041
041 is building AI systems that learn to reason from small datasets rather than scaling up parameters, using a hyperbolic-space architecture and a compiler-based ML framework. Its chess engine is a pure neural network with no search, and its Sudoku model learned to solve extreme puzzles from just 1,000 examples, beating models up to 28 times larger.
- Artificial Intelligence
- Developer Tools
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
Modern AI models are built in flat, Euclidean space, which requires exponentially more parameters and compute as problems branch out, mirroring how knowledge itself expands like a tree. This inefficiency drives steep costs and limits the scalability of complex AI applications that demand ever-larger models and datasets.
Solution
041 provides a radically different AI architecture based on hyperbolic space, which opens up faster the further out it goes, matching the natural expansion of knowledge. Their framework, S-expGPU, is the world's first ML framework designed for agents, enabling fully automated research by compiling a single-file experiment into a runnable program in under 120 lines. The compiler owns the training loop, autodiff, gradient accumulation, evaluation, checkpoints, and precision, eliminating the silent bugs that come from rewriting these components for every idea. Agents can check a full GPT-2 124M run in 45 ms on a laptop with no GPU and no data, and the CUDA device compiles the whole step from IR into fused generated kernels with no Python in the loop, delivering faster step times than torch.compile across all 14 models benchmarked. The company also demonstrates pure-AI reasoning engines, including a chess engine that is a neural policy trained on Stockfish's choices with no search, and a Sudoku model that learned to solve extreme puzzles from 1,000 examples with no rules written in.
Target Audience
Primary users are ML research teams, AI product developers, and autonomous research agents that need to iterate on experiments rapidly without GPU queues or manual training-loop debugging.
Features
- One `.sx` file contains the entire experiment: data, model, objective, optimizer, and run, in under 120 lines
- Compiler owns the training loop, autodiff, gradient accumulation, evaluation cadence, checkpoints, precision, and kernels, keeping them out of the user's file
- Every shape is checked at compile time on a laptop without GPU or data, producing diagnostics with code, line, expected and actual values, and the fix
- Based on Common Lisp with closures, backquote macros, gensyms, and lambda lists, so a transformer block reads like the math and macros can write ablations
- Correctness is validated by matching published loss curves within reference trial spread, not by self-consistency
- CLI verbs (check, explain, run, bundle) return structured JSON output and stable exit codes, built for agent consumption without prompts or dashboards
- Pure neural-network chess engine with no search, trained on Stockfish's choices, with an interactive play interface
- Sudoku model that learned to solve extreme puzzles from 1,000 examples, achieving 99.8% accuracy and beating models up to 28 times larger