Newly launched AI startup Humans& raises $480M round backed by Nvidia, GV – News

Humans& Inc., a three-month-old artificial intelligence startup, today announced that it has closed a $480 million seed round at a $4.48 billion valuation.

The investment was led by SV Angel and Georges Harik, one of the company’s founders. Harik, an early Google LLC employee, helped develop several of the search giant’s core services. Google parent Alphabet Inc. participated in the round through its GV fund alongside Nvidia Corp., Jeff Bezos and more than a half-dozen others.

The initial Humans& team comprises about 20 AI experts. They joined the company from OpenAI Group PBC, Anthropic PBC, Meta Platforms Inc. and other major players in the AI market. Humans& says its researchers are developing neural networks that will make workers more productive.

According to the New York Times, the company’s models will speed up collaboration-related tasks and online research. They are also expected to automate “other tasks that suit machines.”

In a blog post, the company hinted that one of its goals is to equip its AI software with the ability to perform long-horizon activities. Those are complex tasks that take large language models hours or more to complete.

Long-horizon processing is a major focus of machine learning researchers. Google, for example, recently revealed that it has developed an AI model design specifically optimized for long-horizon use cases. The architecture’s flagship feature is a component called a metacontroller. When an AI model is given a long-horizon task, the metacontroller generates software modules that optimize the model’s reasoning workflow.

Humans& also plans to equip its software with support for multi-agent use cases. That means its AI models will be capable of collaborating with other neural networks on multistep tasks. Additionally, the company’s models will proactively ask workers for the information necessary to complete a given task.

Humans& plans to train its algorithms using reinforcement learning. That’s a training approach commonly used by researchers to develop reasoning models. In a reinforcement learning project, an AI completes a set of training tasks and receives positive or negative feedback depending on how well it completes a given chore. The algorithm uses that feedback to hone its output quality.

Unlike certain other development methods, reinforcement learning doesn’t require researchers to enrich their training data with explanatory labels, which lowers costs. However, it can still be capital-intensive because of the large number of graphics cards often needed to train reasoning models. That might be one of the reasons Humans& has raised such a large seed round.

The Times reported that the company plans to launch its first product early this year. 

Image: Unsplash