Generalized Hill-Climbing at Runtime: A Path to Universal Verifiability

In the rapidly evolving landscape of AI, most labs are leveraging a combination of pre-training and reinforcement learning (RL) to develop more generalized models. These models aren't just proficient in one task but can excel in multiple domains and adapt to new challenges. However, this approach requires substantial computational resources, making it accessible primarily to large research institutions.

A Different Path: Generalized Hill-Climbing

For those without access to extensive GPU farms and deep pockets, there's an alternative route to achieving generalized models: generalized hill-climbing. This concept has been simmering in my mind since 2023 or 2024, but it was a tweet by Andrej Karpathy that crystallized it for me:

Software 1.0 easily automates what you can specify. Software 2.0 easily automates what you can verify., Andrej Karpathy

This insight sparked a profound question: How can we make everything verifiable?

Ideal State as a Path to General Verifiability

The concept of an "Ideal State" has been a recurring theme in my work, but the focus on verifiability brought it into sharp relief. The core idea is simple: to achieve any goal, you need a clear target. Whether you call them ladder rungs, footholds, or stepping stones, there must be something to aim for.

Articulating the Ideal State

The challenge lies in clearly defining this Ideal State, especially across various task types. This is where the importance of precise articulation comes into play. In 2024, I wrote "AI is Mostly Prompting," emphasizing that nothing compares to a well-articulated intent. Similarly, my 2025 piece "Coding is Thinking" discussed how writing and coding are fundamentally about expressing ideas clearly.

Most recently, in "How to Talk to AI" (June 2025), I argued that if you can't articulate what you want, no amount of prompting or context will help. The Ideal State concept embodies this principle perfectly.

Technical Details and Implementation

To implement generalized hill-climbing at runtime, the following steps are crucial:

• Define the Ideal State: For each task, clearly specify the desired outcome. This could involve creating detailed specifications, test cases, or even interactive prototypes.
• Iterative Refinement: Use an iterative process to refine the model's performance. Each iteration should bring the model closer to the Ideal State.
• Verification Mechanisms: Implement robust verification mechanisms to ensure that each step of the hill-climbing process is moving in the right direction. This could include automated testing, human-in-the-loop validation, or a combination of both.

Benefits and Challenges

Benefits:

• Flexibility: Generalized hill-climbing can be applied to a wide range of tasks without requiring extensive retraining.
• Verifiability: Each step of the process is verifiable, ensuring that the model remains aligned with the Ideal State.
• Resource Efficiency: This approach is less resource-intensive compared to large-scale pre-training and RL.

Challenges:

• Articulation Difficulty: Clearly defining the Ideal State for complex tasks can be challenging.
• Iterative Overhead: The iterative refinement process may require significant computational resources over time.
• Verification Complexity: Ensuring that each step is verifiable can add complexity to the development process.

Conclusion

Generalized hill-climbing at runtime offers a promising alternative to traditional approaches in AI model development. By focusing on clear articulation of the Ideal State and iterative refinement, we can create models that are both flexible and verifiable. This approach democratizes access to advanced AI capabilities, making it possible for smaller teams and organizations to achieve significant results.