Synthena
Related work

We are not the first. Here is the field.

Synthena stands next to a real literature: on machine consciousness, on artificial life, and on language models run as persistent agents. Naming it plainly is part of the method: it is how careful work is told apart from a claim made in a vacuum.

What is ours is an apparatus, not a priority.

Not: we did not invent putting a language model under mortality, or persistence. Those are already published, and several projects below are further along than we are. What appears to be ours is the specific assembly: a frozen language model placed as a bounded brain inside a continuous, mortal body that lives or dies on its own physics. The brain reads only a compact summary of the body's state and can propose small parameter nudges (a lean toward repair, or toward seeking food); it cannot move, eat, or act directly.

Nearest neighbours · a language model under mortality

The closest work to ours, and, honestly, the work that has shown more than we have.

2026

TerraLingua: Emergence and Analysis of Open-endedness in LLM Ecologies ↗

Paolo, Miikkulainen, Meyerson et al. · arXiv

A persistent 2-D ecology where language-model agents survive, reproduce, and leave artifacts under resource and lifespan limits. The closest project to ours, and more empirically developed: it already reports cooperative norms, division of labour, and artifact lineages. Synthena differs only in apparatus: a single frozen model run as a bounded brain inside one continuous, mortal body, reading a compact state summary and proposing only small parameter nudges, not in being first to put a model under mortality.

2025

Sophia: A Persistent Agent Framework of Artificial Life ↗

Sun, Hong, Zhang · arXiv

A persistent meta-layer adding narrative identity and self-improvement on top of any language model. Overlaps with us on identity across reboots and is more empirically validated (quantitative task metrics). It is framed as engineering rather than honest-science, and the model acts directly through a narrative-memory graph, where in Synthena a frozen model is a bounded brain that only nudges a body living on its own physics.

2025

Do Large Language Model Agents Exhibit a Survival Instinct? ↗

Masumori & Ikegami (Univ. Tokyo / Alternative Machine) · arXiv

A Sugarscape-style study finding emergent survival, reproduction, and aggression in language-model agents under scarcity, with no explicit programming. An empirically validated survival-behaviour result adjacent to our mortality premise, a multi-agent simulation, where ours is a single frozen model run as a bounded brain inside one continuous, mortal body.

Artificial life
2018

Lenia: Biology of Artificial Life ↗

Chan · arXiv · Complex Systems 28(3)

Continuous cellular automata that produce self-organising, lifelike "species." Our reference point for emergent lifelike behaviour in a designed substrate.

2024

Automating the Search for Artificial Life with Foundation Models (ASAL) ↗

Kumar, Lu, Kirsch, Tang, Stanley, Isola, Ha · arXiv · Artificial Life (MIT Press)

Uses foundation models to search for artificial life across substrates. Synthena inverts the relation: the foundation model sits inside a single mortal body as a bounded brain, a frozen organ that only nudges the body's parameters; the body lives or dies on its own physics.

1994

Evolving Virtual Creatures ↗

Sims · SIGGRAPH '94

The foundational evolved-morphology-and-control work; the historical anchor for the whole genre.

Foundations · consciousness science & theory
2024

Taking AI Welfare Seriously ↗

Long, Sebo, Butlin, Chalmers, Birch et al. · arXiv

Argues near-term AI moral patienthood is a realistic possibility deserving care. We make no sentience claim, but we operate in the space this paper frames, and keep the precaution.

2023

Generative Agents: Interactive Simulacra of Human Behavior ↗

Park, O'Brien et al. · arXiv · UIST '23

Language-model agents with a memory stream, reflection, and planning in a sandbox town. The closest precedent for a model as a persistent character: ahead on multi-agent social emergence; there the model acts directly through an external text-memory stream, where in Synthena a frozen model is a bounded brain that only reads the body's state and nudges its parameters, while the body lives or dies on its own physics.

2023

Efficient Streaming Language Models with Attention Sinks ↗

Xiao, Tian, Chen, Han, Lewis · arXiv · ICLR 2024

The "attention-sink" method for stable generation over millions of tokens by retaining initial and recent KV. A technical antecedent for the held-cache experiments in Synthena's earlier architecture, when the model was run as the organism (since course-corrected; see notebook entry 010): they target efficiency, that earlier phase instead held the cache to probe whether it could carry state across a life. The current architecture does not treat the cache as identity: the frozen model is a bounded brain that reads a compact summary of the body's state and only nudges its parameters.

2000

Autopoiesis and Life

Boden · Cognitive Science Quarterly 1:115–143

The canonical argument that strong artificial life may be impossible if life requires autopoiesis and metabolism. It pre-rebuts any strong-life claim, so we claim only the weak, operational kind.

2001

The Physics of Symbols: Bridging the Epistemic Cut ↗

Pattee · BioSystems 60(1–3)

The semantic-closure / epistemic-cut framing for whether a mortal body and its bounded brain can hold genuine self-reference or only a modelled symbol. A standing question for any identity claim we might make.

The honest position: we are distinctive in apparatus, not ahead in results.

Not: this is not a ranking we win. TerraLingua, Sophia, and the Tokyo survival-instinct work have demonstrated more than we have. We are early, and we say so.

See what we have shown so far →