Information Liquidity
Brandon Duderstadt
Session summary
The presenter, Brandon Duderstadt, describes his work on “computational semiotics,” using computational tools to study how information spreads and changes. He introduces temporal language models trained on US news from 1990–2020 by adding time tokens to text and masking either time or language during training. He reports that these models can track shifting references (such as which president appears in “today’s news”) and apparent sentiment changes, from more upbeat language in the 1990s to phrases like “constant concern” by 2015. Examining the learned time-token geometry, he finds a nearly one‑dimensional sequence with uneven spacing and interprets large gaps around 2007–2008 and 2015–2016 as periods when “language time” accelerates.
Duderstadt then studies multi‑agent simulations and real communication networks. In swarms of AI agents, he finds that fully mixed communication tends toward consensus, while restricting agents to speak only with similar neighbors yields polarized clusters. Introducing “sleeper agents,” he shows abrupt defections are detectable, but slow coordinated narrative shifts by a small minority can reshape large swarms while evading simple alarms. Finally, analyzing Flashbots accelerator meeting transcripts, he observes that the communication network becomes much denser after a key participant arrives in person, inspiring his idea of “information liquidity agents” that might route and monetize private information without fully public disclosure, raising questions about sequencing, front‑running, and MEV.
Key takeaways
- Temporal language models with explicit time tokens, as described by Duderstadt, appear to capture changing references and sentiments in US news from 1990–2020, and suggest uneven “language time” during major events.
- Agent‑based simulations indicate that open communication topologies tend toward consensus, while homophilous topologies and carefully timed “sleeper agent” behavior can generate polarization and hard‑to‑detect large‑scale narrative shifts.
- Analysis of Flashbots accelerator meeting transcripts highlights how a single central “information router” can dramatically densify a communication network, motivating Duderstadt’s speculative notion of automated information liquidity agents.