AI chatbots can answer questions and summarize documents, but they're still designed for one user at a time. Now a new startup called Humans& is betting that coordination, not conversation, is the next major frontier for artificial intelligence. Founded by alumni from Anthropic, Meta, OpenAI, xAI, and Google DeepMind, the three-month-old company just raised a $480 million seed round to build foundation models designed for social intelligence and multi-user collaboration - not just information retrieval.
Humans& just became one of the most well-funded AI startups you've never heard of. The three-month-old company pulled in a staggering $480 million seed round this week, and it doesn't even have a product yet. What it does have is a star-studded founding team from Anthropic, Meta, OpenAI, xAI, and Google DeepMind - and a bold thesis that the entire AI industry is about to pivot from chatbots to coordination systems.
"It feels like we're ending the first paradigm of scaling, where question-answering models were trained to be very smart at particular verticals, and now we're entering what we believe to be the second wave of adoption," Andi Peng, co-founder and former Anthropic employee, told TechCrunch. The average user is trying to figure out what to do with all these AI tools, she said, and that's where Humans& comes in.
The pitch is simple but ambitious: current AI models are optimized for single-user tasks like answering questions or generating code. They're not built to handle the messier, more valuable work of coordinating teams, tracking decisions across time, or balancing competing priorities among multiple stakeholders. Humans& wants to build a "central nervous system" for organizations - a foundation model trained specifically for social intelligence and group collaboration.
CEO Eric Zelikman, a former xAI researcher, pointed to something as mundane as choosing a company logo to illustrate the problem. "When you have to make a large group decision, often it comes down to someone taking everyone into one room, getting everyone to express their different camps," he told TechCrunch, laughing as his team recalled the tedious process of getting everyone aligned on their own branding.
That's the kind of coordination work that today's AI can't touch. And it's not because the models aren't smart enough - it's because they're trained for the wrong task. Chatbots are optimized for two things, according to Zelikman: how much a user immediately likes a response, and how likely the model is to answer a question correctly. Neither of those objectives teaches an AI how to ask good questions, remember context over weeks or months, or help a team reach consensus.
To fix that, Humans& is rethinking how foundation models are trained. Yuchen He, a co-founder and former OpenAI researcher, said the company is using long-horizon and multi-agent reinforcement learning - techniques designed to train models for environments where multiple humans and AIs interact over extended periods. "We're trying to train the model in a different way that will involve more humans and AIs interacting and collaborating together," He told TechCrunch.
Long-horizon reinforcement learning trains models to plan, act, revise, and follow through over time rather than just generate a single good answer. Multi-agent RL trains for scenarios where multiple AIs or humans are in the loop simultaneously. Both approaches are gaining traction in recent academic research as AI labs push beyond chatbot responses toward systems that can coordinate actions and optimize outcomes across many steps.
"The model needs to remember things about itself, about you, and the better its memory, the better its user understanding," He said. That's a sharp departure from current models, which treat each conversation as largely isolated.
Humans& is designing its product and model in tandem, which partly explains why the team has been vague about what they're actually building. Peng said the interface and the model's capabilities will "co-evolve" as training progresses. The team hinted that the product could replace tools like Slack for communication or Google Docs and Notion for collaboration, with applications spanning both enterprise and consumer use cases.
What's clear is that Humans& isn't trying to build a plugin for existing collaboration tools. The startup wants to own the collaboration layer itself - becoming the connective tissue across organizations that understands each person's skills, motivations, and needs, then balances those for the collective good.
That ambition puts Humans& on a collision course with some of the biggest names in tech. Anthropic is rolling out Claude Cowork to optimize team collaboration. Google's Gemini is embedded directly into Workspace, enabling AI-driven collaboration inside tools people already use daily. And OpenAI has been pitching developers hard on multi-agent orchestration and workflow automation.
The timing might actually work in Humans&'s favor. LinkedIn founder Reid Hoffman argued this week that companies are implementing AI wrong by treating it like isolated pilot projects. The real leverage, he wrote on social media, is in the coordination layer - how teams share knowledge, run meetings, and make decisions. "AI lives at the workflow level, and the people closest to the work know where the friction actually is," Hoffman said.
That's exactly the space where Humans& is planting its flag. The AI collaboration and productivity space is heating up fast. Granola, an AI note-taking app, raised $43 million at a $250 million valuation as it added collaborative features. But Granola is augmenting existing workflows. Humans& wants to replace them entirely.
That's a much heavier lift, and it comes with significant risks. Training and scaling a new foundation model requires enormous capital and compute resources - both of which are increasingly scarce. Humans& will be competing with Meta, OpenAI, and Google DeepMind not just for market share but for GPUs and top-tier AI talent.
There's also the existential question of whether Humans& can move fast enough. The major AI labs are already working on collaboration features, and none of them appear to be retraining models from scratch around social intelligence. That could give Humans& a structural advantage - or it could make the startup an attractive acquisition target. With Meta, OpenAI, and DeepMind actively hunting for AI talent, M&A is a real threat.
But Zelikman insists acquisition isn't on the table. The team has already turned away interested parties, he told TechCrunch. "We believe this is going to be a generational company, and we think that this has the potential to fundamentally change the future of how we interact with these models," he said. "We trust ourselves to do that, and we have a lot of faith in the team that we've assembled here."
It's a bold bet - one that hinges on the idea that AI's next era won't be about smarter chatbots but about models that can navigate the social complexity of real human work. If Humans& pulls it off, the $480 million seed round might look cheap in hindsight. If not, the company will become a cautionary tale about raising too much, too soon, on a vision alone.
Humans& is making a generational bet that AI's value will shift from answering questions to coordinating people. With $480 million in backing and a founding team pulled from the top AI labs in the world, the startup has the resources and talent to take a real swing at the problem. But it's entering a market where giants like Anthropic, Google, and OpenAI are already moving, and it's doing so without a product or clear go-to-market strategy. The next 12 months will reveal whether training models for social intelligence is truly the next frontier - or whether collaboration is just another feature the incumbents will absorb. Either way, the industry is watching closely.
