train.games.ugrad.ai Hub Lightbright

LLM / tensor / transformer playground

Public spine: Karpathy Zero to Hero + Hugging Face NLP course. Pegs / cells are weights. games.ugrad.ai expands here — Glyph Lightbright is the steno + toy-cipher seat.

Fundamental rungs · playable

T0Scalar autograd · Value / backwardugrad-r0
T1Tensor MLP · batched matmulgo train
T2Activations · visible |w|Lightbright
T3Tokenizer · BPE / pegsalphabet
T4Attention + transformer blockr0 gpt
T5SFT / LoRA / preference cycleslanguage + path-tensor
T6Serve · multiplex · custom enginewebgrid
T7Steno named-spaces · toy XORLightbright

Live 3×3 W · MSE to target (T1)

W random · click SGD

Who pushed the Hugging Face field · by region

France Hugging Face (Delangue, Chaumond, Wolf) — transformers, datasets, Hub, PEFT, TRL, TGI. Mistral (Mensch, Lample) — Mixtral / Apache weights. United States OpenAI GPT-2 as the first mass clone target · Karpathy micrograd/nanoGPT/llm.c (μgrad’s upstream) · Meta FAIR PyTorch + LLaMA/NLLB · Google Attention Is All You Need + T5/Gemma · Stanford HELM · Eleuther / Together / Mosaic · AI2 OLMo · xAI as the default ugrad desk chat. UK DeepMind (Hassabis) · Stability’s early open-weight culture. Canada Mila / Bengio · Vector / Hinton lineage — the textbooks HF packaged. China Qwen (our offline qwen3:8b) · DeepSeek · THUDM/GLM · Yi / Baichuan / BAAI. Germany / CH LAION · Aleph Alpha · EPFL/ETH. UAE TII Falcon · G42 Jais (AR stair portal). Israel AI21. Korea HyperCLOVA · EXAONE · Upstage Solar. Japan rinna · Sakana · PFN. SEA / India SeaLLMs · AI4Bharat · Sarvam — Indic packs on the stair.

train-playground-ugrad.html · spec ugrad/workspace/ugrad-model-train-playground.json · hub