Designing the Physical World with AI

Automating the Physical World with AI

The podcast introduces the core challenge of bringing software's rapid development cycle to the physical world of manufacturing and construction. Historically, industrial power came from scaling design and build processes, but these sectors still face timelines stretching into years due to fragmented workflows and systems resistant to change.

The central question explored is how to use AI to "move atoms, not just bits." While AI can write code, run simulations, and generate designs across thousands of permutations, the goal is to translate these capabilities into faster physical builds rather than just improved plans.

Diode Computers, led by Davide Asnaghi, is one company tackling this by using AI to design and manufacture custom circuit boards. They've developed a compiler that enables AI models to write what feels like Python code, essentially designing circuit boards through a more abstract, model-led approach.

This shift requires generating significant data to enable fully autonomous systems, moving beyond human intervention. The ambition is to allow hardware companies to spin up with the same agility as B2B SaaS firms, confronting the problem of cheap intelligence meeting a slow physical world.

What happens when intelligence gets cheap but the physical world stays slow?

Automating End-to-End Design for Large Infrastructure Projects

Large construction projects, such as hospitals, currently involve an extensive pre-construction design phase that can take over a year. Hundreds of specialized engineers across various disciplines like mechanical, electrical, civil, and structural work together to create a detailed instruction package for the general contractor.

AI technology is now capable of automating this entire initial design process. Users feed site specifications and project requirements into the AI, which then explores tens of thousands of design permutations. This process culminates in the generation of a globally optimized "Issued for Construction" (IFC) package with a single click.

The AI's optimization goes beyond just initial capital expenditure. It considers the total cost of ownership, factoring in long-term aspects such as operation, maintenance, and the overall constructability of the facility. This approach allows for a flexible, parametric design process, similar to software development, where different metrics can be prioritized.

Beyond the current capabilities in design automation, the future of construction envisions widespread integration of robotics. This includes autonomous earth movers, humanoids, and drones operating on construction sites, a development expected to materialize within the next decade if proper incentives are in place.

That first part, that's like line of sight today of how we automate end to end. You will literally feed in a site, a bunch of different requirements about what you're trying to build, and anything you wanna stipulate about how it gets built, and AI is going to explore tens of thousands of different permutations about how to optimally design that facility, a button click, and then what you get back from that is a globally optimized IFC package.

Diode Computers Automates Electronics Design for Manufacturability

Diode Computers is working to fully automate electronics design within a specific segment, particularly designs that inherently produce highly manufacturable outputs. The company aims to leverage AI to bridge the current 80/20 automation gap in manufacturing, enabling 100% automated production today, as the necessary robotic technology is already in place; the primary challenge lies in the design's compatibility with full automation.

The 80/20 gap signifies that while robots handle about 80% of electronics assembly via surface mount technology, the remaining 20% involves manual processes for components that don't fit standard automation, such as large transformers or enclosure assembly. Traditional solutions rely on human labor, as seen with companies like Foxconn, but this limits scalability and slows production cycles, making it insufficient for rapid data center expansion.

Diode Computers envisions empowering software engineers, and even AI agents, to become hardware creators. The company's goal has evolved from enabling human software engineers to design electrical components to giving any code-generating entity the capability to create hardware. This strategy aims to automate both complex PCB designs and simpler, fully agent-driven board creations, leveraging the rapid advancements in code-generation capabilities.

We want to give anything that has the ability to generate code the same ability to generate hardware.

Overcoming Resistance to New Technology in Traditional Industries

Traditional industries like construction and electronics often resist adopting new technology due to entrenched financial incentives. Project finance, for instance, focuses on de-risking and achieving stable internal rates of return (IRR), discouraging innovation that might introduce new variables or perceived risks. This leads to a reluctance to embrace advanced tools, leaving some firms feeling stuck in earlier decades.

To introduce new technologies effectively, companies must consider deep vertical integration. In construction, this means owning enough of the process to create a clean interface with the industry, rather than trying to force individual components onto unwilling participants. Building a team that is multidisciplinary and adaptable to learning new AI tools, rather than relying solely on engineers fixed in traditional methods, is crucial.

Similarly, in electronics, the strategy shifts from selling specialized software, like CAD, to selling the end physical product itself. Companies like Diode leverage AI internally to design and deliver products faster and cheaper. This approach bypasses the need to convince customers to adopt new tools by instead providing a superior final product, with internal processes offering real-time visibility into the design and manufacturing stages.

This end-to-end control and focus on the final product also significantly impacts project financeability. Reducing the build time from years to months can dramatically improve a project's IRR, making speed a critical value proposition. Even if AI tools initially provide 90% efficiency, offering the remaining 10% as a service proves the immediate benefit while models continue to improve towards full automation.

You don't want to convince people to buy your software. You want to convince people to buy the end product, and this is a much harder company to build.

Companies Use Code-First and Model-Led Approaches to Teach AI Physical Constraints

Diode addresses the challenge of limited data in traditional electronics design by reframing the process as code generation. They developed a compiler that provides AI models with enough guidance to write Python-like programs for circuit board design, effectively leveraging the vast amounts of available code training data. This method allows the AI to manage the inherent physical constraints of electronic components.

Diode bootstraps its models by building a library of validated design blocks, which then serve as training data for future iterations. They are optimistic about diffusion architectures for this problem, aiming to create a platform where designs can be freely developed, generating a continuous stream of data to improve model accuracy and enable cost-competitive manufacturing.

Unlimited Industries tackles similar challenges with a "model-led" philosophy. They represent complex systems through robust ontological models that embody all relevant relationships. Within this framework, AI agents and large language models generate code, using deterministic tools in a manner similar to human engineers. This establishes a parametric relationship for the entire system.

This model-led approach transforms design modifications from a complete restart into simple variable updates. It enables highly optimized and iterative development, allowing for continuous improvements in resilience and performance. Both companies also utilize simulation extensively for design validation and to generate crucial data, such as calculating mass flow rates for fluids in pipes with various interdependent variables.

We basically built a compiler that gives the model enough hints that it feels like it's writing a Python program instead of designing a circuit board.

The Evolving Role of Simulation in AI-Driven Design

Simulation has long served as a fundamental calculator in engineering, enabling precise analysis of electron movement, fluid dynamics, and structural integrity. Historically, this involves calculating exact matches for inputs and outputs, akin to solving complex "Legos on hard mode." These tools provide essential grounding for models, particularly in reinforcement learning environments, by verifying designs such as electronic circuits without needing a pre-established "golden reference."

The vision for AI-driven design is to transform simulation from a verification step into a primary training tool for AI models. Currently, experienced engineers rely on their intuition and use simulation mainly for a final check. The aim is for AI to develop a similar "taste" or intuition through extensive training with simulation data, making the design process significantly faster.

Using simulation at training time, rather than inference time, is crucial because simulations are not inherently fast. By providing enough data, AI models can develop emergent properties that allow them to achieve optimal designs in about 95% of cases. This approach aims to accelerate product development, understanding that even the most skilled engineers still require physical prototypes to validate designs, as real-world conditions like enclosures can alter properties.

my eventual goal is that simulation become a, becomes a train time tool. That you use for the model to become better at developing that taste, having that intuition.

Closing the Final Automation Gap in Physical Design

The conversation explores whether the final automation gap in physical design requires a fundamental breakthrough in machine learning or simply more data. One perspective suggests that existing building blocks are sufficient, but a critical barrier is the lack of pooled data, especially for circuit board automation. This data currently resides in large tech companies like Apple, Meta, and SpaceX, which are unlikely to share it.

A competing view proposes that many problems in physical design are well-suited for methods like Monte Carlo tree search and reinforcement learning, potentially allowing for improvements without needing significantly more data. However, the current strategy involves building end-to-end systems and bridging any remaining gaps, with optimism for future advancements in AI architectures and data generation capabilities.

Regarding the necessity of a 'human in the loop' for physical automation, it is argued that despite the complexity of physical systems and external variables, most problems can be bounded by existing industry standards. The bar for improvement over the status quo is considered very low, suggesting that full end-to-end automation is achievable.

A crucial design paradigm is to engineer systems with the explicit requirement of full autonomy, rather than anticipating human intervention. This approach is believed to drive distinct architectural choices that inherently lead to more robust and capable automated systems, with a core bet on continuous model improvement.

I think it's actually an important design paradigm is like, for us specifically, is like making sure that you design the system to actually be fully autonomous and not be human in the loop.

Automating physical labor with humanoids and specialized robots

The discussion highlights the coexistence of specialized robotics and more generalized forms like humanoids in manufacturing. While centralizing manufacturing drives down costs through learning rates, specific custom efficiencies still matter, particularly in giant facilities. In highly automated sectors like electronics manufacturing, the remaining 20% of work often involves tasks that could be significantly improved with smarter robotic arms and advanced computer vision.

A concrete example from electronics is soldering chunky components. Currently, this is either done by hand or with a large, energy-intensive wave reflow oven that is too costly to operate for lower production volumes. This type of marginal, yet critical, task represents a clear opportunity for robotic automation.

Beyond specific manufacturing processes, a larger implication of robotics is the automation of hazardous physical labor. Industries like mining, which traditionally require humans in dangerous conditions, stand to benefit immensely from robots. This shift protects human workers and addresses a vast chunk of the global economy that relies on such foundational labor.

You don't want to put humans in harm's way, like you want to be able to actually have, the robots that do this kind of stuff.

Preserving Tacit Knowledge and Re-Industrializing US Manufacturing

Skilled tradespeople, including manufacturing engineers and electricians, possess invaluable tacit knowledge gained from years of hands-on experience, which enables intuitive problem-solving and efficiency. This expertise is rapidly retiring, creating a critical need to capture it and train the next generation of workers. Industries are already struggling; for instance, Microsoft is resorting to mass manufacturing data center components due to a shortage of available skilled trades for direct construction.

A significant cultural disconnect exists in US manufacturing, where product design is often separated from the practicalities of production. This abstraction, common when designs are sent overseas, causes the "design for manufacturing" muscle to atrophy. In contrast, designers in places like China often prioritize ease of production, even if it requires more upfront design time, due to closer ties with the manufacturing process.

To overcome these challenges and enable re-industrialization, a dual strategy is needed. First, codify existing knowledge and train more skilled workers. Second, leverage AI to generate designs that are inherently optimized for manufacturing, bridging the gap between design and production. This approach aims to restore mass production capabilities in the US, allowing hardware companies to be spun up with the same ease as software businesses.

I want to be able to spin up a hardware company the same way that my friends spin up B2B SaaS.

Enabling industrial capacity and physical innovation in the US

The speakers observe a significant disparity between the rapid advancements in artificial intelligence and software, which have seen remarkable progress in recent years, and the stagnant, often declining, state of physical product delivery. They argue that while AI can perform incredible computational feats, it struggles with the practicalities of manufacturing physical goods.

Their ultimate mission is to bring the same level of step-wise improvement witnessed in software development to the realm of physical design and manufacturing. This transformation aims to empower a new generation of American engineers and innovators, making it easy and accessible for them to build tangible objects, such as CubeSats, with streamlined production processes.

A critical motivation stems from the alarming decline in US construction and manufacturing productivity over the past five decades, a trend evident in labor productivity and adjusted capital expenditure metrics. This contrasts sharply with the continuous progress seen in software, indicating a concerning loss of national capability in undertaking large-scale, ambitious physical projects.

This erosion of "building muscle" has broad implications, affecting essential areas from the energy infrastructure required for AI data centers to the advanced manufacturing needed for re-industrialization and the secure supply of critical minerals. The company seeks to fundamentally improve aspects of this lifecycle to reverse this concerning trend and rebuild the nation's capacity for physical creation.

If you look at any construction metric in the US, like labor productivity or adjusted CapEx numbers over the past 50 years, we're getting worse.