Building Intelligence for the Physical Economy

The next frontier of AI is evolving beyond the world's digital information, into understanding and improving the physical world itself. At Noetive, we are building toward a continuous cycle of recursive self-improvement to develop intelligence for the physical economy.

Public data, targeted data collection, and learning environments have driven major advances in AI so far. But much of the data and expertise needed to understand, reason, and act in the physical economy is not widely available. It lives in observations of people, machines, materials, and processes that often go unrecorded. It lives in proprietary business records and the often unwritten knowledge and judgment of experts who understand what works, what fails, and why.

Learning from the physical world means capturing observations and connecting them with human knowledge and judgment, the actions taken, and the outcomes that follow. Together, these form the real-world experience that models can learn from. Models need to understand complex environments through multimodal observations and act with judgment that builds on human expertise and real-world feedback. They also need to reason across long timescales, connecting actions to consequences that may emerge hours, weeks, or months later. The research challenge is to develop these capabilities so they generalize across tasks, businesses, and domains.

To learn from this experience at scale, we develop self-improving AI to build learning environments, foundation models, and agents that learn from them. The system is designed to automate the development of solutions and evaluations, along with gathering and using human expertise and real-world feedback to improve those solutions. Automating this process allows us to build learning environments across more tasks, businesses, and domains. Experience from these interactions supports the development of stronger foundation models and agents, which enable our self-improving AI to tackle more complex tasks and build better solutions.

We are building intelligence that understands, reasons, acts, and learns alongside people across the physical economy.

Our Approach

Self-improving AI is how we are building intelligence for the physical economy. We are building a flywheel that creates and improves learning environments at scale and develops foundation models from the experience they provide.

Building learning environments at scale. Each environment combines a real-world task, candidate solutions and evaluations, and feedback from people and the physical world. Useful solutions help us learn from data and expertise that are often held within businesses or remain unrecorded, including experts' implicit knowledge, judgment, and preferences. Self-improving AI explores approaches, builds and evaluates solutions, and uses the results to guide further research. Human guidance and real-world outcomes extend this process, refining both the solutions and their evaluations. Automating this research and engineering lets us build environments across more tasks, businesses, and domains.

To capture missing information, we are developing multimodal sensing hardware that records and aligns observations over time, alongside foundation models that interpret these observations in the context of records and human expertise. This helps connect actions to outcomes that may emerge hours, weeks, or months later. The environments retain what was observed and tried, the experts' feedback, and what happened.

Developing models from experience. Experience across these environments provides material for developing foundation models with stronger understanding, reasoning, and action capabilities. Self-improving AI runs experiments to improve how models learn from this experience, using evaluations across tasks and settings to guide further research.

Three connected stages guide this work, with sensing and model development supporting all three.

1. Build real-world solutions at scale. Build self-improving AI that develops solutions and evaluations across diverse tasks, and deploy those solutions to establish real-world learning environments.
2. Learn from human expertise and the real world. Extend these environments so self-improving AI can learn from expert guidance and delayed real-world outcomes.
3. Develop transferable intelligence. Use experience across environments to develop foundation models whose capabilities carry over to new tasks and settings.

Together, these stages form a flywheel of recursive self-improvement. Self-improving AI builds and improves real-world learning environments and uses the resulting experience to develop stronger foundation models. Stronger models enable it to tackle more complex tasks and build richer environments, generating experience for the next round of model development.

Early Results

Today, our work centers on the first stage: building real-world solutions at scale. We develop a general-purpose self-improving AI system that builds and evaluates solutions to specific tasks, including models, data, code, and agents.

We evaluate an early version of this system on public benchmarks in machine-learning engineering and mathematical optimization, which test capabilities relevant to the physical economy. The resulting solutions perform on par with or better than leading published results across these benchmarks, demonstrating the ability of our self-improving AI to tackle problems with different datasets, constraints, and objectives.

Machine-learning engineering. On MLE-Bench Lite, our system achieves an average medal rate of 89.4% across three independent research runs on 22 competitions, with a 12-hour budget per competition. This exceeds MLEvolve’s 80.3% under the same time limit and Arbor’s reported 86.4%. These tasks require selecting and improving algorithms across diverse datasets and objectives.

Mathematical optimization. A single 12-hour search on OptMATH-Train produced a reusable solution. Applied directly across four benchmarks, it achieves results on par with or better than leading published approaches. These tasks require formulating and solving optimization problems from natural-language descriptions. Scores are averaged over three evaluations. On RetailOpt-190, it achieves 87.7%, compared with ReLoop’s published 31.1% using a different model and solver, and 81.2% from directly prompting a frontier model without self-improving AI. It achieves 90.9% on IndustryOR, compared with OptiMind’s 85.2%, and remains competitive on MAMO-Complex and OptMATH-Bench. These results show how self-improving AI can develop reusable methods, allowing one research effort to support many tasks.

These results also highlight the need for evaluations that better reflect the scale and complexity of the physical economy, particularly as some existing benchmarks approach saturation. We are developing these benchmarks as part of our work to build real-world solutions and learning environments.

Real-world use case. While public benchmarks demonstrate technical breadth, the real test of intelligence is real-world impact. We are already deploying our self-improving AI in live enterprise settings, and seeing dramatic performance uplifts. In a recent pilot with a logistics design partner, the challenge was daily freight planning: dynamically consolidating orders, allocating trucks, and holding shipments to minimize total freight spend. Prior to this, a team of product experts and research engineers spent ten weeks hand-crafting a solution using a leading frontier model, and cut freight costs by 2.7%. In under 48 hours, our self-improving AI generated a solution that outperformed that ten-week baseline, and cut costs by 4.5% compared to the partner’s existing production system.

Next Steps

This is just the beginning. Over the coming weeks, we will share dedicated case studies of the real-world use cases, our end-to-end methodology, and how our system solves complex combinatorial operations in the wild.

Our next research focus is the second stage: learning from human expertise and delayed real-world outcomes. This feedback will help refine solutions and evaluations and provide richer experience for developing stronger foundation models.

If you’re excited about building intelligence for the physical economy, we’d love to connect. We also welcome research collaborations to push this frontier forward together.