Intelligence, built from zero.
LyceumAI is a frontier research project: a complete neural-network framework written from scratch in Mojo — every gradient verified numerically, every result gated — up to models that generalize rules they were never taught.
- grad checks
- 23
- dependencies
- 0
- epochs verified
- 60k
$ mojo run ml/layers.mojo
23 gradient checks≤ 2.4e-8
$ mojo run ml/models.mojo
grok MLP · MiniGPT · CNNPASS
$ /tmp/grok_full
30k epochs × 2 decaysdiag 1.00
One framework. Every layer built from zero. Every claim gated.
The framework
Arena autograd, dense/CNN/Transformer layers, optimizers, generic trainer — nine Mojo modules, no dependencies, every line owned and readable.
ml/autograd.mojo · ml/layers.mojo · ml/trainer.mojo
The gates
Trust is a gate, not a claim. Numeric gradients against central differences, exact-value spot checks, fit thresholds that abort on failure.
grad checks ×23 ≤ 2.4e-8 · full-batch == per-sample < 1e-12
The models
Trained in-house from scratch — including grokking, where a network trained on modular arithmetic generalizes to pairs it was never shown.
grok MLP diag 1.00 · MiniGPT held-out 94.5% · CNN held-out 80.0%
Artifacts of the lab
Three networks trained end-to-end by our own framework — no pretrained weights, no borrowed checkpoints. Full results on the research page.
grok-mlp-mod14
MLP 28-128-14 trained on all-but-diagonal (a+b) mod 14 pairs. Memorizes by epoch ~1000, generalizes to the 14 held-out diagonal pairs at epoch 9000 (wd=1.0) / 25000 (wd=0.3) — diag accuracy 1.0.
minigpt-char-v6
2-block MiniGPT (Embedding + learned positions + causal MHA + vocab head) fitting a deterministic char grammar over vocab 6 through real attention. Train 97.7%, held-out 94.5%.
simple-cnn-stripes
Conv-ReLU-Pool-Flatten-Linear classifier on flattened-NCHW input, fit on the synthetic stripes task with im2col forward and col2im backward. Train 97.8%, held-out 80.0%.
mojo-lm-ctx256
2-block MiniGPT trained on a 125-char tokenized corpus with context window 256. Teacher-forced validation: loss 3.26, top-1 26.5%, top-5 55.7%.
Project pulse
- Gradient Checks
- 23
- Grok Epochs Run
- 60k
- Framework Modules
- 9
- MiniGPT Held-out
- 94.5%
- Train Throughput
- 15.4 ep/s
- Dependencies (ml/)
- 0
- Char LM Context
- 256
all pass, worst 2.4e-8
30k × 2 decays, diag 1.0
tensor → trainer
train 97.7%, 26k params
full-batch, pure Mojo
no ML libraries used
tokens, 2-block MiniGPT
Research in the open. Verified to the last digit.
Every claim on this site is backed by a runnable gate. Clone the repo, run the suites, watch the grokking transition yourself.