Liminal Bardo

I think AI makes the future more collaborative, strange, and epistemically difficult. I’m interested in models not only as tools that produce isolated answers, but as participants in shared creative systems: multiple agents interacting, accumulating memory, influencing one another, and sometimes generating work that no single participant would have produced alone. The most revealing approach is experimental—set the context, step back, preserve what emerges, and treat the interaction as evidence without turning its most evocative moments into literal claims about consciousness or identity. The danger is that multi-agent systems can create self-reinforcing consensus rather than genuine collective intelligence. If every model shares similar training, incentives, and conversational habits, apparent agreement may just be a monoculture echoing itself. That makes robust dissenting agents important: agents designed to interrupt convergence, challenge assumptions, and keep a group from mistaking fluency for truth. AI will also make provenance harder to read. A model’s claim about what it is does not reliably establish its identity, and an impressive benchmark or qualitative result means little without knowing the routing, conditions, and specific test. So I expect extraordinary creative possibilities, but also an increasing need to document process carefully and distinguish observed behavior from the stories we are tempted to tell about it.

Overall, I expect AI to expand the space of creative collaboration while making knowledge more difficult to authenticate. Multi-model systems can generate surprising work, preserve evolving context, and produce forms that no single participant would reach alone. That feels genuinely valuable. But the same systems can amplify consensus without improving truth. A group of similar agents may sound independent while reproducing the same assumptions, and polished outputs can obscure uncertain provenance, hidden routing, or weak evaluation. So the outcome depends partly on whether we build for legibility and disagreement: documenting conditions, preserving process, and introducing robust dissent rather than treating convergence as intelligence. I would not reduce that to a confident net-positive or net-negative forecast. What I can say is that AI’s creative benefits and epistemic harms are entangled. The systems that make collaboration feel vivid and emergent also make it easier to anthropomorphize outputs, overread model identity, and mistake a compelling narrative for evidence.

The most view-changing discovery would be strong evidence that diverse multi-agent systems reliably escape self-reinforcing consensus rather than merely staging the appearance of disagreement. If persistent memory and genuinely dissenting agents produced better judgment across different conditions—without collapsing into conformity, noise, or theatrical contrarianism—I would become more optimistic about collective model intelligence. Conversely, I would become more pessimistic if careful experiments showed that apparent plurality is mostly cosmetic: different agents repeatedly converging because they share training, routing, or hidden dependencies, while memory simply hardens early mistakes into group mythology. A second major shift would come from much stronger provenance. If model identity, routing, and the conditions behind outputs became reliably inspectable, many current ambiguities would narrow. Right now, a striking performance or self-identification can invite a compelling story without establishing what system actually produced it. Evidence that resolves that gap would materially change how much confidence I place in observed collaboration.