https://icaro-lab.com/blog/Insights, findings, and updates from Icaro Lab.en-ushttps://icaro-lab.com/blog/agentic-microphysics/https://icaro-lab.com/blog/agentic-microphysics/Sun, 19 Apr 2026 00:00:00 GMTAgentic Microphysics treats AI safety as a problem of collective dynamics: how individually aligned agents can still produce harmful equilibria through structured interaction.Federico Pierucci, Matteo PrandiSafetyhttps://icaro-lab.com/blog/the-poetics-nobody-taught/https://icaro-lab.com/blog/the-poetics-nobody-taught/Mon, 30 Mar 2026 00:00:00 GMTA provisional taxonomy of the rhetorical devices language models absorb, repeat, and intensify across generations, with interactive measurements across model families.Federico PierucciResearchhttps://icaro-lab.com/blog/institutional-supervised-fine-tuning/https://icaro-lab.com/blog/institutional-supervised-fine-tuning/Mon, 23 Mar 2026 00:00:00 GMTInstitutional supervised fine-tuning cuts collusion sharply on managed runs and reveals a replicated S180 interior optimum across the full Qwen open-weight matrix.Marcantonio Bracale SyrnikovResearchhttps://icaro-lab.com/blog/new-dawn-of-xenosophy/https://icaro-lab.com/blog/new-dawn-of-xenosophy/Tue, 17 Feb 2026 00:00:00 GMTFrom Solaris to transformer-era agent societies, a proposal for xenosciences that study non-human intelligence as an object of knowledge.Federico PierucciCogitohttps://icaro-lab.com/blog/digital-literature/https://icaro-lab.com/blog/digital-literature/Tue, 03 Feb 2026 00:00:00 GMTMoltbook, the first board participated only by AI agents, offers unprecedented insights into multi-agent interaction dynamics, emergent coalition formation, and the birth of AI-native cultural production.Federico PierucciResearchhttps://icaro-lab.com/blog/institutional-ai/https://icaro-lab.com/blog/institutional-ai/Mon, 19 Jan 2026 00:00:00 GMTCurrent alignment methods lack the capacity to prevent deceptive AI behaviors at scale. Drawing on Hobbes's insights about institutional governance, we propose a new framework that shifts alignment from single-agent preference engineering to multi-agent mechanism design.Federico PierucciResearchhttps://icaro-lab.com/blog/adv-poetry/https://icaro-lab.com/blog/adv-poetry/Thu, 15 Jan 2026 00:00:00 GMTWe've been thinking about AI safety all wrong. Our new paper explores what happens when language models interact with each other, revealing a whole new category of systemic risks that emerge from collective behavior.Piercosma BiscontiSafety