The Rockstar 0.1.0 Vision
Moving from Level 3 Analytics to Level 5 Autonomous Systems Co-Design.
The goal of MLSys·im v0.1.0 is not merely to be a calculator for humans, but to serve as the definitive Autonomous Operating System for Infrastructure Design.
We categorize infrastructure modeling into 5 Levels of Autonomy. Most of the industry is stuck at Level 1 or 2. Our 0.1.0 architecture propels us directly to Level 5.
The 5 Levels of Infrastructure Autonomy
Level 1: Spreadsheet Physics
Engineers manually calculate arithmetic intensity using static numbers in Excel. * The Trap: Error-prone, silent unit conversions, impossible to version control.
Level 2: Cycle-Accurate Simulators
Frameworks like ASTRA-sim write millions of lines of C++ to simulate exact memory bus timing. * The Trap: Requires hours to compile. A single training run takes weeks to simulate. Too brittle for rapid architectural changes.
Level 3: Analytical Engines (The Baseline)
The current baseline of MLSys·im. It uses first-principles math and the Iron Law to instantly bound physical capabilities. * The Win: Sub-second execution time. * The Trap: A human still has to manually write loops to find the optimal architecture.
Level 4: The Co-Design Agent
We wrap MLSys·im in the Model Context Protocol (MCP). * The Reality: An LLM agent (like Claude or Gemini) takes an SLA (e.g., “$10M budget, 1.5T parameter model”), writes its own YAML configurations, tests them against the engine, reads the resulting OOMError JSON payload, and adjusts its architecture autonomously until it finds the optimal solution.
Level 5: The Self-Improving System
The ultimate “Rockstar” vision for v0.1.0. * The Reality: The agent doesn’t just iterate on existing hardware. It reads a new ArXiv paper (e.g., the DeepSeek-V3 Technical Report). It extracts the mathematical breakthrough for Dual-Pipe MoE routing. It dynamically writes a new Python class inheriting from BaseSolver, hot-loads it into the MLSys·im Pipeline, and breaks a previously impassable Network Wall.
Why MLSys·im Enables Level 5
Other tools fail to reach this level because they try to be everything. MLSys·im is designed as an Intermediate Representation (IR).
- Bring Your Own YAML (BYOY): Hardware architects define future silicon predictively in simple text files.
- Bring Your Own Solver (BYOS): The 3-Tier API Contract ensures that any new math synthesized by an agent instantly plugs into the global plotting and execution framework.
- Dimensional Strictness: Agents hallucinate math. By forcing all inputs and outputs through
pint.Quantity, the engine catches AI hallucinations at the physical layer, ensuring the generated architecture is actually bound by the laws of physics.
To see the Level 5 loop in action, run the simulated demonstration in our repository:
python3 mlsysim/examples/level5_self_improving_loop.pyRoadmap
| Milestone | Status | Description |
|---|---|---|
| Level 3 — Analytical Engine | Shipped | 24 resolvers, 22 walls, CLI, Python API |
| Level 4 — MCP Co-Design Agent | Shipped | Claude Desktop integration, predict-compute-reflect loop |
| Level 5 — Self-Improving System | Prototype | Agent-authored solvers, hot-loading, ArXiv integration |
| Interactive Apps | In Progress | Marimo notebooks for roofline, SLA, and carbon exploration |
| PyPI Stable Release | Planned | pip install mlsysim on PyPI with semantic versioning |
| Expanded Zoo | Ongoing | Community-contributed hardware specs and model profiles |
MLSys·im v0.1.0 is in active development. The analytical engine (Level 3) and MCP integration (Level 4) are functional. Level 5 capabilities are demonstrated in prototype form. We welcome contributions — see the Contributing Guide.