AI's Next Leap: Mastering Human-Like Coordination

Artificial intelligence chatbots have made remarkable strides in recent years, demonstrating proficiency in tasks like answering complex questions, summarizing lengthy documents, and even solving mathematical equations. However, these advancements largely position AI as a sophisticated assistant for individual users. A critical gap remains: the ability to effectively manage the complexities of real-world collaboration – coordinating individuals with differing priorities, tracking decisions over extended periods, and maintaining team alignment. This is where the next major frontier for foundation models lies.

Humans&: A New Approach to AI Collaboration

Enter Humans&, a newly established startup founded by veterans from leading AI organizations including Anthropic, Meta, OpenAI, xAI, and Google DeepMind. The company recently secured a substantial $480 million seed round to build what they describe as a “central nervous system” for the burgeoning human-plus-AI economy. While initial coverage has focused on ‘AI for empowering humans,’ Humans&’s core ambition is far more innovative: developing a novel foundation model architecture specifically designed for social intelligence, moving beyond mere information retrieval or code generation.

“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 where the average consumer or user is trying to figure out what to do with all these things,” explained Andi Peng, a Humans& co-founder and former Anthropic employee, in an interview with GearTech.

The Shift from Chat to Agents and the Coordination Challenge

Humans&’s vision centers on guiding individuals through this new era of AI, shifting the narrative away from job displacement and towards augmentation. Regardless of whether this is a strategic marketing approach, the timing is crucial. Companies are actively transitioning from simple chatbots to more sophisticated AI agents. While these models are becoming increasingly competent, the underlying workflows often lag behind, and the fundamental challenge of coordination remains largely unaddressed. Simultaneously, many individuals feel threatened and overwhelmed by the rapid advancements in AI.

The three-month-old company, like many of its peers, has attracted significant investment based on this philosophy and the impressive credentials of its founding team. Currently, Humans& is focused on product development and hasn’t yet revealed specific details, though the team suggests potential applications as replacements for existing multi-player or multi-user platforms like communication tools (such as Slack) or collaboration platforms (like Google Docs and Notion). They are exploring both enterprise and consumer applications.

“We are building a product and a model that is centered on communication and collaboration,” stated Eric Zelikman, co-founder and CEO of Humans& and a former xAI researcher, to GearTech. “Our primary focus is to create a product that helps people work together and communicate more effectively – both with each other and with AI tools.”

The Pain Point of Group Decision-Making

Zelikman illustrated the problem with a relatable example: “When you have to make a large group decision, it often devolves into a lengthy process of gathering everyone in a room and navigating differing opinions, like choosing a logo for the startup.” He and his team chuckled as they recalled the time-consuming and often frustrating experience of reaching a consensus.

Training AI for Empathetic Interaction

Zelikman emphasized that the new model will be trained to engage in conversations that feel natural and empathetic, mimicking interactions with a friend or colleague – someone genuinely interested in understanding your perspective. Current chatbots often bombard users with questions without considering the value or relevance of those inquiries. This stems from their optimization for two primary metrics: immediate user satisfaction with responses and the accuracy of answering the posed question.

The lack of concrete product details may be due to Humans&’s iterative approach, designing the product in tandem with the model’s development. “Part of what we’re doing here is also making sure that as the model improves, we’re able to co-evolve the interface and the behaviors that the model is capable of into a product that makes sense,” Peng explained.

Owning the Collaboration Layer

It’s clear that Humans& isn’t aiming to create a model that simply integrates with existing applications. Their ambition is to own the collaboration layer – the connective tissue that facilitates seamless interaction between individuals and AI.

The intersection of AI and team collaboration/productivity tools is rapidly gaining traction. Startups like Granola, an AI note-taking app, recently raised a $43 million round at a $250 million valuation as it expands its collaborative features. Prominent figures are also increasingly framing the next phase of AI as one of coordination and collaboration, rather than solely automation. LinkedIn founder Reid Hoffman recently argued that companies are misimplementing AI by treating it as isolated experiments, and that the true potential lies in the coordination layer of work – how teams share knowledge and conduct meetings.

“AI lives at the workflow level, and the people closest to the work know where the friction actually is,” Hoffman wrote on social media. “They’re the ones who will discover what should be automated, compressed, or totally redesigned.”

Humans& intends to occupy this space, positioning its model-slash-product as the “connective tissue” across any organization – from large corporations to families – understanding the skills, motivations, and needs of each member, and balancing those for the collective good.

Rethinking AI Model Training: Long-Horizon and Multi-Agent RL

Achieving this requires a fundamental shift in how AI models are trained. “We’re trying to train the model in a different way that will involve more humans and AIs interacting and collaborating together,” stated Yuchen He, a Humans& co-founder and former OpenAI researcher, to GearTech. The startup’s model will leverage long-horizon and multi-agent reinforcement learning (RL).

• Long-horizon RL trains the model to plan, act, revise, and follow through over extended periods, rather than generating a single, isolated response.
• Multi-agent RL trains the model for environments involving multiple AIs and/or humans.

Both concepts are gaining momentum in academic research as researchers push Large Language Models (LLMs) beyond chatbot responses towards systems capable of coordinating actions and optimizing outcomes over multiple steps. “The model needs to remember things about itself, about you, and the better its memory, the better its user understanding,” He added.

Challenges and Competition

Despite the impressive team behind Humans&, significant risks lie ahead. Training and scaling a new model requires substantial and ongoing financial investment, placing Humans& in direct competition with established industry giants for resources, including access to computing power.

However, the biggest challenge is that Humans& isn’t just competing with platforms like Notion and Slack; it’s challenging the dominant players in AI. These companies are actively developing enhanced collaboration features within their existing platforms, even while speculating about the potential for Artificial General Intelligence (AGI) to eventually render much of today’s work obsolete. Anthropic, through Claude Cowork, aims to optimize work-style collaboration; Gemini is integrated into Workspace, enabling AI-powered collaboration within familiar tools; and OpenAI is actively promoting its multi-agent orchestration and workflow capabilities to developers.

Crucially, none of these major players appear to be prioritizing a complete rewrite of a model based on social intelligence, which either provides Humans& with a competitive advantage or positions it as an attractive acquisition target. Given the aggressive talent acquisition strategies of companies like Meta, OpenAI, and DeepMind, M&A remains a possibility.

Humans& has indicated to GearTech that it has already declined acquisition offers and remains committed to its independent vision. “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,” Zelikman concluded. “We trust ourselves to do that, and we have a lot of faith in the team that we’ve assembled here.”