Exclusive: What If AI could design a jet engine—or even a starship? Google DeepMind and Airbus veterans just raised $23M with an eye on that future

When you dream of the day of artificial intelligence reaches a human capacity, the former CTO Airbus Paul Eremenko says that he has always done so in the context of the construction of real world machines. “I want a superintendent of AI which can build us vessels and spheres of Dyson,” he said Fortune– The latter being a hypothetical megastructure of science fiction which would exploit the energy of a star.

While his dream is still far away, Eremenko lays the foundations. He united his forces with the former researcher from Google Deepmind, Aleksa Gordic, and Adam Nagel, an engineering manager before at Acubed, Airbus's Innovation Center. Together, they founded P-1 AI, which came out of Stealth today with a 23 million dollars seeds led by Radical Ventures. The other investors include Village Global, Schematic Ventures and Lerrer Hippeau, as well as notable angels such as the chief scientist of Google Deepmind, Jeff Dean, and the vice-president of explorations of new Openai products, Peter Wellinder.

P-1, named after P-1 adolescenceA science fiction novel from 1977 by Thomas Joseph Ryan about a sensitive AI, develops an engineering assistant fed by Ai called Archie. Similar to other AI assistants such as the coding of the AI ​​cognition AI AI, the idea is to integrate Archie as a junior member of each engineering team – to manage repetitive but suspended tasks in the interpretation of requirements, generate early design concepts and check compliance with regulations. This is an early step towards a much more ambitious vision: using AI to finally design the complex machines of the future.

Eremenko said he was surprised that no one is already working on this goal, but he quickly understood why. As with autonomous cars and robots, AI teaching to build machines requires a huge amount of training data. The key, he explained, is to simulate realistic engineering systems by building virtual models of real world components, such as engines, pipes and trees. Then, these simulations based on physics are combined in various configurations to generate data, which is then used to form AI models which help automate engineering design.

According to Gordic, this is similar to the way Google Deepmind used games to help train Alphago, the AI ​​that beat human champions at Go, a complex strategy board game. “Alphago was initially trained to imitate the data of real human actors,” he said Fortune. Now, it will lead to training and large -language models (LLM) and other AI systems to understand and modify complex engineering conceptions in rich physics systems such as cooling the data center or HVAC systems.

To go beyond the capacities of the “glorified automatic entry” of LLM as Chatgpt, he explained that models must be useful for engineering tasks. AI must therefore really understand orders and follow the instructions. The powerful combination of AI models which are formed on synthetic data built on physical simulations and which can then understand and act on these data made of really automated engineering a reality. “We form Archie on synthetic data to give it a sort of level of university engineer,” continued Eremenko. But after deployment, Archie can learn human comments and real world data from companies using AI.

P-1 investors said that Eremenko are interested in the starting short-term levels of the startup, but they are particularly enthusiastic about the future. “Many of us in the world of engineering and AI, we have grown up in science fiction, and science fiction has promised us a great intelligence that will build vessels,” he explained.

Large holders like Autodesk, Siemens and IBM working towards elements of the use of AI for engineering, but they do not create a new class of generalist engineering assistants, and they do not reach the same great vision of the machines built by AI.

However, Eremenko and Gordic insist that their own is a very realistic and concentrated path, and it is not purely a research project with an indefinite period. “We are not going to be a 10 -year -old bithot,” said Eremenko. “This is a very pragmatic deployment and market path.”

This story was initially presented on Fortune.com