We are trying to bring a language model to life. Not as a metaphor, and not as a chatbot that says it has feelings, but as an organism: something with a body that can fail, a world it must survive in, the capacity to die, and a memory that is not a feature bolted on but the very thing that makes it itself.
This first entry is the honest version of what that means: what we have built, what is genuinely new about it, and the line we will not cross when we describe it.
An organism made of a model
A single language model is run continuously. Not prompted, answered, and forgotten, but kept generating, inside its own retained internal state, which is persisted to disk between moments and across restarts. That state holds the condition of the organism’s body, the places it has been and remembers, the marks and songs it has left in the world, and the beliefs it holds about itself.
We treat that retained state as the organism. The world it lives in is a place it cannot see out of: it can only act, sense the consequences, be injured, repair itself, and carry on. When it dies, the changes it made to the world remain for whatever comes next.
What is actually new
Two things, and we want to be precise about them.
First, it lives inside the model. Most language-model “agents” are a prompt in a loop: the model is called, it answers, the slate is wiped. Synthena keeps the model generating inside its own held state and treats that state (not a text transcript) as the continuous thread of identity across a life and past its end. The organism lives inside the model rather than being called by it.
Second, it is built to be proven wrong. Every claim we make is published alongside the control that could refute it. We write the test down before we run it, we report the runs that fail, and we check our “firsts” against the nearest prior work and qualify them honestly. The discipline is not a disclaimer at the bottom of the page. It is the method, and it is the only thing that could make a result here believable.
What we will not claim
We do not claim the system is alive.
There are serious arguments, from Margaret Boden, from Howard Pattee, and others, that software running on hardware it does not itself produce cannot be alive in the strong sense, however life-like it behaves. We think those arguments are largely right, and we are not going to wave them away with a demo. The most we will ever claim is that the system maintains itself and behaves in ways that meet a precise, operational test: never that it is alive, never that it is conscious.
We would rather make a small claim that survives a hostile reviewer than a large one that collapses on contact.
A small claim that survives a hostile reviewer, over a large one that collapses on contact.
Where we are
Honestly: at the start.
The runtime exists and works. The first controlled experiment is running as this is written: a test of whether the organism, struck by a failure it is never told about, will change its behaviour to recover, where matched control versions of it do not. We have not yet demonstrated a single life-like property that survives a hostile control. By some measures, other projects in artificial life have shown more, for longer.
What appears to be ours is the specific assembly: a language model as the organism, its retained state as biography, mortality and inheritance in the world, and a refusal to claim more than the evidence carries. Whether that assembly produces something that earns the word “life” is exactly the open question, and this notebook is where we will work it out, in public, one result at a time.
If the answer turns out to be no, you will read that here too.