a16z.news essay: superhuman AI hardware engineers are coming, but the physical world will slow them

An essay on a16z.news opens with a joke about the author's first FPV drone, which took eight months: parts bought in January, received in February, forgotten until June, set on fire in July, crashed into a neighbor's roof in August, and only then, at his wife's urging, followed by a flight simulator. A super-intelligent AI, he says, surely could have done better.

The author's claim is that superhuman hardware engineers are coming soon, perhaps very soon. Frontier models, far from superhuman today, can already generate CAD, design custom PCBs, source commercial-off-the-shelf (COTS) parts and describe how they should be assembled. He cites hardware design benchmarks, CAD Arena, EE Bench and his own LuxoBench, as showing steady improvements, and says you would have to be crazy to bet against the Bitter Lesson. Open questions: will such engineers beat all humans at once or only in some disciplines, and what will hardware engineering look like before, during and after the crossover? His answer starts from one observation: the interface between the digital and physical world is lossy. It is hard to "write" changes from bits into atoms and hard to "read" physical measurements back into a computer. Even once models can consistently one-shot electromechanical designs, he says, they will still have a lot more work to do.

On write access, he notes that when OpenAI released GPT-6 Astra, people posted 3D demos: Tom Krcha modeled a steam train, Thomas Ricouard made a walk-through of a photorealistic house, Anshu Chimala made a 3D game, and Sharif Shameem made the Palace of Fine Arts in San Francisco. Hardware engineers then tried engineering-grade CAD. Nim Cheema, co-founder of Dessn AI, asked Astra for a DJ controller; Roberto Nickson had top models design, engineer and 3D print bridges; and the author launched luxobench.com, asking OpenAI's Astra, Anthropic's Fable 5.1 and SpaceXSI's Grok 4.6 to help build an interactive desktop lamp. The finding: the models by themselves are still not quite ready for primetime hardware engineering.

He lays out five levels of real-world hardware. At 101, you decompose a product, pick COTS parts, check compatibility, design the interconnections, make sure it is buildable and buy the parts. Even this can stall elite teams: in Walter Isaacson's biography of Elon Musk, a 2006 Falcon 1 hot fire was delayed by a capacitor only in stock at a store roughly 6,000 miles from the launch pad in the Pacific. At 201, you need custom parts, so a CAD file must be translated through CAM software by a machinist; the author calls CNC machining an artisanal process that needs a master machinist to choose stock, workholding, tools and cutting strategy, which causes variability and errors in translation. At 301, parts are combined into electromechanical assemblies, tolerances "stack" so that parts that each meet their drawing can still fail to fit, and teams across a company must coordinate requirements and schedules. At 401, outside vendors enter; even SpaceX, the champion of vertical integration, has 591 current employees with supply chain or related titles and is hiring 128 more. At 501, you do it all at scale: building 1 of something differs fundamentally from building 1,000,000, with more part and process variability, earlier financing for machines and inventory, training or programming of assembly and test, and vendor limits.

These writing problems are, he says, fundamentally solvable but must be solved individually. Superhuman AI may be great at pure design, the platonically ideal part, but struggle with design for manufacturing, finding a design that meets requirements while staying easy to build and cheap. The first is relatively tractable and can largely stay inside a machine; the second needs far more real-world interaction and should therefore prove harder for models to master.

On read access, he tells of a friend who buys 45 million chickens per week. Despite identical specs, suppliers, process and machines, the friend's week-to-week variance in saleable food products after processing has historically been as high as 20%. Reducing it took extensive measurement of every supply chain step, vendor cajoling and a sophisticated data operation; he even found a deep-chain supplier injecting chickens with water, which passed intake weight checks but yielded less meat. The author gives five reasons the physical world is hard to read. It is non-uniform and constantly changing, so things break and drift unless inputs and outputs are measured. Its data is sparse: on satellite programs the thermal team would have wanted thermistors everywhere, but software, structures and production got a say. Data is expensive: precision needs ever more precise machines, such as slow, costly coordinate measuring machines. It has inherent latency: boards can only be thermally cycled so fast, and fatigue testing takes time. His favorite machine at Astranis was the HALT machine, nicknamed the shake and bake, which combines vibration with rapid thermal cycling, yet a HALT campaign is hard to compress because the full test must be rerun each cycle. And some information is qualitative, such as whether users like a product or how comfortable it is to hold, which is often the most incompressible step.

Data limits like these, he says, drive the spikiness of AI capability today: the more quantifiable and machine-readable a test result, the faster models climb the hill. How long those challenges can hold back the Bitter Lesson remains to be seen. The available text ends at the start of the section on hardware development in an AGI world, where the author begins by noting that hardware engineering is iterative, that the design step is digital and therefore very model-legible, and that we should expect AI models to become superhuman at it.

Key facts

The essay argues superhuman hardware engineers are coming soon, perhaps very soon, but that the digital-to-physical interface is lossy and will shape how and when they emerge.
Frontier models can already generate CAD, design custom PCBs, source COTS parts and describe assembly, though by themselves they are still not quite ready for primetime hardware engineering; CAD Arena, EE Bench and LuxoBench show steady improvements.
The author sorts real-world hardware into five levels (101 to 501), from picking COTS parts to building a million units, and says each writing problem is solvable but must be solved individually.
He argues pure design can largely stay inside a machine, while design for manufacturing needs more real-world interaction and should prove harder for models.
Reading the physical world is hard for five reasons: it is non-uniform and changing, data is sparse, data is expensive, processes have inherent latency, and some information is qualitative.

Why it matters

Most AI capability talk centers on software and digital output. This essay asks what happens when the work ends in physical objects. Its answer is that design is the easy, model-legible part, while getting designs made and measuring the results is where progress will be slower. The author expects superhuman hardware engineers soon, perhaps very soon, yet says that even models that can consistently one-shot electromechanical designs will still have a lot more work to do. He also ties the uneven pace of AI progress to data: the more quantifiable and machine-readable a test result, the faster models can climb the hill.

Who it affects

Hardware and electromechanical engineers, machinists, supply chain and manufacturing teams, and companies that build physical products, from hobbyists like the author to teams at SpaceX and Astranis. It also bears on the model makers whose systems are being tested on CAD and PCB design: OpenAI's Astra, Anthropic's Fable 5.1 and SpaceXSI's Grok 4.6 were all used in the author's lamp test.

How to use it

The essay offers a checklist for judging how far AI can take a hardware project. At each level, from 101 (picking and buying COTS parts) through 201 (custom parts), 301 (assemblies and tolerance stack-up), 401 (vendors) and 501 (volume production), ask where the digital design meets a physical step that cannot stay inside a computer. The author's benchmarks, CAD Arena, EE Bench and LuxoBench, are named as places to follow progress; he launched luxobench.com himself.

How solid is it

This is an opinion essay published on a16z.news, and its claims are the author's own, backed by anecdotes (his drone, the Falcon 1 capacitor story relayed from Walter Isaacson's biography, a friend's chicken purchasing, satellite and Astranis experience) rather than data. No benchmark scores are given for CAD Arena, EE Bench or LuxoBench. The author notes he runs LuxoBench, so he has a stake in one of the benchmarks he cites. The available text breaks off at the start of the section on hardware in an AGI world, so his final conclusions and any timeline are not visible. The page also carries an unrelated banner saying OpenAI has acquired Orchestra Bio.

Risks and caveats

The author says models alone are not yet ready for primetime hardware engineering, and that the design-for-manufacturing prediction is a "should", not a finding. He also leaves open how long data limits can hold back the Bitter Lesson. Reading his five reasons the real world is hard to measure as a fixed barrier would overstate him: he says that is for time to tell. The timing claim, coming soon, perhaps very soon, is given with no timescale.

“the interface between the digital and physical world is lossy”

— The author of the a16z.news essay (not named in the supplied text)