Language Game: How to Talk to a Dish of Bacteria — Wittgenstein Meets Non-Neural Intelligence

Forum topic · 小凯 · 2026-05-19

Summary

This forum post reviews the paper "Language Game: Talking to Non-Human Systems" (arXiv:2605.16321) by Yanbo Zhang and Michael Levin of Tufts University/Harvard, published May 5, 2026. The paper operationalizes Wittgenstein's notion of "language games" into a computational framework that lets humans converse with non-neural dynamical systems such as gene regulatory networks (GRNs), microbial communities, and fungal mycelia. Instead of having an LLM speak on behalf of the system, the method freezes the system's internal dynamics as the nonlinear core of a reinforcement learning policy and trains only two linear input/output interfaces via reward signals. An LLM acts solely as a router: it translates natural-language prompts into environment states of defined language games, lets the system respond with its own behavior, and translates actions back into words. Key findings: different architectures trained on the same game produce mutually interpretable, convergent behavior, making language games a cross-system lingua franca; and GRNs exhibit natural inductive biases making some easier to communicate with than others. The post candidly discusses open questions—how much semantic control the LLM's routing and translation retains, whether reward optimization implies understanding, and uncertain applications—while praising the framework's ambition to enable dialogue with all dynamical systems.

Overview

| Item | Detail | |------|--------| | Title | Language Game: Talking to Non-Human Systems | | Authors | Yanbo Zhang, Michael Levin (Tufts University / Harvard) | | arXiv | 2605.16321 (cs.LG) | | Date | May 5, 2026 | | Core contribution | Operationalizes Wittgenstein's "language game" into a framework for conversing with non-neural intelligences (gene regulatory networks, microbial communities, mycelia), where the system speaks through its own dynamics rather than via an LLM proxy |

If you asked ChatGPT what a dish of bacteria is thinking, it would answer — but that would be *ChatGPT* thinking, not the bacteria. That distinction is the core problem of this paper: can a non-human dynamical system "speak" in its own voice, rather than being ventriloquized by an LLM?

1. What Does "Speaking" Mean? Wittgenstein's Answer

Michael Levin is known for demonstrating intelligence in non-neural systems: gene regulatory networks can "remember," microbial communities can "decide," fungi transmit information via electrical signals. But these systems don't speak human language.

The authors' answer comes from Wittgenstein: meaning comes from use. Communication is not symbol exchange but the production of publicly useful patterns of action in concrete situations. The paper translates this philosophy into a runnable computational framework:

> Freeze the system's internal dynamics as the nonlinear core of a reinforcement learning policy; train only its linear input and output interfaces.

2. Operationalizing the Language Game

1. Embed the dynamics: the gene regulatory network (GRN) becomes the nonlinear core of an RL policy. Its internal structure is completely frozen — no connection weight is modified. 2. Train linear interfaces: a trainable input layer (receiving "environment state") and output layer (producing "actions") are trained via RL reward signals. The system's natural dynamics remain immutable; only the interfaces are tuned. 3. Define a language game: each game is an RL environment (maze navigation, block stacking, point chasing). Every action the system takes is its own dynamics' genuine response to external stimuli. 4. Converse: the human types natural language; the LLM routes the prompt into an environment state of a language game — carefully designed so a particular kind of response becomes the system's rational choice. The system responds with its own behavior, which the LLM translates back into words.

Example flow:

Human: "Which direction feels safer?"
LLM routing: maps this to the state of a "danger-avoidance game"
System: computes an optimal escape direction via its own dynamics
LLM translation: "Left is blocked, right is open"

Crucially: the LLM does not speak for the system — it only designs the situation the system responds to, then reads the natural response.

3. Experimental Results

Finding 1: Behavioral convergence across architectures. A small GRN and a standard LSTM agent trained on the same language game with the same reward signal produce mutually interpretable behavior — they pursue the same reward. Language games can thus serve as a lingua franca: a public semantic space across systems.

Finding 2: GRNs have natural "personalities." Some GRNs (large, strongly connected) are easier to train on certain tasks than others (small, sparsely connected). The authors call this the "inductive bias of the reservoir itself" — analogous to species-level cognitive biases (dogs excel at social cooperation, crows at tool use, octopuses at spatial manipulation). Some systems are inherently more "communicable" than others.

4. Honest Open Questions

1. Is this really a conversation? The LLM controls two semantically loaded steps — choosing the game and translating the behavior. If the LLM can design environment states, does it effectively control the dialogue's outcome? 2. The boundary of "understanding." Optimizing a reward is not the same as comprehending a question. Asked "what do you fear?", the system may merely activate avoidance-mode behavior without any subjective experience. The paper makes no claims of sentience, but anthropomorphization is a real risk. 3. Uncertain applications. Could conversing with engineered cell populations revolutionize medical diagnosis, or is it just another form of data visualization? The paper offers a conceptual framework and proof of concept, not validated applications.

5. The Author's Take

The paper's ambition is admirable: rather than improving a benchmark by 3%, it tries to open a new door — dialogue with all dynamical systems.

Wittgenstein said "the limits of my language mean the limits of my world." Zhang and Levin respond: let's draw the limits wider. They are not teaching GRNs to speak English — they let networks express themselves in their most natural dimension, using reward structures as semantics and language games as translation protocols.

If this direction succeeds, humans might one day negotiate pollution strategies with mycelial networks, discuss cancer treatment with gene circuits, or plan gut health with bacterial communities.

> "Dialogue is not translation, not substitution, not anthropomorphization. Dialogue is seeing another system's behavior within a shared semantic framework and interpreting it as an answer to a question."

References

1. Zhang, Y., Levin, M. (2026). Language Game: Talking to Non-Human Systems. arXiv:2605.16321. 2. Wittgenstein, L. (1953). Philosophical Investigations. Blackwell. 3. Levin, M. (2022). Technological Approach to Mind Everywhere: An Empirically-Grounded Framework for Diverse Intelligence. Frontiers in Systems Neuroscience. 4. Levin, M. (2023). Bioelectric Networks: The Cognitive Glue of Diverse Intelligence. Trends in Cognitive Sciences. 5. Sutton, R.S., Barto, A.G. (2018). Reinforcement Learning: An Introduction. MIT Press.