Eyoel Lundberg
Swedish builder, dad, designer. Solo riding the AI wave.
What I'm building
Autosim — a domain-agnostic engine that trains tiny AI models from scratch on pure simulation data. No base model. No LoRA. No priors.
The approach
Every serious vertical application of AI today involves taking a large general model and steering it toward a specific domain. You're fighting against cooking recipes and Python tutorials. You're paying per token. You're sending your data to someone else's server. Dependent on a provider who can change pricing, deprecate models, or go down.
The alternative: own your intelligence entirely.
A 50M parameter model trained exclusively on domain reasoning outperforms a 4B general model fine-tuned toward it — because there's nothing to fight against. Every parameter encodes the domain. The model doesn't reason in natural language. It compresses the decision surface of a simulation into 50MB.
The simulation is the reasoning. The model is the fast lookup.
The engine runs an overnight loop: generate simulation data, train, probe against ground truth the model has never seen, score, keep or revert, repeat. ~100 experiments per 8-hour run. Git is the memory. Probe pass rate is the only metric.
- — Trains on Apple Silicon
- — Runs local. Private.
- — Zero API calls for decisions
- — Total footprint ~250MB. Runs on a phone.
The engine is domain-agnostic. Domains plug in. Each one is a simulator and a set of probes. The engine does the rest.