AI Research and Enterprise Solution
Through the lens of manifold covers, we open a mathematical window into neural networks.
deepManifold "eTude"
eTude
A brand-new neural network architecture, grounded in Deep Manifold Theory—built primarily on the mathematical foundations of inequality and calculus. Deep Manifold Comparison
25-50% size of GPT-2 124 M
- Reference : 3.24 Eval Loss, 1B FineWeb Token, < 2000 Iteration, see Modded-NanoGPT SpeedRun
- Target: 25% to 50% size of 124 M, same performance & training configuration above.
- Current Status: 8.3M (1/15 of 124), 5.03 Eval Loss, 1B FineWeb Token, < 2000 Iteration
Is Not
- Not incremental improvement of Transformer, Mamba, Diffusion, any existing architecture
- Not physical theory-inspired approach like flow matching, energy based model
- No data work like distillation,
- No quantization, no model merge,
- No addition GPU/CUDA work
About deepManifold
Our Mission
We are a boutique Gen AI firm (Seattle, San Francisco), founded and led by principal experts—each bringing 30 to 40 years of global experience in research and industry, with deep expertise in healthcare and natural disasters prevention and response