The whole world is racing to adopt AI to a degree I would never have guessed when I started building LLMs in 2019. I am convinced this is one of the most impactful human technologies. But I am also seeing a transformation paradigm that is missing the main point, and one that will eventually lead societies, companies, and individuals into a dead end. To explain this, we first need to look at how we got here.

A Social Algorithm

Biologically, we are still essentially cavemen. A human born millennia ago possessed identical hardware and comparable intelligence to us today. One of those early humans once figured out that a round object could be rolled under a heavy load. The wheel: a fantastic invention.

So how does this exact same species now walk on the Moon, manufacture 2 nm semiconductors, and conduct symphonies? How did we lift billions out of poverty and mortal danger — conditions that have historically been the natural state of humanity?

We got here through a combination of several drivers, stacked on top of one another, each operating at a different frequency and under different guidance.

At the bottom: biology. This is the lowest-frequency algorithm. Evolution tries random mutations to see what sticks — meaning, whatever produces the most surviving offspring. These changes take millennia but eventually yield staggering complexity. However, the underlying exploration process is undirected and purely iterative.

At the top: the individual. This is the highest-frequency algorithm. Every human is a machine driven by emotions, instincts, and a remarkably powerful brain — one capable of adapting to a life radically different from the environment that shaped it. We each start helpless, with very little programming, yet everything we see around us was ultimately mastered by this kind of creature.

In between: the social algorithm. Operating on the timescale of generations, this is the magic that allowed ancient human genes to produce astronauts, neurosurgeons, and conductors. You might call it “culture” or social programming: an environment that unites biology and the individual to create exponential results. This algorithm identifies and preserves knowledge. It builds norms and values, creating a foundation where humans thrive by combining radically different skills and preferences.

No single human — no matter how exceptional — can impact the world all that much in isolation. Transport yourself to an empty, Earth-like planet, and even with everything you know, your possibilities immediately collapse.

Long-Term Game

The power of this middle layer is evident in every touchpoint we have with civilization. If you want to build a trebuchet, you can order a kit and have it arrive at your door in days. That convenience is the output of millions of humans collaborating across generations, time zones, supply chains, legal systems, and languages — most of whom will never meet.

Beyond the functional machinery of institutions, there lies a deeper layer: a social algorithm of collaboration that drives every force that has ever pulled humans together or pushed them apart. We see it in any human group, starting in early childhood education and family structures, but it extends much further. We have built organizations and mechanisms that allow global collaboration. Our social attention algorithms surface great ideas and integrate them into a global conversation. This could be art, social media, academia, politics, or the free market — all giant discovery and coordination games.

What does this social algorithm actually do? I see it as operating across four dimensions:

Ascension of Our Tools

Most of our tools have been created by communities operating like this, standing on the shoulders of giants. But technology has historically mostly served to enhance the capabilities of the individual. In that sense, a hammer is like an oscilloscope: it empowers a single human.

The biggest breakthrough for our social algorithm has undoubtedly been information and communication technology. Modern tech not only extends our reach but helps filter the noise that inevitably accompanies connecting billions of people.

Historically, however, the impulses behind these four dimensions of collaboration were distinctly human. Humans create structures, form teams, ask questions, initiate collaboration, decide on actions, and scale humanity’s global knowledge base.

With the power of modern AI, we are seeing patterns that transcend these historical limits. Local AI agents demonstrate a persistent, almost independent agency to approach their users. AI chatrooms have served as tech demos for unprompted collaboration, and there are playful attempts at allowing AI to initiate contact with humans or other agents autonomously.

With agentic swarms that keep collaborating beyond the lifetime of any single agent, we are seeing aspects of these dimensions emerge for the first time — though still in isolated environments, and without humans.

Yet, in almost all current cases, these systems lack a genuine understanding of teams and human dynamics. If a “think-act” loop occurs, it is generally isolated between one human and one agent. AI currently lacks the necessary knowledge and goals for team-wide analysis, complex decision-making, and execution. We simply haven’t built the technology to master coordination: knowing whom to approach and how to combine the strengths of various intelligent entities.

Furthermore, the long-term aspects of the “build” and “scale” dimensions are only beginning to surface — and entirely absent in human–AI settings. Current AI treats its environment as fixed, leaving the task of scaling new learnings beyond the immediate interaction entirely to the human operator. This keeps AI relegated to the level of a basic tool, operating beneath our social fabric.

The Age of Ingenuity

The same social algorithms that allowed early humans to build civilizations now govern our modern enterprises. A company is simply a micro-culture — a structured group of individuals utilizing collective knowledge to solve complex problems and create value for everyone involved. We are now racing to bring AI into the enterprise but overlook its biggest potential:

Imagine if we had given language model technology to humans thousands of years ago, trained only on the language available at the time. These models wouldn’t know what a wheel was, nor could they conceptualize it based on existing data. This has been explored directly by training models on a deliberately limited knowledge base. Talkie is a 13-billion-parameter language model trained only on pre-1931 text. LittleLearner is a family of models trained only on a U.S. elementary-school (K–5) curriculum. In its experiments, scaling, post-training, and in-context learning amplified what the curriculum had taught — but none of them meaningfully pushed the model’s capability beyond its K–5 training boundary. Together, these projects highlight the systematic limitations of a model confined to a narrow slice of knowledge.

AI’s limited focus on its use as a tool for relatively simple human–machine patterns was understandable while the technology was in its infancy. But this narrow focus now creates a ceiling on its impact. The next monumental leap in value will not come from an iterative increase in tool quality, but from a paradigm shift in how we integrate AI into our work and lives.

There are exciting things happening for AI systems generating new data to adapt themselves, or going out and getting it from connected systems and sources. It’s unclear if this will eventually remove all the limits in areas where humans comparatively excel: conceptualizing truly novel ideas, navigating genuine change, or managing situations requiring deep responsibility and judgment. For today’s systems, these are the cases in which I see AI falling short.

Making today’s knowledge instantly accessible is undeniably useful and can turbocharge knowledge work. It is, in fact, the clearest expression of AI’s real superpower: learning patterns from everything that has ever been written, and combining them at a speed no human can match. That is precisely where humans are considerably worse — while the things humans excel at remain largely out of reach for AI. Two fundamentally different kinds of intelligence, each with its own edge. We have not yet designed systems that take this into account.

Research teams have looked at today’s AI capabilities. In one study analyzing AI-generated stories and narration, the outputs of many different models converged into a similar, predictable structure: the major structural decisions that make a story remarkable are far less diverse than in human-authored texts. These “paint-by-numbers” results mean AI circles around a pleasant average — without genius inspiration and conceptually brilliant strategy (see StoryScope, Russell et al.).

This directly impacts human experts, with chatbot users showing a decline in neural engagement and output diversity (Kosmyna et al., Your Brain on ChatGPT). Giving the lead for problem solving to an AI agent and relegating humans to error correction directly affects human capability, muting creativity and collapsing inspiration. We can even measure “brain fry” in experts tasked with babysitting AI outputs (Bedard et al.). Businesses are drowning in “workslop” — plausible-sounding simulations of work documents that look good at first glance but are full of content that is “almost right” and rarely brilliant (Doshi & Hauser). Current AI integration is uninspiring, and the setups are barely human-compatible.

This is not to doubt that this technology is the fastest industrial revolution humanity has ever experienced, one that comes with outstanding potential. But do we genuinely believe we have invented all the potential wheels — that there is no transformative innovation left to uncover? Is all that is left to improve having AI take over business processes step by step, while humans act as temporary supervisors until the machine learns the quirks?

I am convinced the opposite is true: we are entering an era of ingenuity, where human potential can be uncompromisingly directed toward the contributions that actually shape the future.

We need AI that includes humans not just for error correction, but to provide the vital signals missing from the training data. How can we “prompt” the human to give us the most valuable “completion”? When should an AI initiate what kind of collaboration, and with whom?

Collaborative AI bridges this gap by shifting the machine from a passive servant to an active partner in the social algorithm. Consider automotive manufacturing: an AI can endlessly attempt to optimize the aerodynamics or weight of individual parts. But together with a human, radical new ideas can emerge. If a designer envisions a car with a panoramic display spanning the entire width of the dashboard, they might realize they can move side-mirror and blind-spot warning functionalities directly to that screen, since it is installed right there anyway. This kind of conceptually new idea can come from human intuition and lateral thinking. A collaborative AI can then take that creative leap and help inspire, evaluate, and iterate on the execution in record time.

We have already seen this pattern succeed on a smaller scale. Mathematician Terence Tao has used AI under his own guidance, providing high-level reasoning and architectural direction while relying on AI to complement his blind spots — such as finding obscure papers or rapidly iterating through ideas at speeds no human can match. In his recent paper with Tanya Klowden, Mathematical methods and human thought in the age of AI, Tao highlighted the impact of AI and put the spotlight on the “scale” dimension of collaboration: learning, understanding, and growth. Even if we had a guarantee that an AI solution is correct and complete, if no human understands what is going on, this limits the impact of this result substantially.

In a world where AI is becoming commoditized, many companies will race to the bottom via cost reduction and automation. But ideas that cannot be copied simply by purchasing a chatbot subscription are the only true long-term market differentiators. Ideas that cannot be assembled from patterns in existing training data — not the sheer number of data-center graphics cards — will dictate success. In an environment where all the “boring” parts of work can be scaled more cost-efficiently than ever, and where innovation drives the biggest potential in the market, this might open up opportunities for companies in a different way: there is plenty of evidence across all kinds of industries that small, flexible teams can innovate faster because they are less burdened by the structures and legacy that giant enterprises bring with them. Massive company size may here actually hinder the kind of innovation and speed that AI makes possible.

The winners that will emerge and build new empires through this transformation will be characterized by two main ingredients:

There is not a lot of technology around that even points in this direction yet, but I’m confident that more teams will follow once it becomes evident how impactful collaborative AI will be — not just in coming up with disruptive ideas, but also in helping brilliant humans learn deep knowledge and build teams, and in retaining and scaling rare experts’ experience within the organization. It’s time we pull AI onto eye level, promoting it from just a tool the human uses into a collaborative partner that inspires us and empowers human–machine teams to co-create at the innovation frontier.