Leading AI researchers are leaving major labs to start companies built around long-term research ideas. These new ventures, often called “neolabs,” are attracting large investments even when they have not released a model or paper. Their founders share a central question: can AI advance through approaches beyond building ever-larger language models?
Why build a new lab?
There is no official definition of a neolab. The term describes an independently founded, research-driven AI company organized around a technical bet, such as changing how AI learns, perceives the world, solves problems, is controlled, or computes.
The departures come as public claims about artificial general intelligence grow more confident. Nvidia CEO Jensen Huang said, “AGI has arrived,” while OpenAI president Greg Brockman welcomed humanity to the “AGI era.” The source notes that there is no agreed-upon definition of AGI, leaving the precise meaning of those claims unclear.
Some researchers who helped create current systems appear to question whether the established recipe is enough. That recipe relies on larger language models trained on internet data and refined with reinforcement learning. Their concern is not necessarily that progress will stop; it is that another approach may be needed to make further progress.
The reasons for leaving vary. Some founders want to pursue research without product pressure, while others see room for focused work that might receive little attention inside a large company. Departures can also follow strategic disagreements. And a new company may not be fully separate from established players: Google is a founding investor in Discovery Loop.
Investors are funding a chance to find what giants miss
The current wave of neolabs became visible in 2024, as the race for powerful models accelerated and major cloud companies backed leading AI firms. Ilya Sutskever’s Safe Superintelligence raised a billion dollars within months of its founding, despite having no product.
That kind of funding reflects a belief that large companies can struggle to protect exploratory research when product teams compete for resources and attention. Investors are wagering that small teams can pursue ideas the major labs overlook, whether because the path is uncertain or because it does not fit current priorities.
There are precedents for important work emerging in one place and reaching the public through another. Google invented the Transformer in 2017, and DeepMind developed AlphaGo. OpenAI, founded as a nonprofit in 2015, brought an AI assistant to a mass audience. Anthropic, founded by Dario and Daniela Amodei as an OpenAI spinoff, became another breakout company.
But an early lead may be brief. TypeSafe AI launched its Jev decision model, and two weeks later OpenAI announced its Decisions API. A short window can still demonstrate expertise and create opportunities: the source describes major technology companies hiring teams or founders from smaller AI ventures, as well as acquisition interest.
Six research bets target limits in today’s AI
The source groups 27 labs by their main research bet and also profiles nearly 20 more in brief. Across the landscape, the ideas focus on limitations in existing AI and on possibilities that large model builders may not prioritize.
One gap is continual learning. Today’s models can use new information during a conversation, but typically do not update their weights after deployment. They also need large amounts of data to acquire skills. Six labs pursue ways to improve learning: some focus on adapting models while they are in use, some on learning from fewer examples, and others on learning through experience.
Adaption, founded by Sara Hooker and Sudip Roy, argues that more compute produces better models less reliably than before. Hooker has also argued that research can win because it fits available hardware, rather than because it is inherently better. Adaption’s stated aim is efficient AI that learns continuously, with models adapting to users on the fly.
Its initial tools focus on data: cleaning and visualizing it, along with automated training through AutoScientist. The source says the longer-term goal—a model that learns during deployment—has not yet been achieved. That gap illustrates the distance between a lab’s research ambition and a working product.
The test is whether a distinct idea can endure
Small labs face a difficult market. Established companies have more compute, established data pipelines, and feedback from hundreds of millions of users. Once a small lab releases something useful, a market leader may be able to offer a similar capability soon afterward.
That does not make the research bet pointless. A brief lead can show that a team has valuable expertise, and a major company may hire or acquire that team. But a neolab’s longer-term case depends on building something customers—or the big labs themselves—will pay for.
The founders’ bets are therefore about more than whether AI will become more capable. They are wagers about where the next useful capability will come from, who will build it, and whether a new company can turn a research idea into lasting value before larger competitors catch up.