Source-linked AI summary
A social path to human-like artificial intelligence
Edgar A. Duéñez-Guzmán, Suzanne Sadedin, Jane X. Wang, Kevin R. McKee, Joel Z. Leibo
TL;DR
The paper frames novel data generation, rather than data assimilation alone, as a growing AI challenge. It synthesizes how interacting agents and evolutionary processes can continually reshape learning data, alongside examples of AI systems using these mechanisms. It concludes that integrating such generative processes suggests a path toward compounding innovation, with limitations around self-generated training data and lower-level units.
Problem
As large language models outpace human data generation, human culture’s accumulated output may become an AI bottleneck; consuming more data does not itself create new data.
Source: Synthesis · p. 8Method
The paper synthesizes evolutionary and AI evidence on how structured competition, cooperation, social learning, and multi-scale interactions can enrich agents’ data streams.
Source: Synthesis · pp. 8–9Results
The review links interaction-driven data generation to increasingly complex skills, robust StarCraft II performance, human-level Diplomacy, and task-dependent effects of network topology.
Source: Arms races · p. 4, Machiavellian selection · p. 5, Social learning · p. 5Takeaways & Limitations
Integrating AI’s existing algorithmic components with generative processes that enrich data streams is proposed as a route toward compounding innovation.
Source: Synthesis · p. 9Training models on their own outputs may cause successive loss of data-distribution tails, while lower-level unit simplification may limit ongoing data generation outside cumulative culture.
Source: Major transitions in evolution · p. 7, Cumulative culture · p. 8
Abstract
from arXiv · showhide
Traditionally, cognitive and computer scientists have viewed intelligence solipsistically, as a property of unitary agents devoid of social context. Given the success of contemporary learning algorithms, we argue that the bottleneck in artificial intelligence (AI) progress is shifting from data assimilation to novel data generation. We bring together evidence showing that natural intelligence emerges at multiple scales in networks of interacting agents via collective living, social relationships and major evolutionary transitions, which contribute to novel data generation through mechanisms such as population pressures, arms races, Machiavellian selection, social learning and cumulative culture. Many breakthroughs in AI exploit some of these processes, from multi-agent structures enabling algorithms to master complex games like Capture-The-Flag and StarCraft II, to strategic communication in Diplomacy and the shaping of AI data streams by other AIs. Moving beyond a solipsistic view of agency to integrate these mechanisms suggests a path to human-like compounding innovation through ongoing novel data generation.
Collective living
Collective living can make interactions among agents a source of innovation, while population pressures and arms races can keep data and challenges changing. These mechanisms have limits: competition can leave cooperation unattainable, and two-player zero-sum arms races can stagnate.
Collective living
- Collective interactions can produce innovation beyond single agents, but typical self-organizing multi-agent models converge to static or cyclic equilibria.Such interactions may also provide autocurricula—naturally emerging sequences of data presentation that may aid learning.Source: Collective living · p. 3
Population pressures
- Resource competition can diversify evolutionary strategies, as in adaptive radiations following colonisation events.Source: Population pressures · p. 3
- Genetic algorithms use population pressures to maintain diversity, while Malthusian reinforcement learning [26] varies subpopulation sizes by returns and competition has yielded innovations exploiting misspecified goals or environment errors.Source: Population pressures · pp. 3–4
- Population pressures motivate agents to seek new data, but the search depends on competition and cooperative equilibria often remain unattainable.Source: Population pressures · p. 4
Arms races
- Arms races create autocurricula: agents adapt to peers, scaling challenges to their abilities and supplying learning gradients toward otherwise unattainable optima.In nature, positive feedback in selection can drive extreme traits, as in the cheetah–prey speed race.Source: Arms races · p. 4
- Arms-race autocurricula have produced successions of increasingly complex skills [14] [27] [59] [25] [60] [61].In self-play, agents can discover strategies against peers and then encounter changing opponent data as those peers approach local optima.Source: Arms races · p. 4
- In two-player zero-sum games, policy space can collapse to a small set of intransitive strategies as skill rises, stagnating experience; continued compounding innovation appears possible outside that setting [50].Source: Arms races · p. 4
Social relationships
Social relationships change the data agents encounter and create incentives for social learning and strategic communication, but these processes can settle into collective equilibria. Their effects depend on interaction structure, network topology, and the costs and scope of learning.
Social relationships
- In persistent groups, attending to others as individuals creates social pressures; social learning and strategic communication can alter agents’ data streams, though these mechanisms may become trapped in collective equilibria.Source: Social relationships · p. 4
Machiavellian selection
- The social brain hypothesis links competitive and cooperative relationships to cognitive-evolution feedbacks, including reciprocity, cheating, cheater detection, and evasion that may fuel cognitive autocurricula [24].Source: Machiavellian selection · p. 5
- Multiple opponent pools helped AlphaStar remain robust to diverse strategies and rival champion StarCraft II players, while language-model communication of intentionality supported human-level performance in Diplomacy.In two-player zero-sum populations, training-partner and learner choices can also lead to optimal behavior in the limit [68].Source: Machiavellian selection · p. 5
Social learning
- Social learning spreads skills and can sustain cultural traditions, but when innovation is costly, imitation may outcompete it and reduce innovation overall.AI approaches include imitation learning and behavioral cloning, but learned behaviors must already exist in the data for this form of social learning to work.Source: Social learning · pp. 5–6
- Network topology influences innovation diffusion, belief propagation, and collective memory; decentralized networks suit collective problem-solving, while centralized ones favor innovation and collective action.Source: Social learning · p. 5
- Human-level compounding innovation through social learning emerged from flexible multi-scale interactions among population pressures, social relationships, and other evolutionary forces.Using social learning for ongoing data generation requires understanding the dynamics of those interactions.Source: Social learning · p. 6
Major transitions in evolution
Major evolutionary transitions reorganize interactions and information flows across levels, while cumulative culture uses language and teaching to accumulate knowledge across generations. The paper connects these processes to AI but notes limits in current collective-learning approaches and risks from training on AI-generated data.
Major transitions in evolution
- Major evolutionary transitions form higher-level units through coordinated cooperation, division of labor, and the suppression or channeling of conflict.Information transmission helps groups coordinate and develop group-level individuality.Source: Major transitions in evolution · p. 6
- Higher-level collectives can shape lower-level data streams through cooperative and competitive incentives, including in development and collective foraging.Examples include extracellular matrices, competitive pruning of cells, ant pheromone trails, and honeybee waggle dances.Source: Major transitions in evolution · p. 6
- A higher-level unit could shape lower-level choices to coordinate escape from a social dilemma or promote specialization through competition among local clusters.These are proposed scenarios in which data-stream manipulation serves the higher-level unit.Source: Major transitions in evolution · p. 6
- Collective-learning approaches often generalize poorly to novel situations, while hierarchical reinforcement learning faces interdependent learning signals across controller levels.High-level controllers need competent specialized units, while those units need a signal of their usefulness to the high-level task.Source: Major transitions in evolution · pp. 6–7
- Higher-level processes in AI can generate incentive alignment among lower-level units through population pressures, evolved incentives, or global incentive shaping.The cited approaches include sub-agents, evolving incentives for cooperation, and a global agent shaping incentives for selfish learners.Source: Major transitions in evolution · p. 7
- Coordination through data control can make lower-level units easier to control but may limit ongoing data generation as they become simpler.The paper identifies cumulative culture as an exception to this proposed limitation.Source: Major transitions in evolution · p. 7
Cumulative culture
- Cumulative culture lets learned information accumulate across generations through language and active teaching, supporting compounding innovation in human technology and knowledge.Compositional language supports new expressions built from existing ones, while oral tradition enables efficient distribution and storage of data.Source: Cumulative culture · p. 7
- Human cultural structures use language, networks, norms, and institutions to flexibly organize incentives, cooperation, and specialized work; by contrast, retraining language models on their own outputs may collapse data-distribution tails.The paper also notes that language models may struggle to consistently connect words to physical or social meanings, with proposed remedies including multimodal data, simulators, and search engines.Source: Cumulative culture · pp. 7–8
Synthesis
The paper argues that AI may face a data-generation bottleneck as models outpace human-produced data. Drawing on evolutionary processes, it proposes integrating mechanisms that enrich data streams, while noting that collective transitions can constrain lower-level cognition and data generation.
Synthesis
- As large language models outpace human data generation, human culture’s collective output may become a bottleneck for AI.The passage contrasts data assimilation with data creation.Source: Synthesis · p. 8
- Population pressures, arms races, social relationships, social learning, and cultural transmission can continually alter agents’ data streams and support ongoing data generation.The mechanisms affect learning through distinct interaction structures, including cooperation, competition, and multiscale coordination.Source: Synthesis · p. 8
- The paper favors a middle path: let flexible incentive alignment and data manipulation emerge, while engineering modules such as curiosity or social learning where adequate.It argues that individual mechanisms have achieved only partial success and proposes integrating generative processes that enrich data streams.Source: Synthesis · pp. 8–9
- Social relationships involve varied forms of cooperation and division of labor, while social bonds can reduce the cognitive load of managing relationships.The paper describes bonding neurochemistry as a foundation for other mammalian social bonds and group cooperation.Source: Synthesis · p. 10
- Major transitions can simplify and interdepend lower-level units, potentially limiting ongoing data generation because unified collectives need control.Examples include mitochondria and chloroplasts depending on nuclear genes, and specialized eusocial castes with reduced brain size.Source: Synthesis · p. 10