01 / Planned product
Modular Minimal RLM
A minimal, bloat-free harness built on Recursive Language Models by Alex L. Zhang, Tim Kraska, and Omar Khattab, with extensibility inspired by Pi.
For a young project with a small community, extensibility needs a dependable starting point. The core team will ship and maintain modules, so users can add capabilities without having to evaluate community extensions or work out which ones are still maintained.
- Memory safety: make the runtime dependable for long-running agents.
- Cache efficiency: improve RLM cache hit rates and reduce total cost in day-to-day use, measured against traditional agent harnesses.
02 / Planned product
ASI (Accelerate Self-Learning)
A recursive self-learning system designed to generalize across domains and adapt the learning process itself for faster results.
Learning will be easy to toggle off: explore and improve during development, then freeze learned behavior for production when stability matters more than accelerated learning.
- Simulations: ARC-AGI-3, video games, vending benchmarks (VendorBench), PokerBench, F1Race, strategy games, and SocietyElection. For F1Race, we want agents to infer the underlying physical model from limited visibility (a 60° field of view and 100 m range), with procedurally generated tracks and MCP tools exposing velocity, acceleration, position, and track observations.
- Real world: long-running QA audits that create or rate repository issues; multi-repo coding and CLI/browser skill engineering; trading; Reddit marketing; Luma acceptance prediction; growing faceless social channels; and GPU kernel optimization. For QA and acceptance tasks, agents should report both their classification and confidence.
