In Gennady Muzykantov’s simulation, some internal signals predicted breakdown, while others restored the system more effectively
In Gennady Muzykantov’s simulation, some internal signals predicted breakdown, while others restored the system more effectively
On July 14, engineer and data specialist Gennady Muzykantov described on Habr an independent computational experiment based on models developed by Alexander Mordvintsev and Michael Levin. He ran 50 neural cellular automata, tracked their breakdown, and compared early signs of aging with the channels whose modification returned the system to a stable state.
A neural cellular automaton consists of a grid of digital cells. Each cell can see only its nearest neighbors, and every cell applies the same learned rule. These local actions produce an overall form without central control, in this case an image of a lizard.
Each cell has four visible channels that determine its color and presence, along with 12 hidden channels. The hidden channels provide internal memory and transmit signals. Their roles are not defined in advance. The neural network determines how to distribute across them the information needed to assemble the form.
Muzykantov used an automaton trained to grow a lizard and continued updating it after it reached its adult form. The system remained stable for some time, but its cells eventually died or proliferated without control. Across 50 runs, the resulting curve showed a long stable period followed by a rapid increase in mortality. Biological survival curves often have the same general shape, but this similarity alone does not establish a shared mechanism.
Some hidden channels began to diverge from the adult state hundreds of steps before visible breakdown. They acted as early model biomarkers that made it possible to detect an approaching collapse in advance.
The author then waited until half of the population had died and began moving individual channels slightly toward their adult values at each step. Adjusting the early markers helped, but a brute force search identified a pair of channels, 6 and 9, that worked better. Further deaths among the remaining lizards nearly stopped, and the number of living cells returned to approximately 95–98% of the adult level.
The form recovered much less completely. Its similarity to the adult lizard increased from approximately 10% to 55–60% and then stopped improving. The system became stable again and recovered nearly all of its cell mass, but it remained far from the adult structure.
A good predictor of aging may be only a modestly effective target for intervention. Survival or cell mass may improve more than tissue organization. Evaluating rejuvenation therefore requires separate criteria for function, structure, and stability.
In the published model by Leo Pio-Lopez, Benedikt Hartl, and Michael Levin, aging emerged as the loss of an anatomical target after development was complete, while directed signals initiated restoration. Muzykantov added an engineering question to this idea: which internal coordinates predict breakdown, which control recovery, and which features demonstrate that recovery has occurred.
The specific numerical results cannot currently be verified independently because the article does not include the code, model weights, or raw data, and channels 6 and 9 were initially selected through a brute force search using one aging lizard. The experiment provides a criterion for biological clocks and models that reduce an organism’s state to a set of measurements: a shift in measured age should be accompanied by the restoration of viability, function, and structure.