About
Aether is a minimal, fully-transparent transformer LLM built from first principles. 0.57M parameters, C99 inference engine (350 LOC, zero dependencies), trained on code and knowledge corpora.
"What I cannot create, I do not understand." — Feynman
Training Phases
✅ v0.1: Jot
Syntax corpus. 200 epochs. Foundation.
✅ v0.2: Jung
JIT compiler. 100 epochs. Specialization.
✅ v0.3: Multilang
27MB code. 500 epochs. Loss 0.0947.
✅ v0.4: Knowledge
Balanced corpus. 200 epochs. Loss 0.1233.
🚀 v1.0: Mini
3.5M params. Full scale. Coming next.
Quick Start
Clone & setup:
git clone https://github.com/nulljosh/aether.git && cd nouspython -m venv venv && source venv/bin/activatepip install -r requirements.txt
Train:
python src/train.py --corpus tiny --epochs 200
Inference (C engine, 50K tok/s):
cd inference && make && ./aether ../models/aether.bin "fn " --temp 0.3
Web UI:
python index.py # http://localhost:5001
The Stack
PyTorch Trainer
AdamW, cosine LR, gradient clipping. Full control.
C99 Inference
350 LOC, mmap weights, zero deps. Runs anywhere.
Flask Web UI
Chat, quiz, status. Port 5001. Real-time inference.
Aether Daemon
Continuous training, checkpoints, iMessage notifications.
Benchmarks
| Model | Params | Speed | Capability |
|---|---|---|---|
| aether (v0.4) | 0.57M | 50K tok/s | code + knowledge |
| GPT-2 | 124M | — | coherent paragraphs |
| Claude | ??? | 80 tok/s | reasoning, tools |
Why Aether
- Full stack from scratch: tokenizer → attention → training → C inference
- No black boxes. Every byte visible and understandable
- Learning tool first, production model later
- C99 engine runs anywhere with a C compiler
- Progressive training shows how LLMs learn from data