LIQUID-I: Our Proposed Scientific Direction for the ERC Synergy Project

LIQUID-I is a proposal to find general laws for what happens to agency — the capacity to model, evaluate, and plan — as learning agents (human, microbial, or artificial) start inheriting each other's acquired skills instead of just their genes.

The core intuition: normally evolution only passes on what's baked into DNA. But once agents can copy, teach, or record what they've learned — a skill, a policy, a model of the world — inheritance gets faster and stranger. A population can become smarter collectively while individual members quietly lose the ability to function without the group. This is the paper's animating worry: collective capability and individual agency can move in opposite directions, and we don't have good tools for detecting when that's happening until it's too late. The proposal calls this decoupling "the Lamarckian regime," and wants to study it as a causal, testable mechanism rather than a metaphor — separating "the group did well" from "the group is an agent" from "the group's success required someone losing autonomy."

Three threads carry this. First, real biology: bacterial or yeast colonies coupled to digital controllers, where you can literally delete or transplant "acquired memory" and watch whether performance collapses — a physical testbed for dependence and recovery that's normally impossible to run on humans or societies. Second, artificial societies: language-model agents, whole-brain-inspired agents, and neuromorphic (brain-chip) agents interacting in shared environments, letting the team ask which principles of cooperation and collapse survive when you change the underlying computational substrate — is instability a property of the architecture or just the implementation? Third, human data (no new experiments, just re-analysis of existing neuroimaging/mental-health datasets) to ground claims about what "valence" — the felt quality of an evaluation, borrowed from the authors' Kolmogorov Theory / Algorithmic Neuropsychiatry framework — actually looks like mechanistically, so that claims about AI or collective "well-being" aren't just reward numbers dressed up as feelings.

What makes this more than a grab-bag of experiments is the explicit refusal to conflate related-but-distinct claims: coordinated behavior isn't collective agency, collective reproduction isn't collective planning, a positive average outcome isn't absence of harm to some members, and an LLM saying "I'm happy" isn't evidence of experience. The paper is disciplined about what its own precedents (forest-fire control games, the "cognitive-virus" dependency model, Agent Souper, large LLM-society simulations) actually established versus what LIQUID-I still has to test — notably, that big agent populations don't by themselves prove anything about human realism or multi-generational dynamics.

Worth flagging plainly: this is a discussion draft for an ERC Synergy grant, not a results paper. It reports zero new experiments or findings — it's a map of nine open questions and three appendices of proposed (not run) experiments, meant to get the research consortium arguing about priorities before the actual work starts.