Here’s what’s on our plate today:

• 🤝 Humans& and the $480M bet on collaborative AI.

• 🧠 Brain snack for builders on designing truly collaborative agents.

• 📰 Quick bits on AI quote scandal, trash tech, ChatGPT backlash.

• 📊 Poll on how you want AI to work with you.

Let’s dive in. No floaties needed…

The Laboratory

Why investors are betting $480M on humans&’s vision of collaborative AI

TL;DR

• Wide open, fiercely crowded: No dominant platform exists for human-AI collaboration at enterprise scale, but Microsoft, Google, Salesforce, and Slack are all racing to embed AI agents into their own tools.
  

• The big bet: Humans& argues a coordination model purpose-built with reinforcement learning is fundamentally different from bolting memory and agent features onto a general-purpose LLM.
  

• The startup vs. incumbent dilemma: Will enterprises adopt a new collaboration platform from a startup, or wait for similar capabilities to show up in tools they already pay for?
  

• Stakes beyond the product: Success could reshape how future AI models are designed and adopted. Failure would underscore just how hard it is to engineer the thing humans do instinctively: collaborate.

A generally understood fact about humans is that they are social animals. Born of the necessity to survive in a hostile environment, the human tendency to collaborate has enabled the species to achieve remarkable, sustained growth. In humanity’s collaborative efforts, technology has played an important role.

Before the advent of the internet, communication was slow and constrained by geography. However, as technology has evolved, the global population has access to infinite, immediate knowledge from around the world. Now, as technology continues to evolve, humanity faces a choice: should artificial intelligence be developed to foster greater collaboration, or to exacerbate the loneliness epidemic gripping much of the Western world?

Currently, most AI tools are designed to improve productivity and enhance human abilities; however, without proper intervention, the very tools meant to assist humans can also push them towards isolation, exacerbating existing social problems.

Through 2024 and 2025, the dominant narrative within the AI landscape has been capability, raw intelligence, benchmark performance, and reasoning depth. However, some companies are now proposing a reframing, arguing that contemporary AI models fall short not because they cannot reason, but because they struggle to coordinate tasks, thereby limiting their ability to foster human-to-human collaboration.

One such company recently made headlines not because it released a new model or product, but because it received backing from some of the most prominent names in the industry.

In January 2026, humans&, a company founded in late 2025, publicly announced a successful $480M seed round at a $4.5B valuation, the second-largest seed round in venture capital history. Investors include Nvidia, Jeff Bezos, SV Angel (Ron Conway), GV (Google Ventures), Emerson Collective (Laurene Powell Jobs), Felicis, CRV, and DCVC.

The billion-dollar bet on collaborative AI

The $480M seed round is not just a large number. It is the second-largest seed round in venture capital history, at a valuation that prices the company at roughly 250 times its peer cohort’s typical early-stage valuation. For context: Mira Murati’s Thinking Machines Lab, founded by the former CTO of OpenAI, raised $2B at a $12B valuation, the largest seed round ever.

Driving the valuation is the belief that while the conventional approach of making models more competent has worked, current methods fall short of making them more collaborative with humans and other AI systems. The belief is not the only thing driving up valuation while a product remains absent.

The team is building a collaborative AI

The team working on the idea includes researchers and engineers from leading AI labs, including OpenAI, Anthropic, Google DeepMind, and Meta. Among them is Georges Harik, the seventh employee at Google and a key contributor to Gmail, Google Docs, and Google’s acquisition of Android.

Another co-founder, Eric Zelikman, previously worked at Elon Musk’s AI company, xAI, where he contributed to the Grok-2 model’s training data and conducted research on reinforcement learning methods to improve AI system reasoning.

Bringing it all together is CEO Eric Zelikman, who said his company is working on models that will ”coordinate with people, and other AIs where appropriate, to allow people to do more and to bring them together.”

In simpler words, he argues that today’s AI systems are powerful but forgetful. As such, they can answer complex questions well, yet they lose the context of a conversation once it ends. The models also fail to remember how a specific person or team works over time, and they struggle to manage long, multi-step tasks like drafting, reviewing, and revising without constant human guidance. In other words, they may be intelligent, but they are not good at coordinating work. In most organizations, however, coordination is where much of the real value lies.

To address these problems, humans propose building a new class of AI models trained via multi-agent reinforcement learning to perform tasks that most current models cannot. Which is to work persistently with people, remember what it learns about them, and coordinate across long, complex task sequences.

The technical blueprint for collaborative AI

The plan is to train models using multi-agent reinforcement learning, where multiple AI agents interact with each other and with humans.

In these environments, the system receives feedback on how well the group completes tasks together, allowing the model to learn negotiation, task delegation, and consensus-building behaviors, rather than simply generating good answers.

Another key technical element is long-horizon reinforcement learning, which focuses on training models to handle tasks that unfold over extended periods. Current AI systems typically optimize for the quality of a single response. Humans& want their models to plan, act, revise, and follow through across many steps. In practice, the system could guide a process such as gathering group opinions, synthesizing them into a proposal, revising the proposal based on feedback, and helping the group reach a decision. Training models for this type of sequential decision-making requires reward systems that evaluate outcomes over time rather than the accuracy of a single response.

Additionally, the company emphasizes the importance of persistent memory architecture to enable AI to understand user preferences, past decisions, and the context of ongoing projects. By maintaining continuity between sessions, the system could participate more effectively in long-running collaborative processes.

However, while humans& has the technical roadmap and the dollars to back it, the company faces overwhelming challenges that must be addressed before it can realize its dream of creating collaborative AI systems.

A wide-open but crowded market

Humans& is entering a market that is simultaneously wide open and ferociously crowded.

The wide-open part stems from the lack of a dominant platform for human–AI collaboration at enterprise scale. At the same time, companies such as Microsoft, Google, Salesforce, and Slack are integrating AI into their collaboration tools, often using the same language humans use to describe their vision.

Microsoft has even described 2026 as the year when AI begins to function as a coworker rather than just a tool. Research firm Gartner estimates that by year-end, 40% of enterprise applications will include task-specific AI agents. In other words, the shift toward human-AI collaboration is already underway, whether or not humans& become a major player.

In this scenario, the question arises: if existing enterprise platforms are already embedding AI agents into their workflows, will companies adopt a new coordination model from a startup, or wait for similar capabilities to appear in the tools they already use?

However, what differentiates humans is not just the funding; the company believes that a coordination model specifically trained with reinforcement learning for collaboration is fundamentally different from a general-purpose large language model with a memory feature added.

This belief can manifest in concrete results when the company launches its products, and they are used at scale.

In the meantime, humans& stand on the precipice of either launching a new era of AI development or revealing the limits of AI systems’ integration with human abilities.

If humans& fails, the experiment will reveal how difficult it is to teach machines something humans do instinctively: collaborate. If, however, it succeeds, the idea that AI can actively bring people together, rather than pull them apart, could matter well beyond enterprise use and change the development of future AI models and how they are viewed and adopted by users.

Quick Bits, No Fluff

• Ars AI quote scandal: Ars Technica fires a reporter after fabricated AI interview quotes trigger an internal investigation and public backlash.

• Clear Drop trash reality: Verge reviews a 61-pound soft-plastic compactor that turns bags into tiles, questioning its cost, effort, and whether it meaningfully improves recycling.

• DoD deal user revolt: After OpenAI’s Pentagon contract, estimated ChatGPT uninstalls spike 295%, as some users delete the app over military AI concerns.

Brain Snack (for Builders)

Wednesday Poll

The Toolkit

• Leonardo AI: AI image studio for concept art, product visuals, and styled assets at speed.

• Modal: Serverless Python platform to run AI and data workloads without touching infrastructure.

• QuillBot: AI paraphraser that rewrites, tightens, and clarifies text in multiple languages.

Meme of the Day

Rate This Edition