Former xAI Researcher Raises Money for Human-Centered Artificial Intelligence

Eric Zelikman, who previously worked as a researcher at xAI, is now seeking to raise $1 billion for his new startup, Humans&. The company is discussing a $5 billion valuation, six sources told Forbes. Its founders also include Georges Harik, one of Google’s earliest employees, and other experts from Meta, Anthropic, OpenAI, and DeepMind. Humans& focuses on training artificial intelligence models to interact with people more effectively.

The Founders and Their Experience

Eric Zelikman studied computer science as a doctoral student at Stanford and wrote the first paper on training language models to reason in natural language. He then worked at xAI on data collection for pre-training, reasoning, and agent infrastructure. Georges Harik was Google’s seventh employee, co-created AdWords and AdSense, and now invests in startups.

Noah Goodman, Zelikman’s doctoral advisor and a professor of computer science and psychology at Stanford, worked on the Gemini model’s post-training team. At Anthropic, Andi Peng focused on post-training and behavioral reinforcement learning. Ray Ramadorai designed systems for large data centers at Microsoft.

The Lab’s Goals and Focus

Humans& plans to develop artificial intelligence models that remember and respond to a person’s preferences and interests. The aim is for artificial intelligence to empower people, unlike other labs that focus on replacing them. Zelikman explained this to investor Sarah Guo on the No Priors podcast: "I personally firmly believe that many fundamental human problems are more likely to be solved by models that can collaborate exceptionally well with large groups of people, understand their goals, ambitions, values, and weaknesses, and coordinate them so that everyone can be more effective."

The company told investors that its new approach to model training will require more computing power than current strategies. The exact product is not yet clear, but the emphasis is on ensuring that artificial intelligence collaborates with and empowers people.

Operations and Challenges

Humans& operates as a startup developing new methods for training artificial intelligence models focused on collaboration. Instead of building applications on existing models such as GPT-5 or Claude Opus 4.1, it focuses on foundational training, which requires enormous amounts of money and computing power. This is why it ranks among the small number of labs raising large sums before releasing a product, similar to Mira Murati’s Thinking Machines Lab or Ilya Sutskever’s Safe Superintelligence.

One investor who declined the offer said the amount was too high for a company at such an early stage. Humans& tells investors that its approach should lead to models that understand human weaknesses and coordinate groups to achieve greater efficiency.

Zelikman wrote on his website: "If we bet everything on autonomy and nothing on collaboration, all intelligence and no emotional intelligence, the future will be a colder place." This idea underscores Humans&’s focus on models that not only reason but also interact with respect for human values.

Source: forbes.com