Imagine having to rediscover glassmaking every time you wanted a window, or work out how to cultivate and roast coffee for every cup. We don’t – because civilisation depends on knowledge accumulated by people we have never met. We inherit their techniques, mistakes and discoveries, adapt them and pass them on. Michael Muthukrishna and Joseph Henrich call this the collective brain: innovation from people learning from one another and accumulating improvements across generations (Muthukrishna & Henrich, 2016).
When we talk about and build AI agents, we so often imagine the agent as a singular unit – why? Even Artificial Intelligence: A Modern Approach introduces intelligence largely through the individual agent: an agent perceives its environment through sensors, reasons, learns and acts through actuators. The book certainly includes multi-agent environments, but the individual agent remains the conceptual starting point as well as the main focus (Russell & Norvig, 2021). It doesn’t have to be that way. Unlike humans, an artificial agent need not spend years learning what another has already discovered. A huge advantage of the digital world is that a useful result can be copied and shared with thousands of others in seconds, at a scale and cost impossible for human teaching. So, what if we designed AI around the accumulation of shared knowledge rather than the accomplishments of individual artificial minds?
A collective memory, not a collective mind
Biology offers several ways to think about this. Bacteria can exchange useful genes through horizontal gene transfer, for example through plasmids transferred by conjugation (Arnold, Huang & Hanage, 2022). Honeybees investigate different nesting sites and communicate their findings; and competing signals help the swarm reach a decision (Seeley et al., 2012). Seirian Sumner’s work on wasps likewise illustrates the sophisticated organisation and ecological roles of social insects with tiny individual brains (Brock, Cini & Sumner, 2021). These are different mechanisms, but each shows that a population’s capabilities need not be limited to those of its individual members.
The digital world has already created hive-minds of sorts. Peer-to-peer systems such as IPFS distribute data without a single central store (Benet, 2014), while blockchains show how participants can maintain agreement about a shared record without trusting one administrator (Nakamoto, 2008). Neither creates new knowledge or proves that a scientific claim is true. But they demonstrate that information can be distributed, persistent and shared across many independent nodes. There are therefore many possible architectures in which AI agents could have access to a common, traceable memory without having to think or act identically.
A hundred agents searching for the highest peak
Russell and Norvig explore simple search strategies; breadth-first search, depth-first search, greedy search and A* are useful foundations, but exploring enormous numbers of possibilities still feels cumbersome (Russell & Norvig, 2021). In my mind, the obvious solution to finding the highest peak in a landscape is not to send out one agent to search, but a search party of hundreds of agents.
They all start in different parts of the landscape. One finds a remarkable peak, another discovers a region full of dead ends, and a third develops a better way to decide where to leap. Working independently, each benefits only from its own discovery. But sharing information gives all of them the promising location, the warning and, most usefully, the improved search strategy. This knowledge exchange does not mean every agent should abandon its own exploration to climb the same mountain. Some can investigate its surroundings while others continue exploring distant regions, equipped with the new knowledge. This is the real opportunity in enormous search spaces; a shared constraint that rules out millions of possibilities, an abstraction that turns a difficult problem into a familiar one, or a new heuristic that tells everyone where not to search could be worth far more than a hundred times as many attempts.
Share the knowledge, preserve the diversity
There is evidence that diversity matters. Lu Hong and Scott Page modelled how, under particular conditions, functionally diverse problem-solvers can outperform individually stronger but more similar ones: different representations help a group find solutions others overlook (Hong & Page, 2004). Maxime Derex and Robert Boyd found that partially connected groups developed more diverse and complex solutions in a computer-based innovation task than fully connected groups, where participants rapidly copied successful approaches (Derex & Boyd, 2016). In simulations, David Lazer and Allan Friedman found a similar exploration–exploitation trade-off: rapidly spreading successful strategies could improve early performance while reducing longer-term exploration (Lazer & Friedman, 2007).
This is not an argument for hoarding information or restricting access. Useful knowledge should be shared as widely as possible. What we must preserve is diversity in what people – or agents – do with it.
In fact, research on human cooperation makes the argument stronger. A meta-analysis of 122 studies found cooperation substantially more effective for achievement and productivity than interpersonal competition or individualistic effort (Johnson et al., 1981). A later meta-analysis of 46 studies found that cooperative groups outperformed people competing against one another across linguistic, non-linguistic, well-defined and ill-defined problem-solving tasks (Qin, Johnson & Johnson, 1995). Cooperation and diversity are therefore not opposites. The interesting design problem is how to let many different problem-solvers work toward the same goal, share what they learn, and still keep searching in different ways. Fu and colleagues found that parameter sharing – giving cooperating agents the same underlying policy parameters – can actually be harmful in some multi-modal environments, while separate policies can preserve useful diversity (Fu et al., 2022). That is almost exactly the architecture I am interested in: shared purpose and shared knowledge, but diverse strategies.
Artificial agents make this unusually tractable because we can design their incentives. In fully cooperative multi-agent reinforcement learning, all agents can literally receive the same reward and optimise towards the same team objective (Gronauer & Diepold, 2022). If I create a hundred agents to solve one problem, they do not have competing careers, salaries, patents, egos or institutional interests, so helping the collective solve the problem will be our definition of success.
We already have other technical precedents. CommNet showed that cooperating agents can learn how to communicate alongside their behaviour using backpropagation (Sukhbaatar, Szlam & Fergus, 2016). Population-based training allows multiple neural networks to learn simultaneously, copying successful parameters while continuing to experiment with different settings (Jaderberg et al., 2017). Backpropagation itself is not knowledge sharing between agents, and parameters cannot always be transferred between architectures. Nevertheless, artificial systems can learn both from experience and from one another.
Sure, digital collaboration is not 100% free. Communication costs bandwidth, computation and electricity; a successful strategy may not transfer to another environment, and coordination takes work. But human history is full of technologies designed to reduce the cost of learning from one another. The printing press is one well-studied example: early European cities that adopted printing grew faster than comparable cities that did not (Dittmar, 2011). Digital technologies reduce the costs of storing, copying and transmitting information still further (Goldfarb & Tucker, 2019). AI can potentially take this much further. Once one agent has found and verified something useful, thousands of other agents could have access to it almost immediately. The design challenge is not how to make collective learning less cooperative. It is how to make it cumulative without making it uniform.
What problems should all these minds be solving?
In my earlier essays, I questioned our fascination with artificial systems that resemble humans, win games or generate amusing pictures. I am much more interested in what AI might do that humans cannot: search at enormous scale, notice relationships no individual could remember, and help solve existential crises. Humanity is facing some serious existential crises:
In 2024, around 673 million people experienced hunger (FAO et al., 2025). Some 4.9 million children died before reaching their fifth birthday (UN IGME, 2026). Neglected tropical diseases affect more than one billion people (WHO, 2025). Around one million animal and plant species are threatened with extinction (IPBES, 2019). Climate change is already damaging human societies, livelihoods and ecosystems (IPCC, 2022; WMO, 2025). The economic costs of inequality are similarly vast. The World Bank estimates that closing gaps in lifetime labour earnings between women and men could add US$172 trillion to global human-capital wealth (World Bank, 2020). Racism is harder to quantify globally; in the United States alone, Citi estimates that gaps between Black and white Americans in income, business ownership and homeownership could have cost the economy up to US$3.9 trillion between 2020 and 2024 (Citi, 2024).
These are complex problems, maybe without one universal optimum. What does it take to keep a child alive to adulthood? Which agricultural practices can feed communities without destroying the ecosystems on which they depend? What can we learn from places that have achieved dramatically better outcomes, and how can their solutions be adapted elsewhere? Some parts of these problems already have effective solutions – just not applied everywhere. Others, such as better diagnostics and medicines against neglected tropical diseases, require discoveries we have yet to make. Diversity of thought and strategy is therefore not an inconvenience to eliminate: it is part of solving the problem.
At Helminthix, we are applying AI to one part of this challenge: accelerating better diagnostics and medicines for parasitic worm infections. Different agents could investigate parasite targets, chemical compounds, drug combinations and diagnostic approaches, sharing both promising results and failed experiments. One agent might recognise a molecular relationship another missed. A failed experiment could prevent dozens of agents repeating the same dead end. A useful result in one parasite could generate a hypothesis worth testing in another.
Perhaps the most useful AI will not be one immense mind that knows everything? It will be a population of different minds that can share what they discover, challenge what they inherit and keep addressing extraordinarily complex, sometimes existential, problems worth solving.
Share the knowledge, preserve the diversity. Put both to work on problems that matter.
References
Arnold, B.J., Huang, I.-T. & Hanage, W.P. (2022). Horizontal gene transfer and adaptive evolution in bacteria. Nature Reviews Microbiology, 20, 206–218. doi:10.1038/s41579-021-00650-4.
Benet, J. (2014). IPFS — Content Addressed, Versioned, P2P File System. arXiv:1407.3561.
Brock, R.E., Cini, A. & Sumner, S. (2021). Ecosystem services provided by aculeate wasps. Biological Reviews, 96(4), 1645–1675. doi:10.1111/brv.12719.
Citi (2024). Closing the racial wealth gap. Citi Global Perspectives & Solutions.
Derex, M. & Boyd, R. (2016). Partial connectivity increases cultural accumulation within groups. Proceedings of the National Academy of Sciences, 113(11), 2982–2987. doi:10.1073/pnas.1518798113.
Dittmar, J.E. (2011). Information Technology and Economic Change: The Impact of the Printing Press. The Quarterly Journal of Economics, 126(3), 1133–1172. doi:10.1093/qje/qjr035.
FAO, IFAD, UNICEF, WFP & WHO (2025). The State of Food Security and Nutrition in the World 2025. Rome: FAO. doi:10.4060/cd6008en.
Fu, W., Yu, C., Xu, Z., Yang, J. & Wu, Y. (2022). Revisiting Some Common Practices in Cooperative Multi-Agent Reinforcement Learning. Proceedings of the 39th International Conference on Machine Learning, PMLR 162, 6863–6877.
Goldfarb, A. & Tucker, C. (2019). Digital Economics. Journal of Economic Literature, 57(1), 3–43. doi:10.1257/jel.20171452.
Gronauer, S. & Diepold, K. (2022). Multi-agent deep reinforcement learning: a survey. Artificial Intelligence Review, 55, 895–943. doi:10.1007/s10462-021-09996-w.
Hong, L. & Page, S.E. (2004). Groups of diverse problem solvers can outperform groups of high-ability problem solvers. Proceedings of the National Academy of Sciences, 101(46), 16385–16389. doi:10.1073/pnas.0403723101.
IPBES (2019). Global assessment report on biodiversity and ecosystem services. Intergovernmental Science-Policy Platform on Biodiversity and Ecosystem Services. doi:10.5281/zenodo.3553458.
IPCC (2022). Climate Change 2022: Impacts, Adaptation and Vulnerability. Cambridge University Press.
Jaderberg, M., Dalibard, V., Osindero, S. et al. (2017). Population Based Training of Neural Networks. arXiv:1711.09846.
Johnson, D.W., Maruyama, G., Johnson, R.T., Nelson, D. & Skon, L. (1981). Effects of cooperative, competitive, and individualistic goal structures on achievement: A meta-analysis. Psychological Bulletin, 89(1), 47–62. doi:10.1037/0033-2909.89.1.47.
Lazer, D. & Friedman, A. (2007). The network structure of exploration and exploitation. Administrative Science Quarterly, 52(4), 667–694. doi:10.2189/asqu.52.4.667.
Muthukrishna, M. & Henrich, J. (2016). Innovation in the collective brain. Philosophical Transactions of the Royal Society B, 371, 20150192. doi:10.1098/rstb.2015.0192.
Nakamoto, S. (2008). Bitcoin: A Peer-to-Peer Electronic Cash System. White paper.
Qin, Z., Johnson, D.W. & Johnson, R.T. (1995). Cooperative Versus Competitive Efforts and Problem Solving. Review of Educational Research, 65(2), 129–143. doi:10.3102/00346543065002129.
Russell, S.J. & Norvig, P. (2021). Artificial Intelligence: A Modern Approach, 4th ed. Pearson.
Seeley, T.D., Visscher, P.K., Schlegel, T., Hogan, P.M., Franks, N.R. & Marshall, J.A.R. (2012). Stop signals provide cross inhibition in collective decision-making by honeybee swarms. Science, 335(6064), 108–111. doi:10.1126/science.1210361.
Sukhbaatar, S., Szlam, A. & Fergus, R. (2016). Learning multiagent communication with backpropagation. Advances in Neural Information Processing Systems, 29, 2244–2252.
UN Inter-agency Group for Child Mortality Estimation (UN IGME) (2026). Levels & Trends in Child Mortality: Report 2025. New York: UNICEF.
WHO (2025). Global report on neglected tropical diseases 2025. Geneva: World Health Organization.
WMO (2025). State of the Global Climate 2024. Geneva: World Meteorological Organization.
World Bank (2020). World could achieve gender dividend of $172 trillion from closing lifetime earnings gaps. World Bank