LLM From Scratch: Transformer Workbench

Autoregressive multi-head self-attention, live tensor math, and micro-SGD training

In-Browser TinyGPT Engine

🧠 Transformer Architecture & Live Tensors

Params: 1,648
1. TOKENIZER & EMBEDDING (d_model=16) [2, 16]
2. MULTI-HEAD SELF-ATTENTION (Head 0 Scores) Softmax(Q · Kᵀ / √d_k)
Attention Weights Matrix Head 0 (d_k=8)
Hover/Click matrix cell: Q[1] · K[0]ᵀ / 2.83 → Softmax = 0.30
3. FEEDFORWARD (MLP 16→64→16) + RESIDUAL LayerNorm & ReLU
Residual stream accumulates attention output + FFN projection before next block.
4. OUTPUT UNEMBEDDING & PREDICTED TOKEN Top Logit: Next Token
"sat" (p=0.482)

⚡ In-Browser Training & Compute Scaling

Loss: 2.079
Initial Loss
2.079
Current Loss
2.079
Trained Steps
0
Learning Rate
0.050
Steps 0 → 10
HARDWARE ESTIMATOR (Paul Graham 17yo Premise) Compute Budget
EXPORT VALIDATED PYTHON / PYTORCH SCRIPT
# micro_gpt.py - Autonomous Mini-Transformer import torch, torch.nn as nn from torch.nn import functional as F class TinyGPT(nn.Module): def __init__(self, vocab_size=8, d_model=16, n_heads=2, n_layers=2): super().__init__() self.token_emb = nn.Embedding(vocab_size, d_model) self.pos_emb = nn.Embedding(6, d_model) # Multi-Head Attention & MLP Layers
Enjoy this tool? Build your own with Super