Former OpenAI and DeepMind researchers raise whopping $300M seed to automate science 

Periodic Labs got here out of stealth on Tuesday with a conflict chest of $300 million as a seed spherical, backed by a tech trade who’s who: Andreessen Horowitz, DST, Nvidia, Accel, Elad Gil, Jeff Dean, Eric Schmidt, and Jeff Bezos.

Periodic Labs was based by Ekin Dogus Cubuk and Liam Fedus. Cubuk led the supplies and chemistry workforce at Google Mind and DeepMind, the place considered one of his initiatives was, as an example, an AI software known as GNoME. That software found over 2 million new crystals in 2023, supplies that would sooner or later be used to energy new generations of know-how, researchers say.

Fedus is a former VP of Analysis at OpenAI, and one of many researchers who helped create ChatGPT. He additionally led the workforce that created the primary trillion-parameter neural community.

Its small workforce is likewise stuffed with researchers who’ve labored on different main AI and supplies science initiatives, from constructing OpenAI’s agent Operator to engaged on Microsoft’s MatterGen, an LLM supplies science discovery AI.

The aim of Periodic Labs is nothing lower than to automate scientific discovery, creating AI scientists, the corporate says. This implies constructing labs the place robots conduct bodily experiments, acquire information, iterate, and check out once more, studying and enhancing as they go.

The lab’s first aim is to invent new superconductors that it hopes carry out higher and probably require much less vitality than current superconducting supplies. However the well-funded startup additionally hopes to seek out different new supplies.

One other aim is to gather all of the bodily world information that its AI scientists produce as they combine and warmth and in any other case manipulate varied powers and uncooked supplies of their seek for one thing new.

“Till now, scientific AI advances have come from fashions educated on the web” and LLMs have “exhausted” the web as a supply that may be consumed, the corporate says in an introductory weblog publish. “[A]t Periodic, we’re constructing AI scientists and the autonomous laboratories for them to function.”

The hope is that, not solely will the labs invent next-generation supplies, however they may produce invaluable contemporary information that AI fashions can eat to proceed their evolution.

Whereas this is perhaps probably the most spectacular teams of researchers to assemble a startup for this function, it’s not the one one engaged on AI scientists. AI as a software to automate chemistry discoveries has been a subject of educational analysis since at the least 2023. It’s the pursuit of tiny startups like Tetsuwan Scientific, in addition to nonprofits like Future Home and the College of Toronto’s Acceleration Consortium.