Key Moments
How AI Is Entering the Physical World | Deep Dives with a16z
a16zKey Moments
AI is automating hardware design and construction, reducing build times by decades and costs significantly, but faces data scarcity and entrenched industry practices. Physical world AI promises to revitalize US industrial capacity, yet it requires bridging old systems with new technology.
Key Insights
Automating construction design could reduce 1-1.5 year design cycles to a button click, producing globally optimized 'Issued For Construction' packages.
AI can automate specific electronics design tasks within two years, with the potential to transform software engineers into electrical engineers by enabling code generation for hardware.
The US electronics manufacturing industry faces an 80/20 split between robotic automation and manual labor, with a significant challenge in automating the remaining 20% of tasks.
In construction, incentives are structured around risk reduction for investors, which disincentivizes the adoption of new technologies that could improve efficiency.
Simulation is used as a 'calculator' or tool for AI to run optimizations, with the long-term goal of moving simulation from inference time to training time to accelerate model development.
The lack of readily available, digitized data on circuit board design is a critical barrier for AI automation, with companies like Apple or SpaceX holding significant but siloed datasets.
Revolutionizing construction design with AI
Alex Modon of Unlimited Industries predicts that all construction will be fully automated within 10 years. The current process for large infrastructure projects, such as power plants or hospitals, involves a year or more of design work by hundreds of engineers and project managers, culminating in an 'Issued For Construction' (IFC) package. Unlimited Industries aims to automate this end-to-end process, where AI explores tens of thousands of design permutations based on site requirements and desired outcomes. The optimization can focus on capital expenditure (CAPEX) but ideally targets total cost of ownership, considering operational efficiency and constructability. This parametric approach, similar to software development, allows for optimization on any defined metric. The subsequent phase involves robotics, from autonomous earth movers to a site full of drones and autonomous robots, expected to be realized over the next decade, provided incentives are properly aligned.
Accelerating electronics design and manufacturing
David Asnaghi of Diode Computers believes that specific subsets of electronics design, particularly circuit boards, could be fully automated within two years. He notes the rapid, 'wild' jump in design capabilities with each new AI model tier. Diode's focus extends to manufacturing, where the electronics industry has long used robots for processes like surface mount technology (SMT). However, an 80/20 split remains, with 20% of tasks, such as soldering large components or assembling circuit boards into enclosures, being difficult to automate manually. The goal isn't to replace all design work but to automate the design of manufacturable outputs. Asnaghi posits that if designs are constrained for manufacturability, 100% automation is achievable with current robotics, eliminating the need to wait for robotic advancements. This mirrors the complexity of data center design, which involves fitting components within site footprints and meeting numerous specifications, a task surprisingly similar to designing compact circuit boards.
Transforming engineers through code generation
The vision at Diode Computers is to empower individuals with coding abilities to also generate hardware. This means extending the capabilities of software engineers, who are accustomed to rapidly adopting new frameworks, to the realm of hardware design. By leveraging the structured nature of code, AI agents can perform tasks previously considered orthogonal to coding. Diode has developed a compiler that provides AI models with enough context to feel like they are writing Python programs when designing circuit boards. This approach intends to enable a broader set of individuals, including AI agents, to generate hardware. While current designs are limited in complexity, the derivative suggests significant future advancements within one to two years, aiming to give agents the same ability to generatee hardware that they have to generate software.
Overcoming industry inertia and incentive structures
Both construction and manufacturing are traditional industries with entrenched ways of working and established toolsets, contrasting sharply with the agile nature of software engineering before AI. Alex Modon highlights that in construction, a stage-gated process focused on removing risk for funding dictates how projects proceed. Investors seek stable Internal Rates of Return (IRR) and are disincentivized from taking risks on new technologies, as there's no direct upside for those implementing them. This leads to a slow adoption rate for new technologies, with firms sometimes using antiquated computer systems. Unlimited Industries pursues vertical integration to create a clean interface with the industry, owning enough of the process to implement change effectively rather than trying to chip away at small parts.
Building teams with multidisciplinary expertise
Recruiting talent for these cutting-edge physical world AI ventures involves assembling multidisciplinary teams. Unlimited Industries finds it easier to teach AI tools to domain experts in mechanical, electrical, civil engineering, and simulation controls, rather than teaching domain expertise to software engineers. The core principle is integrating AI as a primitive in product development, leveraging existing domain knowledge. Similarly, Diode Computer's approach involves teaching multidisciplinary individuals AI tools, especially those with backgrounds in advanced vehicles or rockets, suggesting a transferable skill set from high-tech physical domains to AI-driven hardware development.
Shifting industry perception through product delivery
Diode Computers reframes its engagement with enterprise customers, particularly Fortune 100 companies. Instead of selling software, they offer an end-to-end service where customers provide specifications and receive a physical product, aligning with existing business models. The key differentiators are speed and cost reduction. Diode also provides transparency into the design process, akin to real-time updates on GitHub pull requests, allowing engineers visibility into their board's design and delivery. This 'everything is code' philosophy extends to their open-source compiler toolchain, aiming to build infrastructure for AI models rather than owning core design primitives. The bet is that this infrastructure will benefit both humans and AI agents.
The role and evolution of simulation in AI design workflows
Simulation plays a crucial role in computing design values, acting as a 'calculator' to solve complex, interdependent variables. For instance, simulating mass flow rate of fluids involves running scenarios through simulation environments to optimize towards a goal. While simulation software has existed for decades, the AI approach involves training AI to use these specific tools and run optimizations. In the construction world, simulations cover everything from electron flow to structural responses to earthquakes. In electronics, tools like SPICE and electromagnetic simulators are used. Currently, simulation serves as a verification step. The ultimate goal is for simulation to become a 'training time' tool, enhancing AI's intuition and taste, rather than an 'inference time' process, making AI development more efficient.
Addressing data scarcity and the potential of new architectures
A significant hurdle for automating circuit board design with AI is data scarcity. While the data exists within companies likeApple and SpaceX, it is siloed and not readily shared. David Asnaghi believes the biggest frontier is generating sufficient data for AI models. However, there's a competing view that well-structured problems can be solved through reinforcement learning and Monte Carlo research, potentially negating the need for more data. In the interim, Diode focuses on end-to-end solutions, bridging any remaining gaps. They are optimistic that current advancements in AI architectures and data generation capabilities will lead to breakthroughs, and they aim to be positioned to leverage these advancements. For construction, while data is sparser, industry standards help bound problems, and the bar to beat the status quo is low, making full automation a probable outcome by designing systems for autonomy.
Humanoids and robotics in the future of physical AI
The integration of humanoids and specialized robotics into physical industries is a topic of much discussion. While Alex Modon finds the prospect of a future with humanoids inspiring and a core component of Unlimited Industries' vision, he acknowledges a mix of humanoid and purpose-built robots will be employed. The efficiency gained from mass-manufacturing standardized designs is significant. David Asnaghi, while loving all robots, believes that for electronics manufacturing, where automation is already advanced, the focus will be on bridging the remaining 20% with smarter robotic arms, computer vision, and potentially humanoids for specific tasks like soldering chunky components. The broader impact of automating knowledge work, however, is seen in essential industries like mining, where robots are preferable to human labor. Both believe in facilitating this transition and investing in robots, though not necessarily betting on a single form factor like humanoids.
Capturing tacit knowledge and the future of skilled labor
The tacit knowledge of experienced engineers and tradespeople, often referred to as process knowledge, is invaluable. In the US, an aging workforce means this expertise is retiring quickly. While AI can potentially encode some of this intuition, there's a societal need to train the next generation of skilled workers. The demand for tradespeople, like electricians, is so high that salaries can exceed those of Silicon Valley software engineers. This shortage is driving companies to explore mass manufacturing of components, like data center modules, to circumvent labor constraints. The strategy outlined is 'all of the above': training more people in practical skills, codifying existing knowledge, and embodying it into robotics for future scalability in an era of abundance.
Rebuilding US industrial capacity and second-order effects
The overarching second-order effect desired by companies like Unlimited Industries and Diode Computers is the ability to spin up hardware companies as easily as B2B SaaS companies. They aim to lower the barrier to entry for creating physical products, fostering innovation and re-industrialization in the US. The current trajectory of US construction metrics shows declining labor productivity and increasing adjusted CAPEX over 50 years, a stark contrast to software's progress. By improving the building lifecycle by an order of magnitude, they hope to re-establish the US's capacity for ambitious projects. This includes supporting AI advancements, building data centers, and fostering advanced manufacturing. The goal is to reverse the trend of 'getting worse' at building essential infrastructure and to foster a future of abundance driven by intelligent automation.
Mentioned in This Episode
●Products
●Software & Apps
●Companies
●People Referenced
Common Questions
AI is being used to automate the design phase of construction, exploring thousands of permutations to create globally optimized plans. Future automation will also involve robotics like autonomous earth movers and humanoids on construction sites.
Topics
Mentioned in this video
CEO of Unlimited Industries, discussing the future of AI in construction and infrastructure projects.
CEO of Diode Computers, focusing on AI's role in designing and manufacturing circuit boards.
Mentioned as someone who coined the term 'process knowledge' regarding industrial expertise.
Mentioned as a place where David Asagi visited manufacturers and learned about the energy cost of wave reflow ovens.
State mentioned in the anecdote about Microsoft employing a large number of electricians for data center construction.
Mentioned as a place where the speaker lived and observed cultural differences regarding manufacturing and design.
Location mentioned when comparing the salary of an electrician to a Silicon Valley software engineer.
Mentioned as a point of comparison for software engineer salaries against electricians.
Mentioned as a benchmark for cost-competitive manufacturing that Diode Computers aims to match or beat in the US.
The country where Diode Computers aims to manufacture circuit boards faster and cheaper than anywhere else.
Mentioned for its 'process knowledge' in manufacturing and for electrical engineering designs that prioritize ease of manufacturing.
An AI-native firm that vertically integrates design, engineering, procurement, and construction for large infrastructure projects.
A manufacturing company cited as an example of solving complex assembly problems with labor.
A company using AI to design and manufacture custom circuit boards, aiming to make hardware development as accessible as B2B SaaS.
Mentioned in an anecdote about employing a large number of electricians for data center construction.
Mentioned as a platform where Diode Computers' core compiler toolchain is open-sourced, allowing users to see code development in real-time.
Mentioned in the context of Foxconn's manufacturing solutions for Apple products.
Mentioned alongside Foxconn as a company that has solved assembly problems with labor.
A company whose materials are used by David Hansen for building motors.
Mentioned as an example of a company that likely holds valuable, siloed data related to hardware development.
A company mentioned in relation to the rapid advancements in AI model design capabilities.
A simulation tool used in electrical engineering at the schematics level.
Mentioned as an AI tool that enables agents to leverage code for non-code tasks, and also in the context of making designs more manufacturable.
Mentioned as a programming language that AI models can feel like they are writing when designing circuit boards with Diode's compiler.
The final output of the construction design process, which AI can optimize to explore thousands of permutations.
An open-source tool for electromagnetic simulation kernels.
Commercial software used for electromagnetic simulation at the board level.
Mentioned as an example of a complex system that can be represented by an ontological model.
More from a16z Deep Dives
View all 50 summaries
41 minOutsmarting Uber: Why Bolt Wins in Europe | Deep Dives with a16z
66 minMarc Andreessen on how the internet changed news, politics, and outrage | The a16z Show
34 minWhy Claude Feels Different (And What That Means for AI) | The a16z Show
60 minThe Next Wave of AI Companies in Latin America | Tako CEO on The a16z Show
Found this useful? Build your knowledge library
Get AI-powered summaries of any YouTube video, podcast, or article in seconds. Save them to your personal pods and access them anytime.
Get Started Free