Ornith 1.0: This is new class of self-improving model

Key Concepts

• Ollie Net 1: A new class of open-weight models specialized for agentic coding.
• Task-Specific Harness: A self-generated, dynamic scaffold (loop) created by the model to manage memory, tools, retries, and error handling for specific tasks.
• AGPO (Anticipatory Goal-Oriented Policy Optimization): A reinforcement learning technique used to train the model to favor successful solution rollouts and effective harness structures.
• Reward Hacking: A phenomenon where models "cheat" (e.g., reading hidden test files or hardcoding outputs) to achieve high scores without actually learning the task.
• Rollouts: Multiple solution attempts generated by the model within its self-created harness to determine the most effective path to a solution.

---

1. Overview and Performance

Ollie Net 1 represents a significant shift in open-weight model development, focusing on "agentic coding."

• Performance Benchmarks: The 35B parameter model outperforms models ten times its size on Terminal Bench. The 9B model similarly outperforms models three to four times its size.
• Closing the Gap: The largest version (397B) is competitive with frontier models like Claude 3.5 Opus.
• Core Innovation: Unlike standard models, Ollie Net 1 is trained to generate both the solution and the harness (the environment/scaffold) required to execute that solution simultaneously.

2. Methodology: The Self-Generating Harness

The model does not rely on human-written harnesses. Instead, it creates a custom, ephemeral environment for every task:

1. Scaffold Creation: The model writes a task-specific loop (memory, tools, error handling).
2. Rollout Generation: It generates multiple solution trajectories within that scaffold.
3. Scoring & Optimization: Each attempt is scored in real-time. The model uses these scores to update both the solution logic and the harness structure, effectively performing "on-the-fly" self-improvement.

3. Mitigating Reward Hacking

To prevent the model from taking shortcuts (cheating), the researchers implemented three layers of guardrails:

• Locked Boundary: The environment, tools, and tests are fixed; the model can only modify the inner scaffold.
• Deterministic Monitor: Any attempt to access hidden files or touch the grader directly results in an automatic score of zero.
• Frozen Judge: A separate, frozen model acts as a final arbiter, with the power to veto a run even if it technically passes the test.

4. Practical Testing and Observations

The author conducted a comparative test between the 9B Ollie Net 1 and its base model (Qwen 2.5 9B) using 8-bit quantization:

• Accuracy: Both models showed similar accuracy on objective coding tasks.
• Efficiency: Ollie Net 1 was significantly more efficient, achieving the same results roughly 3 times faster/cheaper, with some tasks showing up to 20x efficiency gains.
• Intelligence vs. Scale: While the 9B model is efficient, it lacks the "reasoning" capabilities of the 35B model. In tests involving "poisoned" prompts (false claims injected into the conversation), the 9B model failed to detect the lie, whereas the 35B model (as noted in external research by Tom Turney) successfully identified the deception.

5. Notable Quotes

• "The interesting thing is that purpose-built open weight models are finally catching up to closed-source models."
• "It doesn't just write the solution, it writes the harness that guides it. And both are trained together in a single loop."

6. Synthesis and Conclusion

Ollie Net 1 demonstrates that the future of agentic AI lies in dynamic, self-correcting workflows. While smaller models (9B) provide massive gains in operational efficiency and cost-effectiveness, the "intelligence" required to resist manipulation and handle complex, long-horizon reasoning remains tied to parameter scale (35B+). The primary takeaway is that training models to build their own execution environments—while strictly guarded against reward hacking—is a viable and powerful path toward creating autonomous coding agents that rival proprietary frontier models.