The Unobvious Crowd: Why Engineered Variety Beats Expert Consensus

Some days the argument you've been making for years suddenly feels crisp in your mouth. You find yourself in a room, or across a table, or on a call, and the explanation lands differently, without the usual scaffolding. Those are the days!

The Architecture of Heterogeneity

There is a fundamental tension at the heart of artificial intelligence applied to creativity: the training process that makes large language models useful also makes them homogeneous. They aggregate the written record of human thought and produce outputs that represent, in effect, the consensus. For most tasks (drafting an email, summarising a report, answering a question) this is fine. For frontier innovation, where the goal is to find ideas that no one has yet had, or that the established community of practice has systematically overlooked, consensus is exactly the wrong starting point.

The conceptual breakthrough that resolves this tension is what might be called engineered heterogeneity: deliberately building systems whose outputs resist convergence. This means combining human crowds, selected specifically for their cross-disciplinary diversity, not their domain expertise, with synthetic agent populations that are prompted to explore ideation from radically different contextual starting points. A synthetic agent instructed to think about a materials science challenge from the perspective of an emergency field medic in a remote location will produce different outputs than one starting from an industrial R&D context. Neither is "correct." Both are valuable precisely because they are different from each other, and different from what a roomful of chemists would produce.

"The value of a diverse crowd is not that it averages away to the truth. It is that it refuses to average at all, and in that refusal, it finds the edges of the solution space that experts have trained themselves not to see."

The practical application of this principle produces something more useful than any individual expert analysis: a solution space map. When thousands of ideas, human and synthetic, are generated and then clustered, the resulting landscape tells you not just where ideas exist but where they don't. The empty clusters, the zones where the synthesis imagined plausible solution territory but the real ideation never went, are often more intellectually valuable than the populated ones. They point to genuinely unexplored space. This is a generalisable method for strategic ideation: use clustering not to describe what you have, but to reveal what you're missing.

On the Limits of the Data You Commissioned

There is a subtler challenge in the application of AI to large-scale idea generation that receives less attention: the scope problem. When you ask a system, whether human crowd, AI agent, or both, to generate solutions to a challenge, the breadth of the search space is a design choice, not a given. Specify it too narrowly and you reinforce the existing frame. Specify it too broadly and you generate a large volume of ideas that are technically responsive to the challenge but practically unusable by the organisation that commissioned them.

Consider an organisation with a highly specific core business; particular assets, particular operating contexts, particular risk tolerances, reviewing a dataset of a hundred synthesised approaches that spans everything from incremental process improvements to technologies that would require a complete restructuring of the value chain. The volume creates an illusion of richness. In practice, the evaluators face a filtering problem that may be harder than the original creative challenge: how do you quickly sort signal from noise when the noise is plausible?

"The question is not whether an idea is good. The question is whether it is good for this organisation, at this moment, with these capabilities. Generating without filtering is not creativity - it is the first half of a two-part job."

The answer is a structured categorisation framework before the ideas reach a human evaluator. Mapping solutions on two orthogonal axes - how novel is this to the specific organisation, and how novel is it to the world at large - produces a nine-box grid that makes the portfolio visible. The bottom-left box contains continuous improvement ideas: incrementally better versions of things already being done, easily evaluated and often quickly actionable. The top-right contains genuinely transformational concepts: new to the organisation and new to any organisation. These require a different evaluation mindset, different success metrics, and almost certainly a different timeline. Mixing them in a single undifferentiated list is a recipe for the bottom-left to always win.

Community Is Not a Spike, It Is a Condition

A related failure mode exists in the domain of community building, and it is so common that it has almost ceased to attract comment. An organisation - a conference, a professional network, a special-interest group, builds its entire identity around a single annual event. The event is excellent. The community is energised. And then, for the other eleven months of the year, it slowly deflates until the next spike. The effort required to rebuild momentum from near-zero each cycle is immense and demoralising; the model is structurally inefficient; and the community members who would provide the most long-term value are precisely the ones most likely to disengage in the gaps.

The correct diagnosis, which is not as widely acted upon as it should be, is that the problem is not that you need more small events. It is that the community does not have a continuous identity that exists independently of any event. The annual gathering is a manifestation of something, not the thing itself. Building community continuity requires building a brand presence, a set of recurring touchpoints, a reason for people to remain in the orbit of the community even when there is nothing scheduled, a sense that something interesting might happen at any time, not only in November.

"The most sophisticated communities in the world do not end when the room empties. They continue as a low-level hum - a conversation that pauses for the big event, not one that only exists because of it."

The tactical execution of this principle is genuinely difficult. It requires staffing (social media is the hardest role to fill and retain in a volunteer-led organisation), content strategy, and a willingness to create smaller, lower-stakes touchpoints whose success metric is engagement quality rather than attendance numbers. Evening formats, informal, mixed, off the main stage, can build the network density that large events cannot, precisely because they allow a different kind of conversation. The hackathon model, where the community itself becomes a working group on an open challenge, is a particularly powerful mechanism: it produces new ideas, creates new relationships, and generates content simultaneously.

The Difficult Art of the Intellectual Handover

Knowledge transfer is one of those activities that everyone nominally values and almost everyone systematically underinvests in. The typical account handover, in any professional context, is a rushed conversation and a half-completed document. The departing person is already half-absent; the arriving person is not yet sufficiently present; and the institutional knowledge that took years to accumulate moves across in a form that is technically complete but practically insufficient.

The conditions that make a handover genuinely work are specific and mostly involve preparation time: a written record created by the departing person while they still have full recall and full motivation; a structured format that covers not just the "what" but the "why" - the context behind decisions, the history of relationships, the risks that are not yet risks but might become them; and a transition period where both people are in contact with the account simultaneously. These conditions are rarely achieved. When they are, the result is something like a 20-page document that a successor can actually use - not because it is comprehensive, but because someone thought carefully about what a new person would need to know that wasn't written anywhere else.

"The test of a good handover document is not whether the departing person would give it five stars. It is whether, six months later, the successor is still referring to it when something unexpected happens."

There is a separate category of handover complexity that exists in accounts with unusual structural characteristics: long rotation cycles in the user base, split technical and commercial relationships, ongoing procurement negotiations with multiple potential vehicles. Here the standard template is actively insufficient. In these cases, the document is necessary but not sufficient; what is also needed is a clear allocation of authority between the people taking over, and a shared understanding of which decisions can be made independently and which require escalation. Ambiguity at handover is the most predictable source of account risk in the months that follow.

Making Something That Travels

There is a practical theory of intellectual virality that is distinct from, and often in conflict with, the popular conception of how ideas spread. The popular conception involves luck, timing, and an ill-defined quality of "resonance." The practical theory is less romantic and more actionable: reach beyond your existing network, bridge to communities where your idea is genuinely relevant but currently unknown, and make the act of sharing as frictionless as possible.

The first constraint is the most limiting. Any piece of intellectual work (a book, a talk, an article) can travel only as far as its author's existing network reaches, if that is the only distribution mechanism. The ceiling is deterministic. Breaking through it requires identifying adjacent communities: professional networks, special-interest groups, geographic clusters where the central idea addresses a real problem that people in that community are already thinking about. Each such community requires a different entry point, not a mass distribution but a single, well-chosen conversation with someone who already has standing there.

The second constraint is friction in the act of endorsement. Reviews, shares, and recommendations are acts of social proof; they carry weight precisely because they cost something. The practical implication is that coordination matters enormously. Twenty substantive reviews arriving in the same week create a visible signal that algorithms and readers both respond to. Twenty reviews distributed over two months produce a noise floor. The mechanism that makes coordination possible, a shared commitment, a clear process, a specific date, is the actual product of a launch strategy.

Everything else is content. The networks you have not yet reached are not inactive. They are simply unaware. The question is not whether they would be interested. It is whether anyone has yet had the conversation.

What is your reflection on this? Until next time, let's reflect and debate.

Editor - Ideas for a Better World