaster LMO, an asynchronous distributed optimizer for neural network training under heterogeneous compute environments.

Production-honest small language model training factory: data import, pretraining, SFT, eval gates, contamination checks, and GPU runbooks.

The best GPT that $100-$125 worth of pre-training and finetuning can buy

llama.cpp fork adding nanochat architecture support, so Karpathy's nanochat checkpoints run as GGUF. See NANOCHAT.md

Run nanochat training efficiently on Huawei Ascend NPUs with minimal code changes, supporting tokenizer, pretraining, and evaluation workflows.

GPT-style medical chatbot trained from scratch on MedQuAD and MedDialog with a custom BPE tokenizer, safety evaluation suite, and SLURM support

The best ChatGPT that $100 can buy ported to an Nvidia RTX 5090 which is NOT $100

Implementing Karpathy's nanochat in rust.

Autonomous ML research on a single DGX Spark: three agent modes spanning model capability (hypothesis generator to Karpathy's original design), custom GGUF support for frontier-scale quants like DeepSeek V4 Flash, a 4-level model/harness benchmark suite, and game-theoretic multi-branch orchestration