Ex-OpenAI Researcher’s Startup Targets Up to $1 Billion in Funding to Develop a New Type of AI
The Takeaway
- Startup led by former top OpenAI researcher seeks up to $1 billion in funding
- Core Automation plans new AI models with continual learning
- Startup adds to number of neolabs seeking to overhaul core processes in AI development
The wave of so-called AI neolabs seeking breakthroughs they think incumbents like OpenAI will miss continues to rise.
Core Automation, an AI startup founded by former senior OpenAI researcher Jerry Tworek in recent weeks, aims to raise between $500 million and $1 billion in funding, according to a person who has spoken with the company.
He wants to develop AI models using methods that major firms such as OpenAI and Anthropic aren’t heavily focused on, according to materials shown to potential investors, and to create AI models that can learn on the fly from real-world experience, a capability known as continual learning, which today’s models cannot do, the person said.
The researcher’s plans for the new startup are early and his funding aims and product approach could change.
Tworek represents a small but growing group of AI researchers who believe the field needs an overhaul because today’s most popular model development techniques seem unlikely to be able to develop advanced AI that can achieve major breakthroughs in biology, medicine and other fields while also managing to avoid silly mistakes.
Tworek left OpenAI earlier this month to “try and explore types of research that are hard to do at OpenAI,” he wrote in a post on X. The materials shared with investors said Core Automation intends to use large neural networks, the brain-like math that underlies most frontier AI models. But the company will rethink most of how those models are developed, “up to and including gradient descent,” the standard method for training neural networks.
Tworek wants to develop models that require less data and fewer servers to train, the person said. It would do this by coming up with new model architectures beyond the transformer, which underlies today’s most popular models, according to the materials. Tworek also wants to merge the different steps of model training into a single process, this person said.
By pursuing continual learning, Core Automation seems to be taking a similar approach to Safe Superintelligence, another AI lab cofounded by former OpenAI chief scientist Ilya Sutskever. Sutskever has similarly said that he wants to develop models that can learn from being deployed in the real world.
To be sure, major AI developers such as OpenAI and Anthropic are also interested in continual learning techniques. Some AI researchers believe they can tweak transformer-based AI models so they exhibit such learning attributes without a full-fledged overhaul of the model’s architecture.
Tworek’s ambitious fundraising goals highlight investors’ continued interest in backing neolabs, some of which have raised hundreds of millions of dollars in recent months, despite often lacking any revenue or a product.
These startups pursuing new approaches to AI include Humans&, which this month landed $480 million in seed funding at a $4.48 billion valuation from investors like SV Angel, Nvidia and Jeff Bezos, as well as Mira Murati’s Thinking Machines Lab, which recently was in talks to raise between $4 billion and $5 billion at a $50 billion-plus valuation. Thinking Machines Lab, however, last year released a product for customizing models and is generating some revenue.
Tworek, who joined OpenAI in 2019, envisions the startup’s research scientists working on a single model named Ceres, the name of a dwarf planet and Roman goddess, with a single algorithm. That’s a change from how bigger AI developers handle training, which usually takes place in stages: pre-training the model on broad swathes of the internet before mid-training and post-training on more curated data in fields like coding, medicine and math.
After developing the model—which Tworek aims to need 100 times less data than state-of-the-art models, the company will develop an AI agent to automate development of the company’s products. He envisions the company’s future product will first work on industrial automation, eventually building “self-replicating factories” and potentially building biomachines to automatically create custom designs—or even terraform planets, the materials said.
At OpenAI, Tworek was a vice president of research who led the company’s work with reinforcement learning, a popular model training technique that rewards a model for accomplishing certain goals and penalizes it for other behaviors. He also was a key contributor to the company’s reasoning models, coding tools and agents.