The Power of Collective Intelligence

Current AI products have forgotten the central role of collective intelligence in the success of human societies. The search for a unitary AGI is culturally understandable in the hero’s journey context of many of our nation’s greatest leaders, inventors and scientists. However, the social context in which they were successful is perhaps as important as their accomplishments.

That importance is seen in the many instances of multiple discoveries - when independent parties invent or discover the same thing. Darwin and Wallace independently conceived of natural selection; Newton and Leibniz both discovered calculus; Graham Bell filed his patent for the telephone the same day as Elisha Gray. These are not aberrations. They reveal that discoveries become inevitable when knowledge and tools accumulate in society’s cultural store and the attention of an appreciable number of investigators becomes focused on a problem.

At the core of this social dynamic is collective intelligence, the distributed computation done by groups. A familiar example of this type of computation can be seen in ant colonies. A single ant has minimal capacity to learn, but a colony can exhibit great feats of collective intelligence. This is possible not through the individual ants’ intelligences running in parallel, but through a loop - each ant leaves traces informing others of their experiences, which in turn inform the receivers’ actions, and so on. Human intelligence is more complex than an ant’s, but the core intelligence of humanity - and our effectiveness as a species - also comes from collective intelligence.

Distributed computation across individuals accomplishes much more than the efficiency gain of parallel computation. It is a social loop of model-driven experimentation and learning. It all starts at the individual level, where our brains naturally create models of the world. These are representations of what we believe are the relevant inputs to a situation and how they’re causally connected. We use them to throw a ball (physics model), to make sense of someone’s facial expression (emotional model), and to create a marketing campaign (influence model).

Models that are constrained by universal parameters, like physics, already offer variation - the abundance of golf swing coaches out there reveals different people have different golf swing models. In the context of the social world’s soft constraints, however, this diversity explodes. Our innate differences, different life experiences, and social influences often lead us to work with different mental models of the same situations. This is the key factor that introduces approach variability to our experimentation, accelerating change through iteration. Different models focus individuals’ attention in diverse ways, allow the group to collectively attend to more information than any individual possibly could, and generate a collective dataset with which we refine our own individual models. the core feedback loop of collective intelligence.

The evolution of corporate hierarchy is a great example of such a loop. Post-WWII, the prevailing model was that employees were paid and promoted based on tenure, the key accepted proxy for value. It was only after managers with a different model decided to experiment - Jack Welch at GE, McKinsey’s up-or-out policy - that society realized the benefits of a meritocracy-based model. This steady churn of learning, experimentation and refinement drives our economy and improves society. It is also the feature, not the bug, that makes the collective intelligence of organizations incredibly powerful.

The Chasm: 10x Professionals Don’t Yield 10x Organizations

Despite the importance of collective intelligence to human reasoning, it is not at all in the roadmap of AGI. The quest that motivates the leading AI labs suggests all human problems can be solved by a superintelligence existing in the homogeneous representations of a single model. In a company’s context, this is likely true for individual professionals - with the use of AI, SWEs are orders of magnitude more efficient, CPG researchers can collect customer insights in hours, and law firms can cut headcounts by at least half. Why isn’t every organization in the world, then, cowering in a corner at the certainty that anyone with enough AI agents can just take over their business overnight? Two main reasons: scar tissue and context integration.

The first is simple: every organization has the advantage of knowing how not to do things, by having done them wrong in the past. This may seem trivial, but it is a key source of iterative speed, compared to individuals operating with only their own models. “Do not delete”, timeouts, and retries in code, as well as standard procedures, checklists and scripts in non-code workflows, essentially codify what not to do in a company, creating robust and consistent processes. The resources saved by starting from a refined model, and organizations’ ability to disseminate refinements to the whole group, are enduring sources of competitive advantage.

Context integration is more nuanced. Think of a leadership team at an organization. Over the years, its executives develop a level of social knowledge about the organization and their teams that allows them to increase the speed of collective intelligence’s feedback loop. By coordinating the individuals in their teams under a single context, like a corporate goal, this loop - where different individuals’ attention focused in diverse ways allow the group to attend to more information than any individual could - reaches escape velocity. Leadership iterates on different perspectives about strategy in hours. PMs quickly iterate over client opinions, new features and GTM. Actions are taken but rarely is there finality; the team’s integrated context allows new information to surface and the process continues. Most importantly - we’ll come back to this - everything gets done in the way the organization wants it done, not some generic method.

Both processes require the same two inputs: a sizable, fine-grained memory layer that permeates the organization’s people and allows them to factor in context not directly codified in code, products, and documents; and a myriad of mental models to nurture fast experimentation. The current AI paradigm supports neither of those. Context windows are limited to, in practice, around 130,000 tokens, or a 200-page book, before they become unreliable, and accuracy drops dramatically past the first and last 20% tokens. A medium-sized company easily has 200,000 lines of code, or 2.5 - 3M tokens. It also has a formal memory layer - internal docs, code reviews, Slack, emails, meeting transcripts, and customer calls. This is +100M tokens in raw form. Then there’s the informal memory layer - everything in its employees’s heads. That is hard to quantify, but likely an order of magnitude larger than the memory layer. At the same time, instead of embracing the diverse models that are fundamental to an organization’s experimentation, LLMs average out multiple representations into a single one. This might work in contexts where humanity has already identified a single accurate model, like physics, but for the vast majority of settings - organizations, government, tech - there are no canonical models.

It’s clear that LLMs won’t be able to incorporate organizations’ scar tissue or context integration capabilities any time soon. In other words, having 10x professionals does not translate into 10x organizations, because the (i) fundamental things that make them great can’t be replicated by LLMs and (ii) these are also the things that inform how organizations want their work done - and generalist models just can’t replicate that. No wonder why 95% of enterprise AI solutions never move past the pilot phase, and why 80% of companies that use AI have no bottom-line impact.

Crossing the Chasm: Context Engineering

But what if we treat LLMs as only a small part of the system? We wouldn’t expect a new junior SWE to instantly understand all the context in an organization’s codebase, or a new SDR to instantly identify real sales opportunities during their first day on the job. It is also well-established that, with adequate context and task description, both humans and agents can do amazing individual work - with the advantage that agents are several orders of magnitude faster and cheaper. The question, then, is a different one: how do we distill +100M tokens into the 130,000 tokens over which an LLM can understand context, task, and most importantly, how that specific organization and its specific employees would like the work done?

Context engineering is fundamentally a means of targeting a cognitive model. By focusing the LLM only on the relevant details within each session, the model can exhibit highly specific behavior, bypassing an LLM’s generalistic weights and making unlikely results reliably available. In technical terms, a well-engineered context gives the model a map to guide itself to the right region of the latent space, before looking for the weights to rely on. This approach is highly successful in a wide range of tasks and supported by frameworks like MCP and Claude skills. The beautifully useful aspect of this approach is that it allows us to replicate settings in which there are diverse models of the world, by distilling those models asynchronously and injecting them into a team’s context. In other words - it allows us to typify and simulate the individuals in an organization, both at the cognitive and memory level, recreating the collective intelligence loop that makes them unique and powerful.

The applications of this new product paradigm are endless. An agent empowered by it can (i) gather documented context of the team who would otherwise be doing the work; (ii) query them asynchronously about gaps in aggregate context - different perspectives, prioritizations, and recollections among team members - to reconcile context; and (iii) run the solution by the team’s simulations and iterate on it to make sure it fits what that specific team would do. The secret sauce here is step two: our approach uses an agent that serves as a translator between people’s different representations of context. This agent helps teams effectively align on how they are thinking about complex issues and resolve what the single source of truth about it is, without ever having to jump on a Zoom call. It would be able to not only do things autonomously, faster, cheaper, and more efficiently, but also in a way that addresses the context, scars, politics, and personalities in any team.

Where We Are Starting - Tech Debt

Few domains are as rich in scar tissue, context reconciliation, and solution specificity as tech debt. Every engineering team in the world has it, hates it, and feels frustrated to have to dedicate their time and (very) expensive engineering resources to it. There is also an ever-present tension between engineering managers and business-facing leaders: the former know the impacts of accruing tech debt and want to pay it down, the latter only hear “we’ll diverge engineers from revenue-increasing new features” when tech debt is brought up. Solving it requires doing code, Slack and team “archaeology” to figure out why the debt exists, what breaks if it gets addressed, reviewing the test suite to ensure functionality is preserved, pushing a PR, code review back and forths, commiting, and then chasing down something that broke anyways. Cursor and Claude Code are extremely helpful with execution, but gathering and organizing context is still very much a human task. It’s a messy process, with one important complication - nobody became a software engineer to address tech debt. Everybody knows how crippling it can be, but virtually nobody wakes up in the morning eager to work on it.

The issue is also timely. Vibe coding, both done by professional SWEs and non-code oriented professionals, is a paradigm shift in how humanity develops software and is here to stay. It also leads to code that is bloated (easily 2x the length of optimized code), brittle and often repetitive, since most of it is not properly guardrail with adequate context, documentation, and tests. There are also several companies that are started by non-technical people through vibe coding tools, like Lovable, that enjoy early success but soon need to be rebuilt from the ground up, at enormous engineering cost, when they find their “architecture” does not scale. These, and many other examples, make it easy to see that as adoption of vibe coding continues to grow, tech debt will balloon exponentially - further expanding a problem that already costs $2.4tn per year.

We believe this vision and this product can create the next paradigm shift in how companies and individuals work. If that is appealing to you, we’ love to chat - write us at edu@finegrained.io / jon@finegrained.io.