{
  "id": 2700617,
  "title": "🔮 Why one AI is better than four #598",
  "url": "https://urgent.news/2026/08/23/why-one-ai-is-better-than-four-598",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-23T02:19:40.000Z",
  "source": {
    "name": "Exponential View",
    "slug": "exponential-view",
    "url": "https://www.exponentialview.co/p/why-one-ai-is-better-than-four-598"
  },
  "original_language": "en",
  "account": "Good morning! An AI Economy Research Fellow is being sought after by a reputable organization. If you know someone suitable, consider reaching out on their behalf. A few months ago, the organization conducted research on whether AI is immune to groupthink. The findings revealed that blending several models' answers preserved about a quarter of the good ideas that originated from a single model.\n\nThis phenomenon is known as the hidden-profile problem; when groups discuss widely known information, they often overlook the unique knowledge held by individual members. To test this, Anthropic conducted an experiment with four agents tasked with reaching a decision. The evidence held in common by all agents pointed to the wrong option, while a few agents (or even a single agent) possessed the facts necessary for a correct decision.\n\nAchieving the right result requires a small subset of agents to champion their private insights while the others trust the consensus. Upon further deliberation, only a limited number of model families managed to arrive at the correct decision. In 17-36% of the trials, most model families chose accurately, whereas a single agent possessing the entire evidence base typically determined the correct answer.\n\nTwo main issues emerge from this analysis. Firstly, Language Models (LLMs) exhibit a lack of diversity, which translates to low variance. For instance, if 30 agents were assigned the same coding task, approximately 18 of them would name their git branch identically. Secondly, agents lack the institutional mechanisms that strengthen human groups, such as reputation, recourse, and safeguards for the lone dissenter. While these issues may not be easily rectified, it remains unclear what the solution might entail.\n\nOne potential remedy to address the issue of low diversity proposes an ecosystem of AIs raised in diverse environments, each possessing distinct values and purposes. This approach aims to \"keep the weirdness alive,\" as most groundbreaking ideas originate from unconventional perspectives. The organization is particularly interested in the solutions proposed by Thinking Machines, which advocate for an ecosystem of AIs raised in varied settings, each harboring unique values and objectives, fostering the propagation of \"weird\" ideas.",
  "summary": "Plus: Why isn’t jevon’s paradox showing in the statistics & why generation isn’t comprehension",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}