A thousand airplanes will never train an AI engineer

Sequoia's Sonya Huang and Pat Grady introduce Paul Eremenko as a former DARPA director, the youngest CTO in Airbus history at 35, and now CEO of P-1 AI. The company is trying to build engineering AGI for the physical world: data-center coolers, airplanes, eventually starships. Coding already has Anthropic, Cursor, and Devin. Hardware, they note, has not had that transformation.

The timing of the conversation is not accidental. Sequoia has just run its AI Ascent conference, where Jeff Dean talked about vibe coding and a 24/7 junior software engineer arriving within a year or so. Huang asks the obvious follow-up: what is happening on the physical-engineering side of that same curve? Eremenko's answer is short. Not a lot.

He grew up on hard science fiction that promised AI would help build the world around us, and eventually starships and Dyson spheres. When the deep-learning revolution took off, he asked who was actually doing that work. Nobody was, and it was not on the agenda of the foundation labs. In 2025, he says, it still isn't. Jeff Dean is an angel investor in P-1. Eremenko wants the same discontinuity for metal and thermodynamics that coding AI is already delivering for software.

One of his co-founders, Susmit Jha, wrote a 2011 PhD on program synthesis, a technology that waited more than a decade for packaging, pricing, and a business model. P-1 intends to stand on that coding-AI work: if a physical system can be represented as a program, program-synthesis techniques can generate physical designs. Eremenko does not think this takes ten or fifteen years. The technology bricks, he says, can be stacked this year, with product-market fit possible as early as next year.

The biggest missing brick is training data. Ask a model what happens if you increase the wing area of an A320 by 10%, and it ideally needs to have seen millions of airplane designs. There have not been millions of airplanes designed since the Wright brothers. Even in the fantasy where you had every historical file, they would not be modeled in a coherent, semantically integrated way. Hypothetically you would have maybe a thousand designs. Nowhere near enough to train a large model.

So P-1 builds the corpus itself: a synthetic, physics-based, supply-chain-informed set of hypothetical designs in whatever product domain it is targeting. The design space for most physical products is almost infinitely large. You cannot sample it randomly, and you cannot sample it evenly. You have to sample it cleverly.

The rule Eremenko keeps returning to is density. Sample densely around dominant designs. Sample sparsely around the corners and edges of the design space. Even if that corner is nowhere you would ever want to go, it teaches the model *why* you would not. Throw a million airplane designs and a performance vector for each into an LLM, though, and you still will not magically get a good engineer. Architecture matters as much as volume.

The product is narrowly aimed at cognitive automation: what a human engineer does when designing a physical system. Humans are good at taking a pile of requirements, distilling the key design drivers, posting one or more candidate solutions, and doing first-order sizing. They are also good at picking which physics actually matter. These are multiphysics systems, with electrical, thermal, vibration, and electromagnetic-interference effects that sometimes close the design and sometimes do not. Good engineers know which modalities to keep.

They also know which detailed-design and analysis tools exist, the range of applicability of each, and how to set the problem up. That is the layer P-1 wants. A really important consequence, Eremenko says, is that the company is not playing at the tools layer. Existing CAD, simulation, and analysis packages stay. Archie, the agent, should know they are there and how far they are valid, the same way a human does. It does not replace them, improve them, or compete with them.

Every one of those human steps collapses into a few primitive operations. Design evaluation: given a design, what is its performance, modeling the relevant phenomena. Design synthesis: given a requirements vector, what is the design. A third class is finding errors and infilling inside a design. Any engineering query a human handles, Eremenko argues, reduces to some sequence of those operations. An orchestrator-reasoner LLM takes tasking from people in an organization and decomposes it into the right sequence. Specialists, some neural and some not, then carry the operations out.

Behind that orchestrator sits a graph neural network used as a physics-based surrogate over the performance space, plus a geometric reasoner for relative positioning, packing, and interference. Easy geometric questions can be software 1.0. Harder ones can go to vision-language models. Some higher-order spatial tasks, he admits, still need a coming generation of physical world models. For multiphysics reasoning they also run what he calls a lobotomized LLM: no longer good at English, but very good at programmatic representations of physical designs and at reasoning over them. The whole federation is orchestrated by the same LLM that talks to the user.

P-1 is about nine months old. The pre-seed demo was a toy: residential cooling, air-conditioning units and the like. The domain is rich enough to be convincing, with fluid flow, airflow, thermal coupling, and electrical systems, but the part count is small and a lot of the physics is linearizable. That keeps data generation from swallowing the team. The demo is public. The honest question is how good it is, and other than a vibe test where a human pokes it and says it feels pretty good, there is no accepted answer.

So the company has invested heavily in evals for physical-engineering AIs. By the time the episode airs, Eremenko expects an arXiv paper describing the approach. They call it Archie IQ. The plan is to administer the same eval to an entry-level human engineer, an average one, an expert, and to Archie, then close the loop so Archie climbs that IQ scale.

Residential cooling was never the first market. The first customer deployment is data-center cooling: still a thermodynamic engine, an order of magnitude more complex, and a market that cannot keep up with demand. Cooling systems, Eremenko says, have become the long-lead item pacing data-center development. Delivery is limited by engineering bandwidth for semi-custom solutions. The systems have on the order of a thousand unique parts. The physics is still fairly linearizable, which makes synthetic data manageable, and the customer base is enthusiastic.

From there the map is complexity. The company expects to go roughly an order of magnitude up in product complexity every year. Next is industrial systems that go into a factory: material handling, industrial robots, mills, lathes. Then mobility, which could be automotive, agriculture, or mining equipment. Then aerospace and defense. Data-center cooling is about a thousand unique parts. An airplane is about a million. That is three orders of magnitude, and current projections say roughly one year per order of magnitude.

Non-proprietary synthetic data, Eremenko thinks, can get Archie to entry-level: college-educated, not yet savvy about a specific company's products, processes, or detailed supply chain. The goal is to get him hired. After a data-sharing agreement, Archie can learn behind the firewall: the customer's PLM system, model-based tools, real-world performance, quality escapes. Expertise should then move from entry-level to average to expert, helped along by better underlying models.

There is no generally agreed definition of AGI, so P-1 borrowed Bloom's taxonomy, a cognitive ladder for human learning from the 1950s, already applied to LLMs, and adapted it to engineering. At the bottom is recall. Then semantic understanding: what does this part do. Then evaluation: what is the performance impact of swapping a component or resizing something. Then finding mistakes, the error-correction and infilling step. Then synthesizing a brand-new design or a significant change to an existing one.

The pinnacle, which they call EAGI, engineering AGI, is reflection. Some degree of self-awareness: what process did I just use for the five levels below, what are its limitations, is there an alternative, where could I have gone wrong. Most working engineers, Eremenko says, do not do this well. It is reserved for senior experts and technical fellows in large industrial companies. That self-awareness of the engineering process, including its limits, is the pinnacle of human engineering intelligence as they define it.

A second axis is generalization: can the system switch domains without being trained on the new one. You could call AGI a win on either axis, or on both. P-1 wants both. Getting there, he insists, does not require research breakthroughs. The company sits in applied research, taking what frontier labs already publish and pointing it at this problem. The binding constraints are compute: CPU to generate synthetic data through simulation and sampling, GPU to train, GPU to infer. Today's compute, he does not think, would cover a million-part system.

Synthetic data starts with a catalog of component models and rules for composing them. That catalog needs to be a couple of orders of magnitude bigger than a typical system. A million unique parts in a product implies a catalog of a hundred million or a billion parts. Today they build it by hand, while also writing automation and AI tools to help. Then the components have to be assembled intelligently. It is not a tornado going through a junkyard and assembling a 747. Each assembled design is simulated to produce a performance vector. That vector is the training example.

Supply-chain-informed means every catalog entry either reflects a real part or is a hypothetical the team knows how it would obtain: a new motor, a new compressor. Innovation is not only recombination of what exists. Physics-based means the composition rules model the relevant modalities of interaction, and that the designs produced are realizable. Grady then asks him to put his old Airbus-CTO hat back on. What do all those engineers actually do, and what does life look like before and after engineering AGI?

The engineer's loop, Eremenko says, is fractal. It repeats from the system down through subsystems, assemblies, and widgets, and the organization is designed to mirror the product. That is why Archie is positioned as an agent that joins a team, not an assistant for an individual. One mission statement is an Archie on every team in every major industrial company in the world. The commercial point is to sell work, not software. Selling engineering software into a company like Airbus is brutally hard. There are hundreds if not thousands of tools, connected by intricate and sometimes inelegant glueware, and a new tool is a nightmare to introduce.

On top of that, the labor budget at these companies is much bigger than the methods-and-tools software budget. Tackle labor. Archie shows up on Slack or Teams as a remote junior, with no embodiment and almost no process change. Task him the way you would an engineer at an offshore center. The name is the letter A, so a Bob, a Charlotte, and a Daniel can follow, with echoes of Archimedes and architect. The first use case is product customization, what the data-center cooling world calls specials: taking a platform and bending it to a customer's architecture, function, and building codes, which is where most engineering hours go. Airbus rarely does a clean-sheet airplane. It does derivatives and head-of-variant configs for a particular airline.

Huang asks what an ordinary person will notice once millions of Archies exist. Eremenko will not forecast more than about three years in steeply exponential times. In the first instance, Archie makes existing teams more productive. If every team has one, the Archies may coordinate among themselves better than the humans between teams, speaking their own shorthand. For the average person, the impact is lower-cost goods. Perhaps, he allows when Huang jokes about it, you can buy an airplane. The really interesting work starts when Archie can design things humans cannot: the starships, Dyson spheres, and Matrioshka brains from the books. That is the north star. In the meantime they want a pragmatic, profitable business. The near-term goal is Archies as 10% of the workforce, doing the boring, dull, repetitive jobs.

Sequoia partner Konstantine Buhler talks about a stochastic mindset: classical software is predetermined, models are not. Huang puts the fear plainly. Vibe-coding a web app that breaks is annoying. Vibe-coding an airplane that breaks is a disaster. Eremenko's reply is that humans are pretty stochastic too. A junior engineer will make mistakes, may not do the right thing, and may not be repeatable. The number they need to quantify in this year's pilot is Archie's error rate. Existing organizations already have layers of review, milestones, and tests so that a junior's mistake does not bring down an airplane. If Archie's error rate is comparable or better, slotting in should be seamless.

In a lightning round he says physical AI is getting close, meaning robotics and foundation models that ingest real-world sensor data, both of them important building blocks for P-1. Humanoids get a yes, on the same logic as Archie: they slot into existing environments even if they are not the optimal configuration. Everyone in AI, he thinks, should reread Asimov's robot series, because the laws of robotics were carefully thought out and need to be built somehow very deeply into these models for alignment.

The startups he admires work on models that ingest physical-world data. Neural nets are a rough copy of neurons, but a human baby starts with touch, taste, and hearing, then vision, then language, then higher-order engineering and spatial reasoning. Deep learning went language-first because the internet was there to scrape, then added video and imagery, and skipped touch. Touch, he argues, is particularly important for spatial reasoning about three-dimensional objects. Archetype, founded by a former Google colleague, has a sensor-data foundation model that can already infer some of the underlying physics. Language and vision alone will not get you engineering AGI. His own favorite recent AI app is not ChatGPT or Cursor but a fully generated two-minute Archie biopic from EyeMix, an Israeli-LA studio, delivered in two weeks at about a fiftieth of traditional cost.