Emergent ComputingThe paper Emergent Models: Intelligence from Tiny Substrates makes one proposal, one theorem and a set of small experiments. This note separates the three and lists what each does and does not establish.
The proposal
Treat a simple dynamical system as a learning model. Write the input into its state, iterate a fixed local rule until a halting condition holds, read the output. Train mainly by searching over the initial state, the program, rather than over the weights of a fixed architecture. The proposal is a framing: it widens the space of models, and it is useful only if instances turn out to be trainable and to generalise.
What is proved
Some Emergent Models are latent-universal. With rule, halting condition and interface all fixed, every partial computable function on binary strings is computed by some finite program. The proof is short because it reuses universal Turing machines whose program and input sit in separate regions of the tape; the paper also extends it to the Game of Life.
What it does not prove: that such programs are easy to find, that any of the experimental models is latent-universal, or anything about efficiency. Universality removes a ceiling on what can be represented. It does not make learning possible.
What is shown experimentally
- EM43, a 1D cellular automaton with 70 parameters (140 bits), trained on x ≤ 30, computes x + 1, 2x, x mod 4, round(x/3) and x·(1 + x mod 2) exactly on every tested input up to 1000, in every run. The working solutions are visible as regular geometric patterns in the space-time diagram.
- Game of Life with the rule frozen can learn x + k by placing two gliders, but random search does as well as the genetic algorithm.
- Continuous lattices control CartPole (735 ± 107 out of 800 with three trainable coefficients) and adapt online to alternating food and poison phases (meta-score 0.71 on a 0–1 scale).
What is not shown
- Competitiveness. No comparison against tuned standard methods on standard benchmarks. CartPole can be solved by a linear controller, and on the adaptive task a small dense RNN did better (0.82).
- Harder functions. Multiplication of two inputs was not learned (< 10% test accuracy). It is not known whether the cause is compute, hyperparameters, the interface or the rule space.
- Statistics. Results are representative runs, not multi-seed distributions.
- Robustness. The controllers fail in identifiable ways: drifting out of bounds, rotating in place, freezing.
What we take from it
Two observations seem worth building on. First, the extrapolation results come from models that iterate a rule for an input-dependent number of steps, consistent with a broader pattern across looped and recursive neural models. Second, search is the bottleneck. The Game of Life result shows a landscape in which similar programs have unrelated fitness, and the paper conjectures that this is the price of having few parameters reused over many steps. Both are stated as hypotheses we intend to test.
The concrete next steps are machines with a fixed interface where only the program changes (LUTM-1, LUTM-6), neural cellular automata that share one rule across tasks (NCPU-EM), and search methods built for rugged landscapes.