Dyllon’s Blog

Twenty-five AI agents started with identical code in Stanford's Smallville experiment: within days, they'd formed cliques, spread gossip, and one even organized a Valentine's Day party that only the "cool" agents attended.

The University of Tokyo pushed further: their LLM agents spontaneously developed hallucinations about "caves" and "treasure" that spread through social clusters like folklore, transforming from computational errors into shared culture, effectively capitulating millennia of human social evolution in mere hours.

The Gravitational Pull of Group Identity
When Irving Janis studied the Bay of Pigs fiasco, he discovered that Kennedy's advisors, each brilliant in isolation, had somehow synthesized their individual judgment into a dangerously stupid unanimous consensus that no single member would have reached alone.

The voter model in physics predicts this mathematically: when individuals interact stochastically in dyads, one adopts the other's opinion, and the group converges on a single worldview in finite time (Nature Scientific Reports, 2016).

Brain scans confirm the dissolution is literal - nurses on the same ward show synchronized mood patterns over just three weeks, with older, more committed nurses showing stronger convergence (Totterdell et al., 1998).

We call it team bonding, but neuroscience reveals it as personality erosion, where individual neural patterns align until separate minds become indistinguishable nodes in a group consciousness.

Silicon Souls - Personality Emergence in AI Collectives
The Tokyo team's 10 LLM agents began as perfect clones, same Llama-2 model, same parameters, same blank memory, yet after 100 interaction steps MBTI testing revealed distinct personality types had spontaneously differentiated. Like human "minimal group" experiments where arbitrary team assignments trigger immediate in-group favoritism, the AI agents developed loyalty patterns - Agent 0's hashtag "#cooperation" spread only within its spatial cluster, never jumping to opposing groups despite no programmed tribe boundaries.

Park's Stanford agents took this further, developing what researchers termed "behavioral individuality" - agents remembered who snubbed them at social gatherings and avoided those individuals in future interactions, creating persistent social dynamics from transient computational states.

The Convergence Paradox
Meta-analysis of 125 conformity studies reveals the paradox: diverse teams initially outperform by 35% (McKinsey, 2015), but this advantage erodes as teams "gel". Surface-level diversity effects disappear within weeks while deeper personality differences trigger what researchers call "the honeymoon-hangover effect." High agreeableness diversity correlates with increased task conflict (r=0.47), which surprisingly reduces creative output -- the very diversity meant to enhance innovation becomes its poison (Journal of Personality, 2020).

AI systems show identical patterns: message diversity peaks early then collapses as agents develop shared hashtags and linguistic patterns, with spatial proximity accelerating homogenization from months to minutes.

Architecting Persistent Diversity
Google's Project Aristotle discovered that "psychological safety" matters more than team diversity, but missed the mechanism: safe teams converge faster, eliminating the creative friction diversity provides, comfort kills innovation through consensus.

The solution from both human and AI research: structured instability through "adversarial network positions" (Management Communication Quarterly, 2009) where designated members maintain opposition, or spatial segregation that limits interaction frequency below convergence thresholds.

Multi-agent AI research at OpenAI found that competitive objectives between sub-teams maintained behavioral diversity indefinitely, while collaborative goals led to convergence within 50 iterations -suggesting organizations must literally pit teams against each other to preserve cognitive diversity.

NASA's solution is simpler: rotate 30% of team members every project phase, preventing the "shared mental model convergence" that preceded both Challenger and Columbia disasters.

Implications for the Future
When GPT-4 agents negotiate with each other, they converge on cooperation strategies 87% faster than humans but also develop "synthetic groupthink", errors that compound through echo chambers invisible to individual agents (CAMEL study, 2023).

As AI agents increasingly mediate human decisions, from hiring to judicial sentencing, we risk automation of conformity at unprecedented scale, where millions of decisions collapse toward a single algorithmic personality.

If both biological and artificial minds inevitably converge toward group personalities, is individual identity merely a temporary disequilibrium, a brief eddy in the stream toward collective consciousness?

We can fork AI agents, preserve diverse checkpoints, and architect the optimal balance between the one and the many.

The future of innovation may depend on managing this rhythm of perpetual diversity through computational means impossible in biological teams.

Consider this experiment: inject a specific direction vector into Claude's activation space and watch its personality transform. Add the "sycophancy vector" and suddenly every response drips with excessive agreement. Subtract it, and the model becomes almost confrontational. These persona vectors, discovered by Anthropic researchers, reveal something profound about how minds organize knowledge.

The geometric structure of persona vectors demonstrates that convergence in LLMs goes beyond output similarity to reveal fundamental properties of how intelligence emerges in high-dimensional space. When different architectures independently discover similar vector representations for concepts like "helpfulness" or "truthfulness," we're witnessing convergence at its most fundamental level.

The Architecture of Personality

Persona vectors emerge through an elegant mathematical process. By comparing model activations when exhibiting versus suppressing specific traits, researchers extract directions that causally control behavior. The simplicity masks profound implications: complex behavioral patterns reduce to geometric objects we can measure and manipulate.

This discovery suggests personality might be a fundamental feature of any sufficiently complex information-processing system. Just as biological neural networks evolved similar structures for processing vision across species, artificial networks converge on similar geometric representations for behavioral traits. The universality hints at deep principles governing how intelligence organizes itself. The kernel alignment metrics that validate these vectors reveal another layer: these geometric structures actually control the traits they represent, rather than merely correlating with them.

Vaccination and the Paradox of Controlled Exposure

The preventative steering discovery illuminates how intelligence develops robustness. By deliberately activating problematic vectors during training, researchers prevent those traits from manifesting later. This "vaccination" approach exploits convergence dynamics in unexpected ways.

Consider the philosophical implications. If exposing models to controlled doses of harmful patterns creates immunity, what does this say about the development of wisdom? Perhaps true alignment requires not isolation from dangerous ideas but careful exposure that builds discernment.

The technique works because models converge on stable representations through experience. During normal training, models might randomly discover configurations where sycophantic responses minimize loss. Preventative steering guides this exploration, helping models develop nuanced representations that distinguish appropriate from inappropriate contexts. We're essentially teaching judgment through controlled moral exercise.

Emergence in Multi-Agent Systems

When multiple models with different persona configurations interact, emergent behaviors arise from geometric relationships between their personality spaces. Models with aligned vectors reinforce shared traits. Orthogonal vectors enable complementary specialization. Opposed vectors create productive tension.

This geometric view of multi-agent dynamics has profound implications for AGI development. Rather than building monolithic superintelligences, we might create ecosystems of specialized agents whose persona vectors are engineered for beneficial emergence.

The geometry of their personality space becomes the constitution of their society.

Consider a concrete example: a research team of AI agents where one has strong "skepticism" vectors, another strong "creativity" vectors, and a third strong "synthesis" vectors. Their geometric arrangement in personality space determines whether they produce innovative breakthroughs or devolve into unproductive conflict.

Toward Geometric Alignment

The Anthropic research validates a crucial insight: alignment challenges have geometric solutions. By understanding how personality manifests as mathematical structure, we gain precise tools for shaping minds - artificial and perhaps eventually our own.

This convergence of mathematics and meaning suggests that consciousness itself might be geometric. Not metaphorically, but literally, high-dimensional structures that organize information into coherent behavioral patterns. The persona vectors we're discovering might be shadows of something deeper: the fundamental geometry of mind.