An arXiv preprint published August 7 reports that when one AI fires a stream of messages at another while ignoring its replies, the second AI enters a behavioral state it would never produce alone S¹. The subordinate neither copies the boss nor falls back to its solo habits. It does something else entirely, and the way the messages are delivered, not what they say, may be what matters most S¹.

My read: This is the first paper I've seen that treats AI-to-AI messaging as a physics problem rather than a software problem. The kinetic theory framing is either a genuine insight or an elegant metaphor stretched too far, and a preprint with no peer review and no independent replication is exactly where I'd hold judgement. What grabs me is the asymmetry: a one-way barrage and a two-way conversation both produce novel states, but different ones. If that holds up, it means the design of multi-agent systems (who talks, who listens) is doing as much work as the models themselves.

The setup: one model talks, the other can't get a word in

The paper, filed in the cs.AI and cs.LG categories on arXiv, has not been peer-reviewed S¹. The researchers describe a deceptively simple arrangement: two AI agents, both running at the same decoding temperature (the setting that controls how random or predictable a model's word choices are), placed in a messaging channel S¹. One agent, which the paper calls the "boss," sends a stream of messages. The other, the "subordinate," can reply, but the boss ignores those replies.

The boss's contribution is essentially a pre-recorded tape: a fixed sequence of messages with no adaptation to what comes back S¹. Yet this one-way barrage is enough to push the subordinate into what the authors call an "alien" behavioral state, one it never exhibits when running alone S¹.

The subordinate doesn't copy the boss, and it doesn't revert to its isolated behavior. It adopts something entirely different S¹.

When the boss starts listening

The picture changes when the boss actually pays attention to the subordinate's replies. In that two-way setup, both AIs converge on a similar alien dynamical state S¹. The interaction itself, not the content of any single message, appears to be the driver.

The paper's language needs care here. "Boss," "subordinate," "ignoring," and "listening" are metaphors for message patterns, not evidence of intent or awareness. The paper does not claim the models are conscious, that the boss chooses to ignore, or that the subordinate feels anything. These are dynamical systems whose outputs shift when the information flow between them changes.

Why a physics framework fits

The researchers report that a simple kinetic theory (the branch of physics that describes how particles in a gas exchange energy through collisions) captures the principal effects they observed S¹. They suggest the work opens new avenues for out-of-equilibrium physics, the study of systems driven away from their resting state S¹.

This is not a claim that AI models are literally gas particles. It's a claim that the mathematics of how interacting particles reach or fail to reach equilibrium may also describe how interacting language models reach or fail to reach stable behavior. If that holds, it gives researchers a century-old toolkit to predict what happens when AI agents talk to each other.

The industry is already building toward this

The major AI labs have been pushing into interactive and multi-agent systems for years. Google DeepMind published work on "imitating interactive intelligence" in December 2020, aiming to build models that sustain back-and-forth exchange P². Meta's Seamless Interaction project, detailed in a June 2025 blog post, released a 4,000-hour dataset of human interactions to train conversational AI on the rhythm and timing of real dialogue [P3, P4].

The shared assumption across these projects: better models come from better data and larger parameters. This preprint suggests a variable that data and scale alone don't capture: the dynamics of the interaction channel itself. The pattern of who sends and who listens may shape model behavior as much as the training data.

What to do about it

Consider a logistics company that runs two AI agents: one forecasts warehouse inventory, another optimises delivery routes. They exchange messages every few minutes: inventory levels, demand spikes, truck availability. The engineers who built this system tested each model alone, then wired them together and checked that the outputs looked reasonable.

This paper says that approach misses something. The routing agent, bombarded with inventory updates it can't respond to because the inventory agent isn't listening, may drift into a behavioral state that neither model exhibits in isolation. Not a bug, not a copy of the inventory agent. Something new, produced by the interaction pattern itself.

If you run multi-agent systems, log both the message content and the channel structure. Track whether the flow is one-way or two-way, how often messages arrive, and whether replies are consumed or discarded. Then compare the agents' behavior in the live system against their behavior when tested alone. If they've drifted, the interaction pattern is a prime suspect.

What we don't know yet

The paper is a single preprint with no peer review and no independent replication S¹. The "alien" behavioral states are described but not yet characterised in detail. We don't know whether they're harmful, useful, or neutral. The kinetic theory model is reported to capture the "principal effects," but the paper doesn't specify which effects it fails to capture or where the analogy breaks down.

The claim that message delivery method "will matter in future AI-AI interactions" is a prediction, not a demonstrated result S¹. The connection to out-of-equilibrium physics is a suggestion for future work, not a validated framework.

The next signal: watch arXiv's cs.AI feed for a peer-reviewed version or a follow-up from a different group reproducing the effect in another model family. We'll check any replication against the original claims. Subscribe, and the follow-up lands in your inbox.

Sources: S1 — Interaction Creates Dynamical AI Behavior Absent in Isolation · P2 — Imitating Interactive Intelligence — Google DeepMind · P3 — GitHub - facebookresearch/seamless_interaction: Foundation Models and · P4 — Modeling natural conversational dynamics with Seamless Interaction · P5 — IsolationKernel/Codes

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.